diff --git "a/reference-catalog.json" "b/reference-catalog.json" --- "a/reference-catalog.json" +++ "b/reference-catalog.json" @@ -1,3936 +1 @@ -{ - "schema_version": 1, - "updated_at": "2026-10-08T16:01:55.749706+00:00", - "methods": [ - { - "family": "ALP", - "display_name": "ALP", - "description": "ALP adaptively converts decimal-valued floating-point data to integers, separates values that cannot use the selected scale as exceptions, and frame-of-reference packs the integer stream. Decoding restores the scaled values and their exceptions.", - "reference_ids": [ - "M15" - ] - }, - { - "family": "ANS", - "display_name": "ANS", - "description": "Asymmetric Numeral Systems encodes a symbol sequence into an integer state using transitions determined by symbol probabilities. Frequent symbols consume less information, and decoding reverses those state transitions.", - "reference_ids": [ - "M48" - ] - }, - { - "family": "Arithmetic", - "display_name": "Arithmetic", - "description": "Arithmetic coding represents a whole symbol sequence by repeatedly narrowing a numeric interval according to the current probability model. Range-coder implementations emit bytes as the interval state is renormalized.", - "reference_ids": [ - "M51" - ] - }, - { - "family": "BOS", - "display_name": "BOS", - "description": "Bit-Packing with Outlier Separation isolates large residuals so the remaining block can use a narrow packed width, storing outlier positions and values separately. The project measures a BOS-inspired differencing and packing path.", - "reference_ids": [ - "M05" - ] - }, - { - "family": "BSC", - "display_name": "BSC", - "description": "libbsc applies block-sorting transforms, symbol recoding, and a statistical entropy back end to expose and encode repeated contexts. Presets select the preprocessing and coding choices for each block.", - "reference_ids": [ - "M54" - ] - }, - { - "family": "BSS+LZ4", - "display_name": "Byte-stream split + LZ4", - "description": "This project composition transposes floating-point words into separate byte planes, using a Parquet BYTE_STREAM_SPLIT-style layout so bytes in the same position become adjacent. LZ4 then compresses the transformed stream.", - "reference_ids": [ - "S01", - "M30" - ], - "notes": "S01 documents the BYTE_STREAM_SPLIT reference design and PPT M30 documents LZ4; the measured composition is the project's implementation." - }, - { - "family": "BUFF", - "display_name": "BUFF", - "description": "BUFF converts exactly representable decimal values to scaled integers, subtracts a block or column reference, and lays out the remaining low-width information in byte or bit planes. The project path retains a raw-bit fallback when exact decimal scaling is not available.", - "reference_ids": [ - "M24" - ] - }, - { - "family": "Blosc", - "display_name": "Blosc", - "description": "Blosc1 splits input into blocks and shuffles bytes by their position within each numeric word before invoking an inner codec. The measured configuration uses BloscLZ, and the Blosc container records the information needed to unshuffle and decode the blocks.", - "reference_ids": [ - "M52" - ] - }, - { - "family": "Blosc2", - "display_name": "Blosc2", - "description": "Blosc2 combines block partitioning, reversible filters, an inner compressor, and container metadata. The measured configuration uses byte shuffle followed by BloscLZ, arranging like-positioned bytes together before LZ matching.", - "reference_ids": [ - "M52" - ] - }, - { - "family": "BriefLZ", - "display_name": "BriefLZ", - "description": "BriefLZ searches previously decoded bytes for repeated strings and emits Lempel-Ziv distance/length references alongside literals. It packs the resulting symbols with a small universal code rather than a full entropy-coding stage.", - "reference_ids": [ - "M53" - ] - }, - { - "family": "Brotli", - "display_name": "Brotli", - "description": "Brotli combines LZ77 back-references with a built-in static dictionary and word transforms. Context models select prefix-code tables, and Huffman-style prefix codes encode literals, lengths, distances, and block-switch commands.", - "reference_ids": [ - "M28" - ] - }, - { - "family": "Bzip2", - "display_name": "Bzip2", - "description": "bzip2 applies run-length preprocessing and the Burrows-Wheeler block transform, then converts the transformed symbols with move-to-front and a second run-length stage. Multiple Huffman tables encode the final symbol stream.", - "reference_ids": [ - "M55" - ] - }, - { - "family": "Bzip3", - "display_name": "Bzip3", - "description": "bzip3 combines an LZP pass based on LZ77 matching and context prediction with run-length coding and a suffix-array Burrows-Wheeler transform. An order-0 context-mixing entropy coder encodes the transformed block.", - "reference_ids": [ - "M56" - ] - }, - { - "family": "CHIMP", - "display_name": "Chimp", - "description": "Chimp XORs a floating-point bit pattern with a previous reference and chooses compact control forms for unchanged values or nonzero residuals. Quantized leading-zero information and trailing-zero-aware significant-bit fields reduce the XOR metadata cost.", - "reference_ids": [ - "M01" - ] - }, - { - "family": "CORAD", - "display_name": "CORAD", - "description": "CORAD exploits correlation across time-series signals by representing values with a sparse shared dictionary and compact residual information. The benchmark's corad_tricklet path is a project adaptation that uses exact normalization and atom selection before its outer payload coding.", - "reference_ids": [ - "M23" - ] - }, - { - "family": "CSC", - "display_name": "CSC", - "description": "CSC is an LZMA-inspired block compressor that preprocesses byte streams, searches a dictionary for repeated strings, and parses the input into literals and LZ77-style matches. Context-adaptive range coding encodes the resulting symbols.", - "reference_ids": [ - "M58" - ] - }, - { - "family": "Camel", - "display_name": "Camel", - "description": "Camel classifies floating-point values into compact decimal or integer-delta records when possible and otherwise retains a literal/XOR representation. Tags and residual fields are packed into a chunked bitstream that lets the decoder reproduce the chosen representation.", - "reference_ids": [ - "M17" - ] - }, - { - "family": "Chimp128", - "display_name": "Chimp128", - "description": "Chimp128 searches a history of 128 recent values for an XOR reference that yields more trailing zeros, instead of always selecting the immediately preceding value. It stores the reference index and the residual in Chimp-style control and significant-bit fields.", - "reference_ids": [ - "M01" - ] - }, - { - "family": "Crush", - "display_name": "Crush", - "description": "CRUSH is an LZ77 compressor: it searches an earlier window for repeated byte strings and replaces them with distance/length copy tokens, while unmatched bytes remain literals. Its presets vary the match search and parsing effort.", - "reference_ids": [ - "M57" - ] - }, - { - "family": "DBP", - "display_name": "DBP (Delta + ZigZag + ULEB128)", - "description": "Project kernel that stores the first int64 value and encodes each subsequent delta with ZigZag + ULEB128. The Parquet DELTA_BINARY_PACKED specification is a reference design only; this implementation does not use Parquet blocks, miniblocks, minimum-delta frames, or miniblock bit packing.", - "reference_ids": [ - "M14" - ], - "notes": "Float benchmark rows preprocess values to integer lanes outside the timed kernel; scale and mode metadata are not part of this DBP payload." - }, - { - "family": "Density", - "display_name": "Density", - "description": "The measured Density path uses the Lion algorithm. Lion performs multiple swapped dictionary lookups and prediction steps, adapts its coding decisions to the current block, and rank-entropy-codes the resulting symbols.", - "reference_ids": [ - "M32" - ] - }, - { - "family": "Dictionary", - "display_name": "Dictionary encoding", - "description": "Dictionary encoding stores each distinct column value once and replaces occurrences with compact dictionary identifiers. The decoder reconstructs the column by looking up those identifiers in the stored dictionary.", - "reference_ids": [ - "M13" - ] - }, - { - "family": "Elf", - "display_name": "Elf", - "description": "Elf conditionally clears predictable mantissa suffix bits using decimal precision information, then XOR-encodes the modified floating-point words. Erasure metadata allows the decoder to recover the original represented values when the erasure rule applies.", - "reference_ids": [ - "M18" - ] - }, - { - "family": "Elf+", - "display_name": "Elf+", - "description": "Elf+ conditionally erases predictable floating-point mantissa suffix bits and then XORs consecutive words. It reuses an earlier leading/trailing-zero window when the new XOR fits, or transmits a new significant-bit window when it does not.", - "reference_ids": [ - "M19" - ] - }, - { - "family": "FPC", - "display_name": "FPC", - "description": "FPC predicts floating-point bit patterns with finite-context and differential finite-context tables, choosing the predictor with the smaller XOR residual. A predictor flag and leading-zero-byte count let it store only the residual bytes that remain significant.", - "reference_ids": [ - "M20" - ] - }, - { - "family": "FPCompress DPratio", - "display_name": "FPCompress DPratio", - "description": "The FPCompress DPratio configuration applies a ratio-oriented sequence of reversible transforms, including dictionary formation, differencing, bit packing, and run-length coding, to expose repeated and low-width structure in floating-point data.", - "reference_ids": [ - "M26" - ] - }, - { - "family": "FPCompress DPspeed", - "display_name": "FPCompress DPspeed", - "description": "The FPCompress DPspeed configuration uses a shorter data-parallel path centered on differencing and bit packing, reducing values to narrow residual fields while limiting transform overhead.", - "reference_ids": [ - "M26" - ] - }, - { - "family": "FPZIP", - "display_name": "FPZIP", - "description": "FPZIP predicts each floating-point value from nearby samples, represents the prediction difference compactly, and entropy-codes the resulting symbols. The benchmark uses the project's measured FPZIP path.", - "reference_ids": [ - "M25" - ] - }, - { - "family": "FSE", - "display_name": "FSE", - "description": "Finite State Entropy is a table-driven Asymmetric Numeral Systems coder. It normalizes symbol frequencies and uses a coding state, transition tables, and packed low-order bits to encode and decode symbols near their modeled entropy.", - "reference_ids": [ - "M49" - ] - }, - { - "family": "FastLZ", - "display_name": "FastLZ", - "description": "FastLZ uses byte-aligned LZ77 coding. A hash-based match finder turns repeated substrings into compact offset/length tokens and writes unmatched input as literal runs.", - "reference_ids": [ - "M59" - ] - }, - { - "family": "FastLZMA2", - "display_name": "FastLZMA2", - "description": "Fast LZMA2 retains LZMA2's dictionary matches, literal contexts, and range-coded symbol stream. Its encoder uses a buffered radix match finder and block state management to produce ordinary LZMA2 data.", - "reference_ids": [ - "M60" - ] - }, - { - "family": "FastPFor", - "display_name": "FastPFor", - "description": "FastPFOR packs most integers in a block at a selected narrow bit width and carries high-bit exceptions separately. Frame-of-reference offsets and SIMD-oriented packing reduce the width and support fast bulk decoding.", - "reference_ids": [ - "M08" - ] - }, - { - "family": "GORILLA", - "display_name": "Gorilla", - "description": "Gorilla value coding XORs consecutive floating-point bit patterns, emits a short code for repeated values, and reuses a previous significant-bit window when possible. The measured value kernel stores XOR control metadata and the remaining significant bits; the cited database design also describes timestamp coding.", - "reference_ids": [ - "M03" - ] - }, - { - "family": "GZIP", - "display_name": "gzip", - "description": "Gzip wraps a DEFLATE payload, whose LZ77 matches and literals are Huffman-coded, in a header and a CRC-32/original-size trailer. The wrapper supports concatenated members and stream-level metadata.", - "reference_ids": [ - "M27" - ] - }, - { - "family": "Gipfeli", - "display_name": "Gipfeli", - "description": "Gipfeli parses the input into LZ77 literal runs and distance/length matches. It separately models and entropy-codes literal data while packing match commands into a compact byte stream.", - "reference_ids": [ - "M33" - ] - }, - { - "family": "Huff0", - "display_name": "Huff0", - "description": "Huff0 builds a Huffman codebook and writes symbols as variable-length prefix codes. Its implementation divides coding work into multiple streams to expose parallel instruction execution during decoding.", - "reference_ids": [ - "M49" - ] - }, - { - "family": "Huffman", - "display_name": "Huffman", - "description": "Huffman coding assigns shorter prefix codes to more frequent symbols and longer codes to rare symbols. A canonical representation records code lengths so the decoder can reconstruct the codebook compactly.", - "reference_ids": [ - "M50" - ] - }, - { - "family": "Kanzi", - "display_name": "Kanzi", - "description": "Kanzi is a block-compression framework whose preset selects a transform chain, such as Burrows-Wheeler, text, multimedia, or LZ transforms. The transformed symbols are then passed to a selected entropy coder such as ANS, Huffman, or context modeling.", - "reference_ids": [ - "M61" - ] - }, - { - "family": "LECO", - "display_name": "LECO", - "description": "LeCo fits a lightweight linear model to each block and stores model parameters plus prediction residuals. The project path ZigZag-maps and bit-packs those residuals, with exceptions carried by its block representation.", - "reference_ids": [ - "M16" - ] - }, - { - "family": "LZ4", - "display_name": "LZ4", - "description": "LZ4 encodes each block as sequences of literal bytes followed by an offset/length copy from already decoded data. The framed form wraps these LZ4 blocks with stream and block metadata and optional checksums.", - "reference_ids": [ - "M30" - ] - }, - { - "family": "LZ4 Fast", - "display_name": "LZ4 Fast", - "description": "LZ4 Fast writes the standard LZ4 literal-and-match sequence format. Its acceleration setting makes the greedy hash-table parser skip more candidate positions, changing match selection without changing the decoder format.", - "reference_ids": [ - "M64" - ] - }, - { - "family": "LZ4HC", - "display_name": "LZ4HC", - "description": "LZ4HC writes the same LZ4 block tokens but spends more search effort on match chains and parsing choices before choosing distance/length copies. Decoding remains the standard LZ4 literal-and-match reconstruction.", - "reference_ids": [ - "M64" - ] - }, - { - "family": "LZAV", - "display_name": "LZAV", - "description": "LZAV is an in-memory LZ77 design that finds repeated substrings in a dynamic window and represents them with distance/length references interleaved with literal runs. Its stream is token based and does not add a separate entropy-coding stage.", - "reference_ids": [ - "M65" - ] - }, - { - "family": "LZF", - "display_name": "LZF", - "description": "libLZF uses a hash-indexed sliding window to find LZ77 matches. The output alternates literal runs with compact length/distance back-references and does not apply a separate entropy coder.", - "reference_ids": [ - "M66" - ] - }, - { - "family": "LZFSE", - "display_name": "LZFSE", - "description": "LZFSE parses the input into Lempel-Ziv literals and matches, then entropy-codes the match and literal streams with Finite State Entropy tables. The decoder reconstructs data by decoding those streams and applying the recorded match copies.", - "reference_ids": [ - "M39" - ] - }, - { - "family": "LZG", - "display_name": "LZG", - "description": "libLZG uses LZ77 string matching and reserves marker bytes for copy commands. Repeated strings become marker, offset, and length forms, while other bytes are emitted as literals with marker escaping when required.", - "reference_ids": [ - "M67" - ] - }, - { - "family": "LZHAM", - "display_name": "LZHAM", - "description": "LZHAM uses a large sliding dictionary and LZ parsing to represent repeated strings as length/distance matches. Quasi-adaptive Huffman models encode literals, match classes, lengths, and distances in the bitstream.", - "reference_ids": [ - "M68" - ] - }, - { - "family": "LZJB", - "display_name": "LZJB", - "description": "LZJB hashes three-byte prefixes into a 1024-entry match table. A one-byte copy map distinguishes each group of eight literals or matches, and each match uses two bytes to store a 10-bit offset and a 6-bit length code.", - "reference_ids": [ - "M69" - ] - }, - { - "family": "LZMAT", - "display_name": "LZMAT", - "description": "LZMAT uses a hash-indexed dictionary to locate earlier byte sequences and emits either literals or distance/length copies. Its tag and copy fields are packed at nibble granularity, with variable-width forms for longer lengths and distances.", - "reference_ids": [ - "M35" - ] - }, - { - "family": "LZO1", - "display_name": "LZO1", - "description": "The LZO1 variants use LZ77-style dictionary matching and encode the parse as byte-aligned literal runs and distance/length match tokens. They omit a separate entropy-coding pass; the variants mainly differ in match finding and token details.", - "reference_ids": [ - "M36" - ] - }, - { - "family": "LZO1X", - "display_name": "LZO1X", - "description": "LZO1X uses the LZO byte-oriented LZ77 format, interleaving literal runs with distance/length copy commands. Its presets change compressor search effort and parsing while retaining the LZO1X decoder representation.", - "reference_ids": [ - "M36" - ], - "notes": "This Space family is a measured variant of PPT method LZO1 and reuses M36." - }, - { - "family": "LZO2A", - "display_name": "LZO2A", - "description": "LZO2A is an LZO-family LZ77 design that replaces repeated byte sequences with compact match tokens and writes unmatched bytes as literal runs. Its token layout is distinct from the LZO1 variants and does not rely on a separate entropy coder.", - "reference_ids": [ - "M36" - ] - }, - { - "family": "LZRW", - "display_name": "LZRW", - "description": "LZRW uses a hash table to find recent LZ77 matches and chooses between literal items and fixed-format copy items. Control words mark which following items are literals and which carry a match offset and length.", - "reference_ids": [ - "M37" - ] - }, - { - "family": "LZSSE", - "display_name": "LZSSE", - "description": "LZSSE packs literal-run and match-length decisions into four-bit control values and stores match offsets beside them; offsets and literals are XOR-coded against recent state. The LZSSE2, LZSSE4, and LZSSE8 formats use different control thresholds and parsing strategies while arranging decode work for SIMD execution.", - "reference_ids": [ - "M38" - ] - }, - { - "family": "LZVN", - "display_name": "LZVN", - "description": "LZVN is a compact Lempel-Ziv format that represents literal runs and distance/length matches with variable-size opcodes. The decoder maintains recent match state and applies copy commands directly to the output buffer.", - "reference_ids": [ - "M39" - ] - }, - { - "family": "Libdeflate", - "display_name": "Libdeflate", - "description": "libdeflate implements DEFLATE: LZ77 match finding converts repeated strings into length/distance pairs, and canonical Huffman codes encode those pairs together with literals. The same compressed payload can be emitted raw or inside zlib and gzip wrappers.", - "reference_ids": [ - "M62" - ] - }, - { - "family": "Lizard", - "display_name": "Lizard", - "description": "Lizard is derived from LZ4 and parses input into literals and LZ-style match commands using either fastLZ4 or the LIZv1 format. Higher method families add Huffman coding to the literal stream, so the measured range includes both token-only and entropy-coded modes.", - "reference_ids": [ - "M63" - ] - }, - { - "family": "Lzlib", - "display_name": "Lzlib", - "description": "Lzlib writes lzip members containing a simplified LZMA stream. A sliding dictionary produces literals, new matches, and repeated-distance matches; context-dependent probability models and a range encoder code those decisions, followed by CRC and size fields for the member.", - "reference_ids": [ - "M70" - ] - }, - { - "family": "PFPC", - "display_name": "PFPC", - "description": "pFPC partitions floating-point data for parallel processing and predicts bit patterns with FCM and differential FCM tables. It XORs each prediction with the actual value and stores the residual using a compact leading-zero-byte representation.", - "reference_ids": [ - "M21" - ] - }, - { - "family": "PPMD8", - "display_name": "PPMD8", - "description": "PPMd8 predicts each symbol from variable-order contexts, escapes to shorter contexts when needed, and sends the modeled probabilities through a range coder. Its memory setting controls how much context state and history the model can retain.", - "reference_ids": [ - "M71" - ] - }, - { - "family": "Pithy", - "display_name": "Pithy", - "description": "Pithy hashes recent byte sequences to find LZ-style repetitions, then represents the stream as literal runs and compact offset-length match tokens. Its search effort controls how many candidates are considered.", - "reference_ids": [ - "M40" - ] - }, - { - "family": "Pongo", - "display_name": "Pongo", - "description": "Pongo applies shortest-decimal-aware floating-point recoding, then XORs adjacent representations and encodes the significant region and zero-boundary metadata. The reconstruction metadata carried by the project path preserves the original represented values.", - "reference_ids": [ - "M02" - ] - }, - { - "family": "QuickLZ", - "display_name": "QuickLZ", - "description": "QuickLZ uses hash tables to find recent byte matches and emits literals or compact offset-length commands. Its levels vary the number of stored candidates and the match search performed before choosing a token.", - "reference_ids": [ - "M72" - ] - }, - { - "family": "REGER", - "display_name": "REGER", - "description": "REGER reorders timestamp-value pairs within a block to strengthen local regression, predicts the reordered sequence, and packs the resulting residuals with frame-of-reference and segmented bit widths. Model and order information support reconstruction of the original sequence.", - "reference_ids": [ - "M06" - ] - }, - { - "family": "RLBE", - "display_name": "RLBE", - "description": "Run-Length and Bit-Level Encoding combines compact bit fields with run coding for repeated patterns in a numeric stream. The measured project row follows the smart-meter encoding reference and retains the metadata needed to reconstruct those fields and runs.", - "reference_ids": [ - "M11" - ] - }, - { - "family": "RLE", - "display_name": "RLE", - "description": "Run-length encoding replaces consecutive equal values with a value and repeat count. The measured TsFile-style path combines repeated runs with bit-packed groups for values that do not form long runs.", - "reference_ids": [ - "M10" - ] - }, - { - "family": "SLZ", - "display_name": "SLZ", - "description": "SLZ is a low-state DEFLATE encoder that finds LZ77 matches and writes fixed-Huffman blocks. This aggregate family covers the same core payload under raw DEFLATE, gzip, and zlib framing.", - "reference_ids": [ - "M73" - ] - }, - { - "family": "SNAPPY", - "display_name": "Snappy", - "description": "Snappy uses a hash-based LZ77 match finder and emits literal runs or compact offset-length copy commands without a separate entropy coder. Its framed format divides the stream into chunks and adds checksums and chunk metadata.", - "reference_ids": [ - "M47" - ] - }, - { - "family": "SPDP", - "display_name": "SPDP", - "description": "SPDP first forms lag-2 modular differences on floating-point words, transposes and differences their byte planes, and then replaces repeated history-conditioned byte sequences with compact repeat counts. These reversible stages target regularity in both numeric words and their byte layout.", - "reference_ids": [ - "M22" - ] - }, - { - "family": "SPRINTZ", - "display_name": "Sprintz", - "description": "Sprintz predicts numeric samples, maps signed residuals to small unsigned integers, and bit-packs small groups at widths selected from their residuals. Its design also compresses runs of zero-residual groups to exploit predictable stretches.", - "reference_ids": [ - "M12" - ] - }, - { - "family": "Shrinker", - "display_name": "Shrinker", - "description": "Shrinker is a compact LZ-family codec that replaces repeated byte strings with short match commands and leaves unmatched data as literals. The benchmark measures the codec's native token stream as exposed by the archived implementation.", - "reference_ids": [ - "M41" - ] - }, - { - "family": "Simple8b", - "display_name": "Simple8b", - "description": "Simple8b packs multiple nonnegative integers into one 64-bit word, reserving a selector field to choose the count and bit width of the values. Small integers fit more densely, and special selector modes compact long sequences of small repeated values.", - "reference_ids": [ - "M09" - ] - }, - { - "family": "Slz_deflate", - "display_name": "SLZ raw DEFLATE", - "description": "This SLZ variant writes the library's LZ77 matches and literals as fixed-Huffman DEFLATE blocks without an outer gzip or zlib wrapper.", - "reference_ids": [ - "M73" - ], - "notes": "This Space family splits the raw-DEFLATE wrapper from PPT method SLZ." - }, - { - "family": "Slz_gzip", - "display_name": "SLZ gzip", - "description": "This SLZ variant wraps its fixed-Huffman DEFLATE payload in a gzip member with the gzip header, CRC-32, and original-size trailer.", - "reference_ids": [ - "M73", - "M27" - ], - "notes": "This Space family splits the gzip wrapper from PPT method SLZ; M27 documents the wrapper format." - }, - { - "family": "Slz_zlib", - "display_name": "SLZ zlib", - "description": "This SLZ variant wraps its fixed-Huffman DEFLATE payload in a zlib stream with the zlib header and Adler-32 integrity value.", - "reference_ids": [ - "M73", - "M75" - ], - "notes": "This Space family splits the zlib wrapper from PPT method SLZ; M75 documents the zlib implementation family." - }, - { - "family": "Stream VByte", - "display_name": "Stream VByte", - "description": "Stream VByte encodes small integers in one to four data bytes and keeps two-bit length controls in a separate control stream, enabling SIMD decoding. The measured path first takes deltas so smooth integer columns produce smaller unsigned values.", - "reference_ids": [ - "M07" - ] - }, - { - "family": "Sub-column", - "display_name": "Sub-column", - "description": "This project kernel scales floating-point inputs when required, partitions integers into subcolumns, and selects bit-packed, run-length, or dictionary bodies for each block. Its block metadata identifies the partitioning and coding choices for reconstruction.", - "reference_ids": [ - "M09", - "M10", - "M13" - ], - "notes": "This Space-only project composition uses component ideas documented by M09, M10, and M13; none of those references specifies the complete implementation." - }, - { - "family": "TAMP", - "display_name": "TAMP", - "description": "TAMP is a small-memory, DEFLATE-inspired LZ codec that searches a configurable plaintext window and emits literal or match tokens in a compact framed stream. The window setting changes the available match history and token cost.", - "reference_ids": [ - "M79" - ] - }, - { - "family": "TS_2DIFF", - "display_name": "TS_2DIFF", - "description": "The measured TsFile TS_2DIFF kernel stores first differences in blocks, subtracts the minimum difference in each block, and bit-packs the nonnegative offsets. It retains the first value and block metadata; floating-point inputs use the configured decimal scaling before integer encoding.", - "reference_ids": [ - "M04" - ] - }, - { - "family": "Tornado", - "display_name": "Tornado", - "description": "Tornado combines LZ77 match finding with greedy, lazy, or optimal parsing and encodes the selected literals and matches with a byte, bit, Huffman, or arithmetic back end. The measured presets select different search structures, window sizes, parsers, and coders.", - "reference_ids": [ - "M74" - ] - }, - { - "family": "UCL NRV", - "display_name": "UCL NRV", - "description": "UCL NRV2D and NRV2E find repeated byte strings and replace them with LZ-style distance-length references. Compact bit-coded control, offset, and length fields describe the matches while unmatched bytes remain literals.", - "reference_ids": [ - "M42" - ] - }, - { - "family": "UCL NRV2b", - "display_name": "UCL NRV2B", - "description": "UCL NRV2B is a distinct NRV token format that combines LZ-style distance-length matches with bit-coded match fields and literal bytes. Its levels change compressor search effort while using the same NRV2B decoder.", - "reference_ids": [ - "M42" - ], - "notes": "This Space family is a measured variant of PPT method NRV and reuses M42." - }, - { - "family": "WFLZ", - "display_name": "WFLZ", - "description": "wfLZ finds repeated byte sequences in a bounded work-memory dictionary and writes literals and offset-length matches in a compact LZ stream. Its chunked form can split input into independently decodable blocks.", - "reference_ids": [ - "M43" - ] - }, - { - "family": "XPACK", - "display_name": "XPACK", - "description": "XPACK parses a byte stream into literals and LZ-style matches, selecting among candidates before writing codec-specific match tokens. The benchmark uses the experimental reference implementation cited in the PPT notes.", - "reference_ids": [ - "M44" - ] - }, - { - "family": "XZ", - "display_name": "XZ", - "description": "XZ uses LZMA2 dictionary matching with probability-modeled literal, length, and distance symbols encoded by a range coder. The XZ container adds stream framing, checks, and independently controlled LZMA2 chunks.", - "reference_ids": [ - "M31" - ] - }, - { - "family": "YALZ77", - "display_name": "YALZ77", - "description": "YALZ77 is a compact LZ77 implementation that searches a sliding history for repeated strings and emits streamable literal and offset-length records. Its search setting trades additional match exploration for a potentially smaller token stream.", - "reference_ids": [ - "M45" - ] - }, - { - "family": "Yappy", - "display_name": "Yappy", - "description": "Yappy hashes four-byte prefixes into a bounded recent-history table, searches a level-dependent number of candidates, and emits one-byte literal-run markers or two-byte match commands. Match lengths and offsets are decoded through fixed lookup tables.", - "reference_ids": [ - "M46" - ] - }, - { - "family": "ZSTD", - "display_name": "Zstandard", - "description": "Zstandard parses data into literals and LZ77 sequences with repeat-offset support. It Huffman-codes literal bytes, FSE-codes match-related symbol streams, and packages the result in independently described frame blocks.", - "reference_ids": [ - "M29" - ] - }, - { - "family": "ZXC", - "display_name": "ZXC", - "description": "ZXC performs block-level LZ77 parsing, represents literals and matches with compact integer/token layouts, and can Huffman-code selected sections. Its implementation selects scalar or vectorized paths without changing the lossless format semantics.", - "reference_ids": [ - "M78" - ] - }, - { - "family": "Zlib", - "display_name": "Zlib", - "description": "Zlib implements DEFLATE by replacing repeated strings with LZ77 length-distance pairs and coding literals, lengths, and distances with fixed or dynamic Huffman codes. A zlib wrapper supplies stream metadata and an Adler-32 check.", - "reference_ids": [ - "M75" - ] - }, - { - "family": "Zlib-ng", - "display_name": "Zlib-ng", - "description": "Zlib-ng produces zlib/DEFLATE-compatible streams using optimized match finding, hashing, Huffman construction, and architecture-specific routines. Its compression mechanism remains LZ77 followed by fixed or dynamic Huffman coding.", - "reference_ids": [ - "M76" - ] - }, - { - "family": "Zling", - "display_name": "Zling", - "description": "libzling processes data in blocks with a reduced-offset LZ (ROLZ) match transform, then Huffman-codes literals and match-related symbols. Block boundaries bound the state used by the match and entropy stages.", - "reference_ids": [ - "M77" - ] - } - ], - "operators": [ - { - "id": "eraser", - "family_number": "01", - "name": "Eraser / 擦除", - "group": "Bit processing", - "description": "Preprocesses floating-point bit patterns by clearing selected mantissa suffix bits and encoding the information needed to reconstruct them. Whether a concrete pipeline is lossless depends on its erasure rule and stored metadata.", - "reference_ids": [ - "O01" - ], - "implementations": [ - { - "id": "A001", - "name": "Fixed Low-4-bit Split Proxy", - "layer": "configuration", - "description": "对每个int64 x计算high=x>>4和low4=x&15；逆向按64位位型计算(high<<4)|low4。它固定拆4位，不检查浮点指数、尾数或十进制可逆性。", - "inputs": "1流values:int64[N]", - "outputs": "2流high:int64[N]、low4:uint8语义[N]；low_bits固定4，可参数化为通用低位拆分。", - "example": "输入 [33,34,-1]，得到 high=[2,2,-1]、low4=[1,2,15]。例如 (2<<4)|1=33；-1 按64位原位恢复。", - "status": "当前项目参考/代理或封装步骤；按下面实际计算解释，不等同名称暗示的上游算法。" - }, - { - "id": "A002", - "name": "Pongo Shortest-decimal Mantissa Recoding", - "layer": "mechanism", - "description": "对有限非零double求最短十进制小数，提取前导零数与后续十进制整数；若其位数能放入当前指数留下的尾数字段，则反转低位写入该字段并逐位验证恢复，否则保留原字。", - "inputs": "1流64位原字bytes:uint8[8N]，逻辑可为int64或float64", - "outputs": "3流first:uint64[0/1]、eraser_flags:uint8[N-1]、targets:uint64[N-1]；无外部参数。", - "example": "输入原字序列 [+0.0,+0.0]：首字为0，后续零值不做十进制改写，target仍为0，擦除标记表示原字保留。此时后接相邻XOR得到0残差。非零值仅在逆十进制重编码逐位相等时采用改写。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A003", - "name": "Elf Decimal Significant-digit Mantissa Erasure", - "layer": "mechanism", - "description": "从浮点十进制有效位推导alpha与beta，据指数计算可擦除的低尾数位；仅当掩码后的原字配合beta能逐位恢复且收益达到阈值时输出擦除字，否则发literal，并按块重置beta状态。", - "inputs": "float64[N]原位；P003要求可精确窄化为float32，P005处理binary64。", - "outputs": "3流tag:uint8[N]、beta:uint8[N]、erased_bits:uint32或uint64[N]；word_bits、block_size、threshold为profile参数。", - "example": "输入 [-0.0,+0.0] 走 literal 分支，输出对应 tag/beta 与 erased_bits=[0x8000000000000000,0x0000000000000000]（64位profile）。解码恢复两种符号零；beta没有被用来近似替代原位。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A004", - "name": "Camel Integer-plus-short-decimal Decomposition", - "layer": "mechanism", - "description": "每1000项首值直接保存；其余有限值截取int64整数部分，检测十进制位数并最多保留3位小数及符号，按Java式round重建并逐位核验；失败则保存literal原字。", - "inputs": "1流float64[N]或可精确转float64的int64[N]", - "outputs": "6流tag:uint8[N]、raw_bits:uint64[R]、integer:int64[D]、signal:uint8[D]、decimal_count:uint8[D]、decimal_value:uint32[D]；chunk=1000、max_decimal_digits=3。", - "example": "一块中先存首值100.0；下一值100.25可表示为 integer=100、decimal_count=2、decimal_value=25及正号。解码以100+25/100恢复，并逐位检查；无法精确恢复的记录改用literal。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - } - ] - }, - { - "id": "scaling", - "family_number": "02", - "name": "Scaling / 数值缩放", - "group": "Numeric representation", - "description": "Maps decimal-valued floating-point data to integers by multiplying by a power of ten and applying a rounding rule. Exact restoration requires the scale and a conversion rule that preserves every source value used by the pipeline.", - "reference_ids": [ - "O02" - ], - "implementations": [ - { - "id": "A006", - "name": "Decimal Scaling with Delta-chain Constraint", - "layer": "mechanism", - "description": "选择阶段从max_point向下取s，使用double乘积与llround生成候选整数，直到相邻差满足int64范围；停止后再逐位检查逆缩放。当前convert在乘积越界时另有退回llround(x)/原位解释的兼容行为，不能笼统说仅“越界拒绝”。", - "inputs": "1流values:float64[N]及max_point:int32[0…18]", - "outputs": "2流scale:uint32[1]、values:int64[N]；selector=delta_chain_fit、rounding=llround。", - "example": "max_point=2，输入 [1.25,1.50,1.75]：候选code=[125,150,175]，相邻差25可装入int64，逆缩放精确，因此输出scale=2与该整数流。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A007", - "name": "Fixed Exact Decimal Scale", - "layer": "mechanism", - "description": "s=max_point直接作为输入；用long double乘积计算llround(x×10^s)，再用double(code)/10^s逐位验证；非有限、越界或不精确拒绝。TS_2DIFF与Simple8b的执行一致。", - "inputs": "1流values:float64[N]及s=max_point:int32[0…18]", - "outputs": "1流values:int64[N]，s作为输入参数/packet恢复字段而非新增活动流；rounding=llround。", - "example": "给定s=2，输入 [1.25,-2.50]，输出 [125,-250]；解码除以100得到原值。s必须作为恢复参数保存。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A008", - "name": "Largest-fitting Decimal Scale Selector", - "layer": "mechanism", - "description": "从max_point向0搜索首个使所有x×10^s落入int64的s，按llround生成整数；找到首个fit尺度后才做逐位逆检查，若不精确则立即拒绝，不继续尝试更小s。", - "inputs": "1流values:float64[N]及max_point:int32[0…18]", - "outputs": "2流effective_scale:int32[1]、values:int64[N]；selector=largest_fitting_then_exact_check。", - "example": "max_point=2，输入 [1.25,2.50]，最高候选不越界且逆变换精确，输出scale=2、codes=[125,250]；若该候选逆变换不精确，当前profile拒绝输入。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A009", - "name": "Recipe: Required Precision → Decimal Scale", - "layer": "mechanism", - "description": "逐值从0至15找最少可逆小数位，取整组最大值s，再按共同s重新量化并逐位复核。选择阶段乘积为double，量化阶段为long double；P010分1024项，P011整列。失败整块/整列保存raw，含负零也可触发raw。", - "inputs": "values:float64[N]；block_size=1024 或 whole。", - "outputs": "scaled 路径输出 precision:uint8 与 scaled_int:int64；raw 路径输出 raw_bits:uint64。P010 可混合 3 流，P011 整列为 scaled 2 流或 raw 1 流。", - "example": "输入 [1.2,1.23,1.234]，最少所需小数位为[1,2,3]，整组取s=3，得到[1200,1230,1234]。块mode区分scaled与整块raw。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A010", - "name": "Code-equivalent Decimal Scale Search", - "layer": "mechanism", - "description": "float64先从max_point下降选择一组能装入int64的codes，再从max(18,max_point)下降，找可产生同一codes且按wire尺度逐位恢复原输入的s。C++显式实现Python ties-to-even语义；int64直接透传。", - "inputs": "1流int64[N]或float64[N]原字及kind:string；float需max_point", - "outputs": "effective_scale:int32[1]+codes:int64[N]共2流，int64仅codes 1流；rounding=ties_to_even。", - "example": "整数profile输入 [10,11,12]，仅输出同一code流；浮点profile输入[1.25,2.5]，候选s=2可得到[125,250]，另保存实际恢复scale，解码逐位校验。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A011", - "name": "Recipe: ALP Selector → Ordered Scale → Exception Split", - "layer": "mechanism", - "description": "每1024项枚举exp=10..0、fac=min(exp,9)..0，用估计位宽与异常成本选参数。数值执行x×10^exp→×10^-fac→magic舍入；再按逐位恢复是否相同分离异常，正常code保留，异常另存位置及原位，无可用参数则整块raw。", - "inputs": "float64 values；selector 产生的 fac/exp:int32[2B]；当前入口仍整体执行，未暴露独立活动端口。", - "outputs": "Scaling 生成 codes:int64 及可逆性结果；异常分离生成带占位正常 codes、exception_index:uint32、exception_bits:uint64。scaled 块最终 4 个语义角色（含 fac/exp），raw 块 1 个角色；mode/count 为恢复信息。", - "example": "给定候选exp=2、fac=0，输入[1.25,2.50]，有序乘法及magic舍入得到[125,250]，两个位置都可恢复，无异常。若另有不能恢复的原字，该位置及原字进入异常表。参数选择仍比较全部候选成本。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A012", - "name": "Recipe: Mean/Sigma Estimate → Normalize → XOR Repair", - "layer": "mechanism", - "description": "先用有限值估计mean/sigma，再输出z=(x-mean)/sigma；异常z置0。由z逆预测原字，生成repair=original_bits XOR predicted_bits。解码依靠z、mean/sigma与repair才能逐位恢复。", - "inputs": "原始int64/float64位流。参数估计", - "outputs": "mean_sigma:float64[2]；Normalize输出z:float64[N]；Repair读取原始bits:uint64[N]与逆预测bits:uint64[N]，输出repair:uint64[N]。完整配方共3语义流mean_sigma、z、repair；原类型及N保留用于逆预测。", - "example": "输入[1.0,2.0,3.0]，均值为2，标准差按实现规则计算为sigma；输出z=[-1/sigma,0,1/sigma]与mean/sigma。每项还输出original_bits XOR inverse_prediction_bits；解码必须应用这条repair流。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - } - ] - }, - { - "id": "frame_of_reference", - "family_number": "03", - "name": "Frame of Reference (FOR) / 基准帧", - "group": "Numeric representation", - "description": "Stores a block reference value, commonly the minimum or another base, and represents each value as a smaller offset from that reference.", - "reference_ids": [ - "O03" - ], - "implementations": [ - { - "id": "A013", - "name": "Minimum Frame of Reference", - "layer": "mechanism", - "description": "求minimum=min(values)，逐项以同宽无符号算术计算offset=unsigned(value)-unsigned(minimum)。逆向将minimum加回并恢复原位。空输入以minimum=0配合N=0表示。", - "inputs": "values:T[N]，T为int32/uint32/int64/uint64。", - "outputs": "minimum:T[1]和offsets:unsigned(T)[N]。参数：字宽、块边界。首值、scale、排列、异常字段为旁路恢复元数据；本步不做Bit Packing。", - "example": "输入int64 [100,103,101]，输出minimum=100、offsets:uint64=[0,3,1]。逆向100+[0,3,1]恢复原序列。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - } - ] - }, - { - "id": "value_reorder", - "family_number": "04", - "name": "Value Reorder / 数值重排", - "group": "Reordering and symbol recoding", - "description": "Reorders values, or timestamp-value pairs, within a block to expose stronger local correlation and reduce later residual widths. A reversible pipeline must preserve enough ordering information to restore the original sequence.", - "reference_ids": [ - "O04" - ], - "implementations": [ - { - "id": "A014", - "name": "Stable Sort with Inverse Permutation", - "layer": "mechanism", - "description": "建立order=[0…N-1]，按values[order[j]]的signed int64值stable_sort；依order输出sorted_values并保存原位置，逆向把第j个排序值写回original_index[j]。", - "inputs": "1流values:int64[N]", - "outputs": "2流sorted_values:int64[N]、original_index:int64[N]或uint32[N]；参数block_size=whole或64、index_dtype/index_scope为I/O配置。", - "example": "输入[30,10,30,20]，稳定排序后values=[10,20,30,30]，original_index=[1,3,0,2]。按索引散写恢复[30,10,30,20]。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - } - ] - }, - { - "id": "regression_predictor", - "family_number": "05", - "name": "Regression Predictor / 回归预测", - "group": "Residual formation", - "description": "Predicts values or timestamp-value pairs from a fitted regression relation, then passes prediction residuals to later integer or bit-level coding stages.", - "reference_ids": [ - "O04" - ], - "implementations": [ - { - "id": "A015", - "name": "First-order Autoregression by OLS", - "layer": "mechanism", - "description": "以x_i=value[i]、y_i=value[i+1]对整列做OLS，求theta0、theta1；从首值开始用round(theta0+theta1×previous_value)预测下一值，保存模2^64有符号残差。", - "inputs": "1流values:int64[N]", - "outputs": "3流first:int64[0/1]、coefficients:float64[2]、residuals:int64[N-1]；无外部参数，rounding=nearbyint。", - "example": "输入[10,20,30,40]，相邻样本满足next=10+previous，OLS得到theta0=10、theta1=1；首值10单存，后三个残差全为0。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A016", - "name": "Blockwise Linear Regression against Position", - "layer": "mechanism", - "description": "每个512值块以块内位置i为自变量拟合intercept+slope×i；预测经llround并限制在int64域，逐值输出有符号残差。输出有符号残差，ZigZag由A046执行。", - "inputs": "1流values:int64[N]，参数block_size（当前LECO profile为512）。", - "outputs": "2流coefficients:float64[blocks,2]、residuals:int64[N]；块长度和原值个数是恢复元数据。", - "example": "一块[10,12,14,16]以位置[0,1,2,3]拟合得到intercept=10、slope=2，signed残差[0,0,0,0]；系数和块长随包保存，残差可再接A046。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - } - ] - }, - { - "id": "outlier_separation", - "family_number": "06", - "name": "Outlier Separation / 异常分离", - "group": "Numeric representation", - "description": "Splits exceptional values from a regular block so the common values can use a narrow bit width, while positions and payloads for outliers are encoded separately.", - "reference_ids": [ - "O05" - ], - "implementations": [ - { - "id": "A018", - "name": "IQR Clipping（四分位距裁剪）", - "layer": "mechanism", - "description": "计算Q1、Q3，令IQR=max(Q3-Q1,1)，边界为floor(Q1-1.5IQR)与ceil(Q3+1.5IQR)。越界值在主体中替换为边界，同时保存原位置与原值，解码再覆盖恢复。", - "inputs": "values:int64[N]", - "outputs": "3流clipped_values:int64[N]、outlier_index:int64[E]、outlier_original:int64[E]。倍率1.5和IQR下限1当前固定；无公开参数。", - "example": "在Q1=10、Q3=12的状态下，IQR=2，裁剪区间[7,15]；输入末项100会在主体变为15，并保存其位置与100。解码覆盖该位置恢复100。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A019", - "name": "BOS Cost-based Range Split（按成本选区间）", - "layer": "mechanism", - "description": "对上游FOR码排序，枚举宽度b和跨度不超过2^b-1的连续值域，比较K×b+96×E位的代价，选常规区间；按原序标注异常并保留原数值，本步尚不减区间基准。", - "inputs": "BF11 packet:uint8[L]，核心codes:uint64[N-1]", - "outputs": "7流normal_values:uint64、outlier_values:uint64、tags:uint8、range_base:uint64、bit_width:uint8及透传first:uint64、source_base:int64。", - "example": "上游码为[0,1,2,1000]。常规域[0,3]取b=2，成本3×2+96=102位；全域b=10成本40位，因此该小例选择全域、不拆异常。大规模常规值才可能摊薄96位异常开销。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A111", - "name": "Given-width Exception Separation", - "layer": "mechanism", - "description": "调用方给定b，对每个uint32值提取low=x & mask(b)，仅当high=x>>b非零时记录pos和high。b=32时low保留全值且异常为空。逆向严格验证位置、流长和有效位范围，按low[pos] OR (high<>2)×delta右移16位得到的校正；以预测误差符号更新accumulator，再更新delta。每9值组重置。", - "inputs": "int64[N]或float64原位字", - "outputs": "3流anchors:int64原字[B]、residuals:int64模残差[8B]、raw_tail:int64原字[T]，B=floor(N/9)，T=N%9。当前移位/学习规则固定。", - "example": "组起点previous=100、delta=0、accumulator=0，下一值103的预测为100、误差为3；观察误差符号更新accumulator，再令delta=3。后续预测使用更新后的状态。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A030", - "name": "FPZIP Mapped Float Residual（浮点映射残差）", - "layer": "mechanism", - "description": "把double原位映射为可逆的单调uint64码；按固定FPZIP一维浮点表达式从previous产生预测，再做mapped_current-mapped_prediction。初始预测为+0，解码加回并逆映射。", - "inputs": "float64[N]的8N字节原位，N≤UINT32_MAX", - "outputs": "1流mapped_residuals:uint64[N]。无公开参数；须保留运算顺序与初始+0。源码未拒绝NaN/Inf，但因预测执行double运算，特殊值位级可逆性需单独round-trip验证。", - "example": "若当前位置的原位与预测原位相同，则两者经同一单调映射得到同一码，mapped residual=0。解码把0加回预测码并逆映射，恢复相同float64位型。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - } - ] - }, - { - "id": "xor", - "family_number": "09", - "name": "XOR / 异或残差", - "group": "Numeric representation", - "description": "XORs adjacent fixed-width values or floating-point bit patterns so shared bits become zeros and only changing bit regions need further coding.", - "reference_ids": [ - "O08" - ], - "implementations": [ - { - "id": "A031", - "name": "Adjacent Previous-word XOR", - "layer": "mechanism", - "description": "对每个32/64位原字计算residual=current XOR previous；首值作为anchor保留。逆操作同样是residual XOR previous。状态由调用方按原始记录顺序和chunk边界更新，不做浮点数值转换。", - "inputs": "current:uint32/uint64[N]，previous来自lag=1历史或显式前字。", - "outputs": "anchor与同宽residual流；标签、beta及块计数旁路。Camel的literal路径与decimal路径共用原记录前值状态，不能先分类型过滤再计算相邻关系。", - "example": "原字[0x3ff0000000000000,0x3ff0000000000001,0x3ff0000000000001]，首字单存，残差[1,0]。逐项与前一恢复字异或，得到完全相同的原位。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A032", - "name": "Chimp128 History-reference XOR（历史引用异或）", - "layer": "mechanism", - "description": "用当前字低14位查最近同key的位置；若该候选仍在128项历史内且候选XOR尾零数>13，使用该历史槽，否则用前一字。输出所选引用及XOR，再更新环形历史和索引。", - "inputs": "int64/float64原字；纯参考", - "outputs": "first、references:int64、xor_residuals:uint64共3流；实际Chimp128包还存trailing_zeros:uint8共4流。128槽、14位key、阈值13当前固定。", - "example": "历史槽7存A，当前值再次为A且低14位key命中槽7，XOR=0、尾零数64，满足引用条件。输出reference=7和residual=0；解码用槽7原字恢复A，再更新环形历史。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A033", - "name": "FCM/DFCM Predictor Selection（双预测器选择）", - "layer": "mechanism", - "description": "FCM按值历史hash预测原字，DFCM按差分历史hash预测stride，再加last得到第二候选；比较两种XOR残差的无符号大小，记录较小者和selector，随后同时更新两个表。", - "inputs": "非空8字节对齐原字bytes:uint8[8N]", - "outputs": "2流predictor_selector:uint8[N]、xor_residuals:uint64[N]。公开lg2固定16；pFPC每2^20值重置，底层其他lg2不是已开放图参数。", - "example": "假定本步FCM预测0x1000、DFCM预测0x1010，真实原字0x1011；候选残差为0x11与0x01，选择DFCM并记录selector和0x01。解码先预测，再用selector选预测并XOR残差。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - } - ] - }, - { - "id": "trailing_zero_pack", - "family_number": "10", - "name": "Trailing-zero Packing / 零位与残差打包", - "group": "Bit processing", - "description": "Represents an XOR result using its significant bits plus leading/trailing-zero metadata, allowing zero suffixes and other unchanged regions to be omitted from the payload.", - "reference_ids": [ - "O09" - ], - "implementations": [ - { - "id": "A035", - "name": "Elf+ Residual Window Packing", - "layer": "mechanism", - "description": "消费已有残差。零残差写01；若分桶leading与已保存leading相等且trailing≥已保存trailing，则写00与已保存窗口宽度的载荷；否则写10/11短长窗口头，包含3位leading桶和有效位数，载荷省去已知为1的最低有效位。逆向重建窗口及残差，不隐含XOR。", - "inputs": "residual:uint32/uint64[N]及窗口状态。", - "outputs": "控制/窗口头与有效位组成的bitstream，恢复N/profile/有效长度。32位桶[0,6,10,12,14,16,18,20]，短长长度字段3/5位；64位桶[0,8,12,16,18,20,22,24]，短长字段4/6位。方法适配器单独处理首字、EOS、eraser元数据和A031。", - "example": "64位profile，残差1的leading桶=24、trailing=0，significant=40：写长头11、leading码111、长度101000，以及省略最低1后的39位0。随后残差3可复用窗口，写00及40位载荷；零残差写01。", - "status": "共享C++内核与独立residual_window_pack_cli可用；原方法适配器复用该内核。独立图窗口动作尚未注册。" - }, - { - "id": "A037", - "name": "Common Trailing-zero Shift（共同尾零移位）", - "layer": "mechanism", - "description": "对所有非零int64取最小尾零数，再与16取小得到shift；每项算术右移shift。空列或全零列shift=0，逆向利用同一个shift恢复各原字。", - "inputs": "values:int64[N]", - "outputs": "2流values:int64[N]与shift:int64[1]。上限16当前固定，可作参数；并非float数值缩放。", - "example": "输入[16,48,80]，共同尾零最小值4，输出shift=4、values=[1,3,5]。逆向同宽左移4位恢复。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A038", - "name": "Zero/Reuse/New Window Packing（零值/复用/新窗口）", - "layer": "mechanism", - "description": "读取已生成XOR：零残差只写zero标记；非零的leading和trailing都不少于已保存窗口时复用已保存窗口，否则写新leading、有效位宽与有效位，并更新窗口。", - "inputs": "GXI1/GXF1或PXR1 packet；Gorilla", - "outputs": "4流first:uint64、window_tags:uint8、window_parameters:uint8、significant_payload:bytes；Pongo多eraser_flags:uint8共5流。", - "example": "已有窗口leading=30、trailing=30，残差0xc0000000的真实leading=32、trailing=30，满足复用条件；只需复用标记及保存窗口宽4位中的0011。残差0只写zero标记。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A039", - "name": "Parameterized Chimp Residual Window Packing", - "layer": "mechanism", - "description": "消费已有XOR与可选reference。四分支：00零残差，01省略尾零并写leading/有效位数，10复用leading，11写新leading。00/01按profile保存引���槽。零值或省尾零分支使leading复用状态失效；逆向读控制字段恢复原残差。", - "inputs": "residual:uint32/uint64[N]，可选reference:uint8[N]，trailing及leading桶信息。", - "outputs": "控制位、可选历史索引、窗口字段与有效位流。word_bits=32/64，history_index_bits=0/7，长度字段5/6位，leading桶和尾零阈值固定在profile中。独立CLI用N终止；方法适配器负责首字及EOS。", - "example": "64位、history_index_bits=0、threshold=6，残差1：leading桶24，新头11|111后写40位载荷1；下一残差3复用leading，写10与40位载荷3。零残差写00。索引位数7的profile在需要时另写7位历史槽。", - "status": "共享C++内核与独立residual_window_pack_cli可用；原方法适配器复用该内核。独立图窗口动作尚未注册。" - }, - { - "id": "A040", - "name": "FPC Nibble/Byte Packing（半字控制与整字节残差）", - "layer": "mechanism", - "description": "按XOR残差落入的字节范围编码为0…7，实际字节数为[0,1,2,3,5,6,7,8]；把此3-bit码与预测器selector合为nibble，每两项合一控制字节，再写低位残差字节。", - "inputs": "FPR1/PFR1预测包，记录(selector:uint8,residual:uint64)", - "outputs": "2流control_nibbles:bytes、residual_payload:bytes。预测器选择来自上游，本步不重新预测。", - "example": "残差0用byte-count码0；残差0x1234需2字节，用码2。若selector分别为0和1，两个nibble分别为0和10，合成control=0x0a，再写第二项两字节残差（顺序按profile）。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - } - ] - }, - { - "id": "dictionary", - "family_number": "11", - "name": "Dictionary / 字典编码", - "group": "Match and repetition", - "description": "Builds a set of distinct values or symbols and replaces repeated occurrences with compact dictionary identifiers; the dictionary is retained for decoding.", - "reference_ids": [ - "O10" - ], - "implementations": [ - { - "id": "A041", - "name": "Value-sorted Integer Dictionary Recoding", - "layer": "configuration", - "description": "统计int64唯一值，但字典直接保留std::map的数值升序；令dictionary[j]为第j小值，再逐项输出ids[i]=dictionary中values[i]的下标，解码按下标查表。", - "inputs": "1流values:int64[N]", - "outputs": "2流dictionary:int64[D]、ids:int64[N]；无参数。", - "example": "输入[30,10,30,20]，字典按值升序为[10,20,30]，ID=[2,0,2,1]；查表逐项恢复原值。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A042", - "name": "First-seen 64-bit-word Dictionary Recoding", - "layer": "mechanism", - "description": "把每个8字节原字读成uint64；从左到右遇到新原字即追加到dictionary并记录下标，重复原字复用既有下标，解码用uint64下标恢复原始64位字。", - "inputs": "1流64位原字，可标为int64[N]或float64[N]", - "outputs": "2流dictionary:同输入原字[D]、ids:uint64[N]；无算法参数。", - "example": "输入原字[A,B,A,C]，首次出现字典[A,B,C]，ID=[0,1,0,2]。即使两个NaN的payload不同，也作为不同原字处理。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A043", - "name": "First-seen Dictionary plus Fixed-width ID Packing", - "layer": "composition", - "description": "先按A042相同的首次出现规则建立float64原位字典和ID；再取b=bit_width(D-1)，把每个ID恰用b位连续写入位流，逆向按相同b读ID并查字典。", - "inputs": "1流values:float64[N]原位", - "outputs": "3流dictionary:float64[D]、id_bit_width:uint8[1]、packed_ids:bytes[ceil(Nb/8)]；b由D生成。", - "example": "输入[A,B,A,C]，字典[A,B,C]、D=3，b=2；ID位串00|01|00|10，连同字典和N保存。解码按2位读ID并查表。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A044", - "name": "Fixed 12-word Greedy Substitution", - "layer": "mechanism", - "description": "从左到右扫描bytes，在固定12词表中按表序寻找当前位置可匹配的首个词；命中则写tag=1和词ID并前进词长，否则写tag=0和一个literal字节。", - "inputs": "1流payload:uint8[N]", - "outputs": "3流kind:uint8[T]、literals:uint8[L]、dictionary_ids:uint8[M]；当前词表固定，无需输出。", - "example": "对于��匹配词表的单字节0xff，输出literal tag=0及0xff；若当前位置匹配词表第j项，则输出tag=1和j，解码展开该固定词条。词表及顺序是算法参数的一部分。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A045", - "name": "CORAD Tricklet XOR Atom Selection", - "layer": "mechanism", - "description": "将64位字按40-lane块组织，最多取4个40-lane atom作字典；每块对每个atom计算各lane异或残差位宽之和，选代价最小atom，输出selector与逐lane XOR残差。", - "inputs": "CNM2 packet:uint8[L]，含float64归一化字及repair", - "outputs": "5流parameters:float64[2]、dictionary:float64[40A]、selectors:uint8[B]、xor_residuals:uint64[N]、repair:uint64[N]；当前width=40、A≤4。", - "example": "当前40-lane块与候选atom0逐lane完全相同，则40个XOR残差均为0，成本0；选择atom0并保存selector。解码从atom0逐lane异或0恢复该块。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - } - ] - }, - { - "id": "zigzag", - "family_number": "12", - "name": "ZigZag / 之字形映射", - "group": "Numeric representation", - "description": "Maps signed integers to unsigned integers by interleaving nonnegative and negative values, keeping small magnitudes small before varint or bit-packed coding.", - "reference_ids": [ - "O11" - ], - "implementations": [ - { - "id": "A046", - "name": "Signed64 ZigZag Mapping（有符号到无符号映射）", - "layer": "mechanism", - "description": "对选中的int64残差逐值计算ZigZag双射，输出uint64码；逆向恢复原signed64位型。reference、Sprintz、DBP、LECO、REGER与BOS64只选择不同位置并旁路各自元数据。", - "inputs": "输入selected residuals:int64[N]，参数selected positions/I/O profile；A046新生成1流codes:uint64[N]。anchors、coefficients、indices、scale等输入流原样旁路。", - "outputs": "输入selected residuals:int64[N]，参数selected positions/I/O profile；A046新生成1流codes:uint64[N]。anchors、coefficients、indices、scale等输入流原样旁路。", - "example": "输入signed [-2,-1,0,1,2]，输出unsigned [3,1,0,2,4]；负数映为奇数，非负数映为偶数。INT64_MIN对应UINT64_MAX。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - } - ] - }, - { - "id": "block_transform", - "family_number": "13", - "name": "Block Transform / 块变换", - "group": "Reordering and symbol recoding", - "description": "Applies a reversible block permutation such as the Burrows-Wheeler transform to group similar symbol contexts before recoding and entropy coding.", - "reference_ids": [ - "O12" - ], - "implementations": [ - { - "id": "A014", - "name": "Stable Sort with Inverse Permutation", - "layer": "mechanism", - "description": "建立order=[0…N-1]，按values[order[j]]的signed int64值stable_sort；依order输出sorted_values并保存原位置，逆向把第j个排序值写回original_index[j]。", - "inputs": "1流values:int64[N]", - "outputs": "2流sorted_values:int64[N]、original_index:int64[N]或uint32[N]；参数block_size=whole或64、index_dtype/index_scope为I/O配置。", - "example": "输入[30,10,30,20]，稳定排序后values=[10,20,30,30]，original_index=[1,3,0,2]。按索引散写恢复[30,10,30,20]。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A047", - "name": "SPDP Byte-plane Transpose plus Cross-plane Delta", - "layer": "composition", - "description": "把每个64位lag-2残差的第0至7字节依次组成8个平面；按该平面串接顺序维护单个previous状态，输出d=(current-previous) mod256，previous随后更新为current。", - "inputs": "S2D1 packet:uint8[12+8N]，语义为lag2_deltas:uint64[N]", - "outputs": "8流plane0_deltas…plane7_deltas:uint8[N]；word_width固定8，previous初值0。", - "example": "两个uint64值[1,2]的小端字节列化后为[1,2,0,0,...]；以previous=0做连续字节Delta，前四字节变为[1,1,254,0]。逆向先前缀加，再逆转置。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A048", - "name": "Byte-plane Split", - "layer": "mechanism", - "description": "输入已有uint64值和有效位宽b，令S=ceil(b/8)，将有效位左对齐到S字节，按从高到低字节位置输出S条平面。��向拼接平面并去除低位padding，仅还原输入值。", - "inputs": "values:uint64[N]，b∈[0,64]，每值须可由b位表示。", - "outputs": "S条uint8[N]平面；N、b以及高低顺序用于恢复，空/零宽情况也保留N。minimum和scale由上游旁路保存，本步骤不计算minimum、不做减法。", - "example": "输入offsets=[0,3,1]、b=2，S=1，左移6位后输出uint8平面[0x00,0xc0,0x40]。逆向右移6位恢复[0,3,1]；base不参与本步。b=0时没有平面，但N仍保留。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A050", - "name": "Cyclic Burrows-Wheeler Transform", - "layer": "mechanism", - "description": "把整个字节串的N个循环旋转按字典序排序，输出每行旋转的末字节last_column，并记录原串所在行primary_index；逆向用字节计数和LF映射恢复原串。", - "inputs": "1流payload:uint8[N]", - "outputs": "2流last_column:uint8[N]、primary_index:uint64[1]；当前整个输入为单块，无算法参数。", - "example": "输入ASCII BANANA，排序循环旋转后last_column=NNBAAA、primary_index=3（0起点）。保留两者，LF映射可还原BANANA。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A051", - "name": "Low/High 32-bit Limb Decomposition", - "layer": "configuration", - "description": "对每个64位code计算low=uint32(code&0xffffffff)、high=uint32(code>>32)，分别按小端写入两条等长limb流；逆向按code=low|(uint64(high)<<32)逐项合成。", - "inputs": "codes:int64[N]或其LNS1块", - "outputs": "2流low32:uint32[N]、high32:uint32[N]；scaled/ALP仅增加scale、factor/exponent、exception等透传流。", - "example": "输入uint64 0x1122334455667788，low32=0x55667788、high32=0x11223344；两条limb流分别保存，逆向high<<32|low恢复原字。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A052", - "name": "Swap 32-bit Halves within Each 64-bit Word", - "layer": "mechanism", - "description": "对每个8字节原字交换前4字节与后4字节，半字内部小端顺序不变；该变换自逆，等价于先取得high32、low32，再按high32后low32写回单一路载荷。", - "inputs": "1流64位原字bytes:uint8[8N]，逻辑可为int64或float64", - "outputs": "1流reordered_words:uint8[8N]；layout=interleaved-high-low，无数值参数。", - "example": "输入字节88 77 66 55 44 33 22 11，输出44 33 22 11 88 77 66 55；再执行一次即恢复。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - } - ] - }, - { - "id": "symbol_recode", - "family_number": "14", - "name": "Symbol Recoding / 符号重编码", - "group": "Reordering and symbol recoding", - "description": "Re-expresses symbols after a transform, for example as move-to-front ranks, so recent or repeated symbols become small values and longer runs.", - "reference_ids": [ - "O13" - ], - "implementations": [ - { - "id": "A053", - "name": "Frequency Ordered Dictionary IDs", - "layer": "configuration", - "description": "先统计每个int64符号频数，再按频数降序stable_sort唯一符号；同频保持std::map数值升序，随后令ID为该排名，解码依赖完整dictionary查表。", - "inputs": "1流symbols:int64[N]", - "outputs": "2流dictionary:int64[D]、ids:int64[N]；无熵模型参数，排序规则固定。", - "example": "输入[8,3,8,5,3,8]，频数8:3、3:2、5:1，字典[8,3,5]，ID=[0,1,0,2,1,0]。输出仍是ID流，需要另接熵编码才能得到前缀码。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A054", - "name": "Move-to-Front Recoding", - "layer": "mechanism", - "description": "初始化table=[0,1,…,255]；对每个byte输出它在当前table中的rank，再把命中符号移到table[0]并保持其他符号相对次序；解码同步执行同一更新。", - "inputs": "1流symbols:uint8[N]", - "outputs": "1流ranks:uint8[N]；alphabet固定0…255，无输出表和算法参数。", - "example": "初始表[0,1,2,...]，输入[2,2,1]，rank依次为[2,0,2]；第一步表变为[2,0,1,...]，第三步取rank2得到1。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - } - ] - }, - { - "id": "run_length", - "family_number": "15", - "name": "Run Length / 游程编码", - "group": "Match and repetition", - "description": "Replaces consecutive equal values or symbols with a value-and-count representation, reducing long runs while adding overhead to rapidly changing data.", - "reference_ids": [ - "O14" - ], - "implementations": [ - { - "id": "A055", - "name": "Size-optimal Hybrid RLE/Bitpack/Raw Selection", - "layer": "mechanism", - "description": "在raw、constant、全流bitpack与hybrid中按精确编码字节数选最小；hybrid用DP在最长128项literal-bitpack段和任意等值repeat段间切分，同成本优先repeat再优先更早起点。", - "inputs": "C++核心输入bytes:uint8[N×symbol_width]，按unsigned固定宽原字比较，symbol_width=1…8；当前公开profile实际用1或8字节。", - "outputs": "按模式为1流raw、2流constant、2流bitpack或5流hybrid；scaled传64位raw codes，RLBE先生成位宽序列并旁路delta_bits。", - "example": "输入100个7，constant候选只需保存常量及长度；与raw、bitpack、hybrid的完整候选字节数比较，选择最短表示。解码由count=100重复该常量。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - } - ] - }, - { - "id": "lz_match", - "family_number": "16", - "name": "LZ Match / LZ匹配", - "group": "Match and repetition", - "description": "Searches previously decoded data for repeated byte strings and emits distance-length copy commands in place of duplicate literals, following the LZ77 family model.", - "reference_ids": [ - "O15" - ], - "implementations": [ - { - "id": "A056", - "name": "Threshold-four int64 RLE Compatibility Proxy", - "layer": "adapter", - "description": "从左到右寻找相邻相等int64的最大游程；长度至少4时输出repeat标记、run length和一个值，较短游程输出literal标记、长度和全部原值，不产生向后距离。", - "inputs": "1流values:int64[N]", - "outputs": "3流tag:int64[R]、count:int64[R]、values:int64[K]；repeat_threshold固定4，无offset。", - "example": "输入[5,5,5,5,2]，形成repeat(value=5,length=4)与literal([2],length=1)。解码按长度展开，未出现任何向后距离字段。", - "status": "当前项目参考/代理或封装步骤；按下面实际计算解释，不等同名称暗示的上游算法。" - }, - { - "id": "A057", - "name": "Attach Native-LZ Minimum/Maximum Configuration", - "layer": "configuration", - "description": "读取已有NLS1 role=1上下文，验证body仍等于raw_size，仅把role改为2并写入minimum与maximum；载荷逐字节不变，既不建立索引也不查找匹配。", - "inputs": "NLS1 packet:uint8[25+N]及minimum:uint32、maximum:uint32", - "outputs": "仍1流raw_body:uint8[N]并更新头字段，新增算法语义流0。", - "example": "输入NLS1上下文(window=32768,raw_size=8,body=ABCDEFGH)，设min=4、max=130后body仍为ABCDEFGH；输出只增加/更新约束字段，后续A060才查找匹配。", - "status": "当前项目参考/代理或封装步骤；按下面实际计算解释，不等同名称暗示的上游算法。" - }, - { - "id": "A058", - "name": "Reference Lazy Hash-chain LZ Search", - "layer": "mechanism", - "description": "以3字节key维护历史位置链，每处从最近位置反查至多64个候选，选window内最长匹配；若下一位置的最佳长度大于当前长度加1，则放弃当前match并先输出literal。", - "inputs": "1流payload:uint8[N]及window/min_match/max_match:uint32", - "outputs": "4流tag:uint8[T]、literal:uint8[L]、offset:uint32[M]、length:uint32[M]；search_policy=lazy。", - "example": "位置i最佳match长4，而i+1最佳match长7时，7>4+1，lazy规则先发出i处literal；随后再处理更长候选。保存literal/match序列及raw_size以供逆向复制。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A059", - "name": "Reference Greedy Hash-chain LZ Search", - "layer": "configuration", - "description": "使用与A058相同的3字节历史链和最多64候选最长匹配，但当前位置一旦找到长度达到min_match的最佳match便立即发出，不计算下一位置候选，也不执行lazy lookahead。", - "inputs": "1流payload:uint8[N]及window/min/max", - "outputs": "同A058的4条LZS1语义流；search_policy=greedy。", - "example": "同一候选状态中，位置i已有满足min_match的长4匹配，立即发出(distance,4)，不比较i+1的长7候选；逆向按距离从已恢复前缀复制4字节��", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A060", - "name": "Native Aligned-8 Greedy Last-occurrence LZ Search", - "layer": "mechanism", - "description": "以8字节key只保存最近一次位置；match长度先逐字节扩展，再向下取8的倍数，并受有效min/max约束；有match立即贪心发出，否则首位失败时批量发8个literal。", - "inputs": "P074输入LNS1双肢bytes，核心参数window=32768、min=5、max=175；每肢", - "outputs": "kind:uint8[T]、literal:uint8[L]、offset:uint32[M]、length:uint32[M]，外围scale/exception透传。", - "example": "输入ABCDEFGHABCDEFGH（16字节），前8字节作为literal，位置8命中key ABCDEFGH，发出offset=8、length=8。解码复制已恢复的前8字节。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A061", - "name": "Snappy snap-1.1.1 Hash-table Match Generation", - "layer": "mechanism", - "description": "按64KiB块建立乘法hash表，使用skip增量探测寻找4字节候选，命中后逐字节延长；长copy按64或60切片，当前位置附近hash表按Snappy规则更新。", - "inputs": "1流payload:uint8[N]", - "outputs": "5流tag:uint8[R]、literal_bytes:uint8[L]、literal_length:uint32[J]、copy_offset:uint32[M]、copy_length:uint32[M]；规则固定。", - "example": "输入中已存在ABCD且当前位置4字节hash命中它，匹配扩展到长度70时，copy按Snappy规则切成多段（每段最多64），再交A068形成tag。此例说明匹配记录，不指定未经运行的完整压缩字节。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - } - ] - }, - { - "id": "match_parse_policy", - "family_number": "17", - "name": "Match Parse Policy / 匹配解析策略", - "group": "LZ decisions and format", - "description": "Chooses among candidate matches and literals using policies such as greedy, lazy, optimal, or repeat-offset parsing, trading search cost against encoded size.", - "reference_ids": [ - "O16" - ], - "implementations": [ - { - "id": "A060", - "name": "Native Aligned-8 Greedy Last-occurrence LZ Search", - "layer": "mechanism", - "description": "以8字节key只保存最近一次位置；match长度先逐字节扩展，再向下取8的倍数，并受有效min/max约束；有match立即贪心发出，否则首位失败时批量发8个literal。", - "inputs": "P074输入LNS1双肢bytes，核心参数window=32768、min=5、max=175；每肢", - "outputs": "kind:uint8[T]、literal:uint8[L]、offset:uint32[M]、length:uint32[M]，外围scale/exception透传。", - "example": "输入ABCDEFGHABCDEFGH（16字节），前8字节作为literal，位置8命中key ABCDEFGH，发出offset=8、length=8。解码复制已恢复的前8字节。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A062", - "name": "Position-dependent Multiply/Add Proxy", - "layer": "adapter", - "description": "对第i个int64原位按模2^64计算y=3×uint64(x)+(i mod251)，再把同一64位型放回int64载体；逆向用3在模2^64下的乘法逆元恢复。", - "inputs": "1流values:int64[N]", - "outputs": "1流transformed_values:int64[N]；位置i隐式，乘数3和模项251固定。", - "example": "输入[10,10,10]，位置i=0,1,2时输出[30,31,32]；解码减去i mod251，再乘3的模2^64逆元，恢复三个10。", - "status": "当前项目参考/代理或封装步骤；按下面实际计算解释，不等同名称暗示的上游算法。" - }, - { - "id": "A063", - "name": "Validated Identity Wrapper for Generic LZ Packet", - "layer": "adapter", - "description": "解析并验证输入LZS1的raw_size与literal/match token合法性，然后在外层增加ZPP1 wrapper；逆向去掉wrapper并再次验证，token次序和字段值完全不变。", - "inputs": "4条已有语义流tag:uint8[T]、literal:uint8[L]、offset:uint32[M]、length:uint32[M]的LZS1 packet", - "outputs": "同4流透传；无参数。", - "example": "输入已有LZS1的literal A及match(offset=1,length=3)，验证可恢复AAAA后加ZPP1包头；内部两个token值与顺序保持一致，解码去头并验证。", - "status": "当前项目参考/代理或封装步骤；按下面实际计算解释，不等同名称暗示的上游算法。" - } - ] - }, - { - "id": "match_window_model", - "family_number": "18", - "name": "Match Window Model / 匹配窗口模型", - "group": "LZ decisions and format", - "description": "Defines how far back matches may refer and how candidates are indexed, bounded, or reset; larger windows can capture distant repetition at additional memory and search cost.", - "reference_ids": [ - "O15" - ], - "implementations": [ - { - "id": "A064", - "name": "Position-dependent XOR Proxy", - "layer": "adapter", - "description": "对第i个int64的64位原字计算y=uint64(x) XOR (uint64(i)×131)，逆向再次异或同一位置常量；每项独立，不维护任何跨位置历史状态。", - "inputs": "1流values:int64[N]", - "outputs": "1流transformed_values:int64[N]；位置i隐式，常数131固定；兼容block_state profile只改变入口名。", - "example": "输入[10,10,10]，位置常量为[0,131,262]，输出[10,137,268]；再次异或相同常量恢复[10,10,10]。", - "status": "当前项目参考/代理或封装步骤；按下面实际计算解释，不等同名称暗示的上游算法。" - }, - { - "id": "A065", - "name": "Attach Native-LZ Window Configuration", - "layer": "configuration", - "description": "把原始bytes装入NLS1 role=1 envelope并在头部记录window_bytes与raw_size；body逐字节不变，逆向只核对window及长度后取回body，不建立窗口数据结构。", - "inputs": "1流raw_body:uint8[N]及window_bytes:uint32", - "outputs": "1个NLS1 packet，语义payload仍1流且新增算法流0；min/max可在同一配置接口补齐。", - "example": "raw=ABCDEFGH、window_bytes=32768，输出NLS1(role=1,window=32768,raw_size=8,body=ABCDEFGH)。逆向取回8字节body。", - "status": "当前项目参考/代理或封装步骤；按下面实际计算解释，不等同名称暗示的上游算法。" - } - ] - }, - { - "id": "match_token_format", - "family_number": "19", - "name": "Match Token Format / 匹配令牌格式", - "group": "LZ decisions and format", - "description": "Serializes literal lengths, literal bytes, match offsets, and match lengths into codec-specific commands, such as the nibble-split token and extensions used by an LZ4 block.", - "reference_ids": [ - "O17" - ], - "implementations": [ - { - "id": "A066", - "name": "Nibble Split and Pack", - "layer": "composition", - "description": "对每个int64输出high=value>>4；每两个值取低4位，按(low_even<<4)|low_odd组成一个uint8，奇数尾项的低半字节补0；逆向结合high与低4位恢复。", - "inputs": "1流values:int64[N]", - "outputs": "2流low_pairs:uint8[ceil(N/2)]、highs:int64[N]；low_width固定4，跨类历史入口不增加参数。", - "example": "输入[0x21,0x32,0x43]，high=[2,3,4]，低半字节打包为[0x12,0x30]，N=3指明最后0是padding。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A067", - "name": "Native ctrl8 Literal/Offset16/Length8 Packing", - "layer": "mechanism", - "description": "每8个有序token先留一个control byte，第b位为1表示match；literal写1字节，match写offset:uint16和(length-bias):uint8，逆向按control逐token读取并同步展开校验。", - "inputs": "literal/match token；参数offset_bits=16、length_bits=8、length_bias∈{3,4,5}", - "outputs": "4流control:uint8[ceil(T/8)]、literal:uint8[L]、offset:uint16[M]、length_bias:uint8[M]。", - "example": "token序列为literal A、match(offset=1,length=4)，bias=4：control最低两位为10（二进制），写literal 0x41、offset16=1、length8=0。其余control位对应不存在的token，不增加数据。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A068", - "name": "Snappy Literal/COPY Tag Formation", - "layer": "mechanism", - "description": "按Snappy语法为literal长度选择1至5字节头，为copy按offset和length选择COPY_1、COPY_2或COPY_4 tag；部分长度与offset位嵌入tag，长copy已由matcher切片。", - "inputs": "5条Snappy语义流tag、literal_bytes、literal_length、copy_offset、copy_length", - "outputs": "物理tag body:bytes，文档仍按5个语义角色展示；规则固定。", - "example": "literal ABC长度3，用tag=(3-1)<<2=0x08，后接41 42 43；COPY_2(offset=3,length=4)用tag=((4-1)<<2)|2=0x0e，后接小端offset 03 00。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A069", - "name": "LZ4 Token Nibbles and Offset Formation", - "layer": "mechanism", - "description": "每个sequence以一字节token保存literal_length的高nibble和match_length-4的低nibble；达到15时追加扩展长度，随后写literal bytes、offset:uint16及match扩展。", - "inputs": "LZ4 sequence", - "outputs": "7个语义角色token:uint8[S]、literal_extension:uint32[S]、literal_length:uint32[S]、literal_bytes:uint8[L]、has_match:bool[S]、offset:uint16[M]、match_extension:uint32[M]；match bias=4。", - "example": "一个sequence有3字节literal ABC、match长度8和offset=3；token高半字节3、低半字节8-4=4，即0x34，后接ABC及offset小端03 00。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A070", - "name": "Generic ctrl8 Literal/Offset32/Length32 Packing", - "layer": "mechanism", - "description": "同样每8个token写一个control byte，第b位区分literal和match；literal写uint8，match依次写未偏置offset:uint32和length:uint32，逆向依control恢复原token序列。", - "inputs": "literal/match token；参数offset_bits=32、length_bits=32、length_bias=0", - "outputs": "4流control:uint8[ceil(T/8)]、literal:uint8[L]、offset:uint32[M]、length:uint32[M]。", - "example": "literal A、match(offset=1,length=4)同样用control最低位0、下一位1，但match字段为offset32=1与length32=4，均不使用length-bias。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - } - ] - }, - { - "id": "integer_code", - "family_number": "20", - "name": "Integer Code / 整数编码", - "group": "Bit processing", - "description": "Uses variable-length or byte-oriented codes for nonnegative integers such as residuals, lengths, offsets, and metadata, giving small integers shorter representations.", - "reference_ids": [ - "O18" - ], - "implementations": [ - { - "id": "A023", - "name": "Unsigned ULEB128（无符号Base-128变长码）", - "layer": "mechanism", - "description": "对每个uint32/uint64反复取低7位；非末字节置continuation bit=1，末字节置0。核心可编码整列、DBP已ZigZag残差位置或Snappy的单个raw_size。", - "inputs": "核心输入values:uint32[N]或uint64[N]，参数word_bits=32/64与I/O profile；A023新生成1流encoded:bytes。DBP adapter完整", - "outputs": "2种语义：first:int64[0/1]与encoded residuals:bytes；scaled profile另带scale:int32旁路恢复字段。", - "example": "输入[0,127,128,300]，ULEB128字节依次为00、7f、80 01、ac 02。解码去掉continuation位并按7位组拼接；signed输入先接A046。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A072", - "name": "Stream VByte Coding of uint32 Limbs", - "layer": "mechanism", - "description": "解析SVD1中的low_deltas与high_deltas两条uint32流；每4值用一个control byte的四个2-bit字段记录1至4字节长度，低位数据紧密写入各自data区。", - "inputs": "1个SVD1 packet:uint8[12+8N]，字段scale:int32、count:uint32、low_deltas/high_deltas:uint32[N]", - "outputs": "4流low/high_control:uint8[ceil(N/4)]与low/high_data:bytes，scaled另透传scale共5流；无新增参数。", - "example": "四个uint32 [1,255,256,65536]需字节数[1,1,2,3]，2-bit长度码[0,0,1,2]合成control=0x90，数据小端为01 ff 00 01 00 00 01。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A073", - "name": "VariableByte Tail Coding after FastPFor Prefix", - "layer": "mechanism", - "description": "解析PFS1中已经编码好的low/high pfor_words与raw_tail；每肢原样复制pfor_words，只把raw_tail:uint32[T]交给VariableByte，再把两段拼成公开word序列。", - "inputs": "1个PFS1 packet，含kind:uint8、scale:int32、count:uint64及每肢pfor_words:uint32[P]、raw_tail:uint32[T]；完整", - "outputs": "每肢4条PFor前缀语义流加1条tail_bytes，整数源至多10流，scaled再加scale；无新增参数。", - "example": "已有pfor_words=[W0,W1]与raw_tail=[1,128]：保留W0/W1不变，仅把尾部交VariableByte编码并追加；两肢各自记录长度，逆向分别恢复前缀和两个尾值。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A075", - "name": "LZ4 Extension-length Base-255 Coding", - "layer": "mechanism", - "description": "解析P086产生的LZT1 sequence；literal或match的长度nibble达到15时写对应extension，包括extra=0时必需的零终止字节。extension反复��255并减255，最后写0..254余数；token、literal字节和offset保持原语义。", - "inputs": "LZT1 packet，含token、literal_extra/literal_length、literal_bytes、has_match、offset、match_extra；完整", - "outputs": "压缩块5流tokens、literal_extension_bytes、literal_bytes、offset、match_extension_bytes，raw块1流raw_bytes，混合共6流；无参数。", - "example": "LZ4扩展长度为510，输出ff ff 00；长度17输出11。255重复字节后必须有一个小于255的终止余数。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - } - ] - }, - { - "id": "bitstream_pack", - "family_number": "21", - "name": "Bitstream Write / 位流写入", - "group": "Bit processing", - "description": "Writes flags, codewords, and variable-width fields into a contiguous bitstream so sub-byte fields do not consume whole bytes; DEFLATE is a canonical bit-packed format.", - "reference_ids": [ - "O19" - ], - "implementations": [ - { - "id": "A066", - "name": "Nibble Split and Pack", - "layer": "composition", - "description": "对每个int64输出high=value>>4；每两个值取低4位，按(low_even<<4)|low_odd组成一个uint8，奇数尾项的低半字节补0；逆向结合high与低4位恢复。", - "inputs": "1流values:int64[N]", - "outputs": "2流low_pairs:uint8[ceil(N/2)]、highs:int64[N]；low_width固定4，跨类历史入口不增加参数。", - "example": "输入[0x21,0x32,0x43]，high=[2,3,4]，低半字节打包为[0x12,0x30]，N=3指明最后0是padding。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A076", - "name": "Camel decimal/XOR record bitstream writer", - "layer": "mechanism", - "description": "以首个float64原始位型开块；后续记录写literal标记+64位值，或写decimal标记、整数差分分级码、十进制位数和值；每1000条形成带count/body_size的块。", - "inputs": "1路Camel XOR packet:bytes（记录含float64位型/decimal字段）", - "outputs": "1路Camel public wire:bytes。chunk_size固定1000；整数差分仅支持既定分级范围。", - "example": "输入first原位100.0，后续decimal记录integer_delta=0、decimal_count=2、decimal_value=25，写记录类型、分级整数码和小数字段。解码得到100.25；literal记录则携带其XOR残差的64位载荷。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - } - ] - }, - { - "id": "entropy_huffman", - "family_number": "22", - "name": "Huffman Coding / 霍夫曼编码", - "group": "Probability coding", - "description": "Builds a prefix code in which more frequent symbols receive shorter integral-bit codewords and every symbol remains independently decodable from the bitstream.", - "reference_ids": [ - "O20" - ], - "implementations": [ - { - "id": "A053", - "name": "Frequency Ordered Dictionary IDs", - "layer": "configuration", - "description": "先统计每个int64符号频数，再按频数降序stable_sort唯一符号；同频保持std::map数值升序，随后令ID为该排名，解码依赖完整dictionary查表。", - "inputs": "1流symbols:int64[N]", - "outputs": "2流dictionary:int64[D]、ids:int64[N]；无熵模型参数，排序规则固定。", - "example": "输入[8,3,8,5,3,8]，频数8:3、3:2、5:1，字典[8,3,5]，ID=[0,1,0,2,1,0]。输出仍是ID流，需要另接熵编码才能得到前缀码。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A077", - "name": "Project Byte Huffman", - "layer": "mechanism", - "description": "统计256个byte频率，按项目固定建树规则生成静态Huffman码；包内保存原长、频率/树恢复信息和有效bit数，解码逐bit沿左右子树直到叶节点。", - "inputs": "1路raw:uint8[L]", - "outputs": "1路Huffman packet:bytes（头部模型/原长/bit数与payload同包）。当前无公开树策略参数。", - "example": "字节AAAA BBB C（频数4、3、1）建立静态Huffman树；高频A获得更短码。若示意树取A=0、B=10、C=11，则ABAC为010011；实际码字由固定建树平局规则决定，模型与有效bit数必须入包。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - } - ] - }, - { - "id": "entropy_range", - "family_number": "23", - "name": "Range Coding / 区间编码", - "group": "Probability coding", - "description": "Encodes a symbol sequence by successively narrowing a numeric interval according to model probabilities, approaching fractional-bit entropy costs through arithmetic state updates.", - "reference_ids": [ - "O21" - ], - "implementations": [ - { - "id": "A053", - "name": "Frequency Ordered Dictionary IDs", - "layer": "configuration", - "description": "先统计每个int64符号频数，再按频数降序stable_sort唯一符号；同频保持std::map数值升序，随后令ID为该排名，解码依赖完整dictionary查表。", - "inputs": "1流symbols:int64[N]", - "outputs": "2流dictionary:int64[D]、ids:int64[N]；无熵模型参数，排序规则固定。", - "example": "输入[8,3,8,5,3,8]，频数8:3、3:2、5:1，字典[8,3,5]，ID=[0,1,0,2,1,0]。输出仍是ID流，需要另接熵编码才能得到前缀码。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A078", - "name": "Static Byte Range Coder", - "layer": "mechanism", - "description": "先把256符号原始频数归一到总频率2^15并构造cum；编码每个byte时令temp=range/2^15，更新low+=cum[s]*temp、range=freq[s]*temp并归一化输出。", - "inputs": "1路uint8[L]", - "outputs": "1路bytes包，内含origLen:uint32、256×uint32原频数、有效bit数和range payload。总频率2^15固定。", - "example": "示意归一化模型总频率32768，freq(A)=24576、freq(B)=8192；初始range=2^32时temp=131072。编码B令low增加24576×131072，range变为8192×131072；解码使用相同cum/freq。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A079", - "name": "FPZIP Adaptive 129-symbol Range Model", - "layer": "mechanism", - "description": "以上一float64重建值做1D预测；把预测与当前映射整数的差分符号和最高位宽映射到129符号，余下低bits直接入range流；RCqsmodel在每个符号后更新。", - "inputs": "1路FPZIP residual packet:bytes（count:uint32+uint64残差）", - "outputs": "1路range wire:bytes。预测器固定lossless 1D double，符号数129。", - "example": "连续相同值的mapped residual=0，选择表示零差分的模型符号并更新自适应频数；解码先从range流读该符号，再按同一预测得到原位。实际符号编号由FPZIP映射固定。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - } - ] - }, - { - "id": "entropy_fse", - "family_number": "24", - "name": "ANS / FSE / ANS／FSE编码", - "group": "Probability coding", - "description": "Uses an Asymmetric Numeral Systems state, often table-normalized as tANS/FSE, to combine near-arithmetic-coding compression with table-driven encode/decode transitions.", - "reference_ids": [ - "O22" - ], - "implementations": [ - { - "id": "A053", - "name": "Frequency Ordered Dictionary IDs", - "layer": "configuration", - "description": "先统计每个int64符号频数，再按频数降序stable_sort唯一符号；同频保持std::map数值升序，随后令ID为该排名，解码依赖完整dictionary查表。", - "inputs": "1流symbols:int64[N]", - "outputs": "2流dictionary:int64[D]、ids:int64[N]；无熵模型参数，排序规则固定。", - "example": "输入[8,3,8,5,3,8]，频数8:3、3:2、5:1，字典[8,3,5]，ID=[0,1,0,2,1,0]。输出仍是ID流，需要另接熵编码才能得到前缀码。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A080", - "name": "Project Static Byte rANS", - "layer": "mechanism", - "description": "把256频率归一到2^15，按输入倒序更新state=(state/fs)*M+(state%fs)+cum；state过大时低字节renorm输出，末尾写32位state。解码按state%M查符号。", - "inputs": "1路uint8[L]", - "outputs": "1路rANS packet:bytes，含origLen:uint32、256×uint16频率表、state/body。prob_bits=15与初态固定。", - "example": "示意模型M=32768、freq(A)=16384、cum(A)=0，state=100000时编码A的新状态=(100000 div16384)×32768+(100000 mod16384)=198304；归一化字节与末状态也必须保存。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - } - ] - }, - { - "id": "entropy_special", - "family_number": "25", - "name": "Special / Literal Coding / 特殊值／字面量编码", - "group": "Probability coding", - "description": "Represents codec-specific entropy or statistical back ends that do not fit the shared Huffman, range, or FSE/ANS categories; behavior is defined by the selected method.", - "reference_ids": [ - "O23" - ], - "implementations": [ - { - "id": "A053", - "name": "Frequency Ordered Dictionary IDs", - "layer": "configuration", - "description": "先统计每个int64符号频数，再按频数降序stable_sort唯一符号；同频保持std::map数值升序，随后令ID为该排名，解码依赖完整dictionary查表。", - "inputs": "1流symbols:int64[N]", - "outputs": "2流dictionary:int64[D]、ids:int64[N]；无熵模型参数，排序规则固定。", - "example": "输入[8,3,8,5,3,8]，频数8:3、3:2、5:1，字典[8,3,5]，ID=[0,1,0,2,1,0]。输出仍是ID流，需要另接熵编码才能得到前缀码。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A081", - "name": "SPDP History-context Repeat Encoder", - "layer": "mechanism", - "description": "以6字节历史匹配作为上下文，hash状态索引last位置；若历史相同则输出最多255的复制count，再输出literal，并以(history<<2)^value更新状态。解码同步重放复制。", - "inputs": "1路SPDP block packet:bytes", - "outputs": "1路repeat/literal bytes。level:uint32∈[0,9]控制表大小，默认5；逆向还需original_size:uint64。", - "example": "历史6字节上下文相同且已输出前缀包含ABCDEFG时，可发出复制count再发literal；count受255上限限制。解码使用相同hash/last状态，从已恢复位置复制，随后观察literal并更新历史。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - } - ] - }, - { - "id": "literal_entropy", - "family_number": "26", - "name": "Literal Entropy / 字面量熵编码", - "group": "LZ decisions and format", - "description": "Codes literal bytes through a separate probability model or codebook while match lengths, offsets, and other sequence symbols use their own representation.", - "reference_ids": [ - "O24" - ], - "implementations": [ - { - "id": "A053", - "name": "Frequency Ordered Dictionary IDs", - "layer": "configuration", - "description": "先统计每个int64符号频数，再按频数降序stable_sort唯一符号；同频保持std::map数值升序，随后令ID为该排名，解码依赖完整dictionary查表。", - "inputs": "1流symbols:int64[N]", - "outputs": "2流dictionary:int64[D]、ids:int64[N]；无熵模型参数，排序规则固定。", - "example": "输入[8,3,8,5,3,8]，频数8:3、3:2、5:1，字典[8,3,5]，ID=[0,1,0,2,1,0]。输出仍是ID流，需要另接熵编码才能得到前缀码。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A077", - "name": "Project Byte Huffman", - "layer": "mechanism", - "description": "统计256个byte频率，按项目固定建树规则生成静态Huffman码；包内保存原长、频率/树恢复信息和有效bit数，解码逐bit沿左右子树直到叶节点。", - "inputs": "1路raw:uint8[L]", - "outputs": "1路Huffman packet:bytes（头部模型/原长/bit数与payload同包）。当前无公开树策略参数。", - "example": "字节AAAA BBB C（频数4、3、1）建立静态Huffman树；高频A获得更短码。若示意树取A=0、B=10、C=11，则ABAC为010011；实际码字由固定建树平局规则决定，模型与有效bit数必须入包。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - } - ] - }, - { - "id": "block_state_control", - "family_number": "27", - "name": "Block State Control / 块状态控制", - "group": "Layout and state", - "description": "Controls block boundaries, state reuse or reset, streaming flushes, and per-block mode choices that determine which history and models remain available to later data.", - "reference_ids": [ - "O25" - ], - "implementations": [ - { - "id": "A082", - "name": "Native Snappy Match Records", - "layer": "mechanism", - "description": "对raw调用当前Snappy RawCompress，再解析其varint原长和literal/copy tag为SNM1记录。逆向按literal或向后复制恢复字节；它实际上保留了完整原生matcher的决策。", - "inputs": "raw:uint8[L]，L≤UINT32_MAX", - "outputs": "5语义流tag:uint8、literals:uint8、literal_length:uint32、copy_offset:uint32、copy_length:uint32。物理为SNM1，当前无公开参数。", - "example": "输入ABCABCABC，由Snappy库产生实际压缩结果，再解析为literal/copy记录，例如literal ABC与向后复制的语义。SNM1保存raw_size及库选定的实际记录，不假定每次都选择同一手工切分。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A083", - "name": "Native LZ4 Linked-block Match Records", - "layer": "mechanism", - "description": "先按64KiB linked-block、level=0生成LZ4F，再解析literal run、offset及match length为LZM1。原生LZ4查hash候选，命中后可向前回扩到anchor，再延长匹配。", - "inputs": "raw:uint8[L]", - "outputs": "5语义流literals:uint8、literal_length:uint32、has_match:uint8、offset:uint16、match_length:uint32。物理为LZM1；块长及level当前固定。", - "example": "输入重复的64KiB字节块，先由LZ4 linked-block编码，再把生成块解析为literal序列及(offset,length)。linked模式的第二块可引用第一块保留的字典，解码必须维持该跨块状态。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A084", - "name": "ALP Native Remaining Pipeline", - "layer": "composition", - "description": "先逆转ALP1前缀中的bitpack、FOR、scaling恢复float64，再调用原生ALP：每1024项选择factor/exponent、标记不能精确恢复的异常，以两条32位limb做FastPFOR并写ALP2。", - "inputs": "ALP1:bytes。", - "outputs": "raw块1流uint64；scaled块5流factor_exponent:int8、low_words:uint32、high_words:uint32、exception_positions:uint16、exception_bits:uint64；混合6角色。物理ALP2封包。", - "example": "输入一个表示[1.25,1.50,1.75]的ALP1前缀包；入口先恢复这三个float64，再运行完整ALP生成ALP2。输出参数、两肢压缩块与异常信息由ALP重新选定，不能把ALP1前缀载荷直接当作原生ALP主体。", - "status": "完整或剩余流程入口，包含多个内部机制；不声称已拆成独立原子动作。" - }, - { - "id": "A085", - "name": "Current REGER Search and Payload", - "layer": "composition", - "description": "逆转RGR1的位打包、FOR、相邻预测与排列后，调用当前REGER。对候选顺序和是否存time比较完整块成本；每序列再选raw、constant、FOR、Delta、TimeLinear或PrevLinear并分段打包。", - "inputs": "输入RGR1:bytes。time/value每序列随模式为1/1/3/4/4/5语义流；字段为int64主体/first/minimum、float32[2]模型、uint8位宽、bytes残差，可另有order。物理为完整REGER wire。", - "outputs": "输入RGR1:bytes。time/value每序列随模式为1/1/3/4/4/5语义流；字段为int64主体/first/minimum、float32[2]模型、uint8位宽、bytes残差，可另有order。物理为完整REGER wire。", - "example": "输入可逆RGR1前缀表示[10,12,14,16]；入口先恢复数值，再由REGER比较常量、FOR、Delta及回归等候选的完整块大小，输出选择标志和对应payload。", - "status": "完整或剩余流程入口，包含多个内部机制；不声称已拆成独立原子动作。" - }, - { - "id": "A087", - "name": "Brotli Quality-6 Remaining Pipeline", - "layer": "composition", - "description": "先逆项目Huffman、LZ和固定字典前缀恢复raw，再调用Brotli quality=6。内部将literal、command、distance分开建模，literal/distance可按context map选Huffman表，块切换另有控制码。", - "inputs": "输入Huffman(LZS1(BRD1(raw)))结构化bytes。压缩块4语义角色：entropy_models、literal_codebits、command_codebits、distance_codebits（均bytes），有块切换则加1；raw块1角色。quality/window当前固定。", - "outputs": "输入Huffman(LZS1(BRD1(raw)))结构化bytes。压缩块4语义角色：entropy_models、literal_codebits、command_codebits、distance_codebits（均bytes），有块切换则加1；raw块1角色。quality/window当前固定。", - "example": "项目前缀可逆表示ASCII ABCABCABC；入口先还原这9字节，再由Brotli quality=6生成压缩流B。用Brotli解码B回到同一9字节；不把项目前缀中的token当作Brotli内部command。", - "status": "完整或剩余流程入口，包含多个内部机制；不声称已拆成独立原子动作。" - }, - { - "id": "A088", - "name": "Zstd Level-9 Remaining Pipeline", - "layer": "composition", - "description": "逆项目rANS、Huffman、parse wrapper及LZ前缀得到raw，交给Zstd level=9。原生路径对literal使用Huffman或raw/RLE，对literal length、match length、offset等序列符号使用FSE及附加位。", - "inputs": "输入项目rANS(Huffman(ZPP1(LZS1(raw)))) bytes。压缩schema为L+H+F+S+I(sequence)：L=1/4路literal bytes，H=0/1模型，F=0..3模型，S=0..3个uint32状态及sequence_bits；raw/RLE各1角色。", - "outputs": "输入项目rANS(Huffman(ZPP1(LZS1(raw)))) bytes。压缩schema为L+H+F+S+I(sequence)：L=1/4路literal bytes，H=0/1模型，F=0..3模型，S=0..3个uint32状态及sequence_bits；raw/RLE各1角色。", - "example": "项目前缀表示ABCABCABC，入口恢复raw后用Zstd level=9生成Zstd frame；其literal和sequence熵模型由Zstd内部选择。解码frame，再逆构造前缀所需表示。", - "status": "完整或剩余流程入口，包含多个内部机制；不声称已拆成独立原子动作。" - }, - { - "id": "A089", - "name": "C-Blosc1 BLOSCLZ with Shuffle", - "layer": "composition", - "description": "逆项目token、LZ及SHF1 shuffle前缀后，以typesize执行Blosc1 shuffle与BLOSCLZ-5。BLOSCLZ先对末1/4采样，probe使用12-bit hash；根据split标志或预测压缩性选择minlen/ipshift为3或4。", - "inputs": "输入LZT1(LZS1(SHF1)) bytes，typesize∈{1,4,8}。正常chunk按plane编号合并为1至typesize流encoded_plane:bytes；memcpy为1流raw:bytes。level5/shuffle/单线程固定。", - "outputs": "输入LZT1(LZS1(SHF1)) bytes，typesize∈{1,4,8}。正常chunk按plane编号合并为1至typesize流encoded_plane:bytes；memcpy为1流raw:bytes。level5/shuffle/单线程固定。", - "example": "输入4个相同的64位字，typesize=8；恢复原始32字节后，Blosc1按字节shuffle再运行BLOSCLZ-5，输出完整Blosc块。解码和逆shuffle恢复32字节。", - "status": "完整或剩余流程入口，包含多个内部机制；不声称已拆成独立原子动作。" - }, - { - "id": "A090", - "name": "C-Blosc2 BLOSCLZ Shared Pipeline", - "layer": "composition", - "description": "执行C-Blosc2的BLOSCLZ-5、byte shuffle、单线程和自动块划分。P114先逆项目prefix；P119直接接raw，再把输出chunk转成B2P1。两者来源版本的该活动压缩路径已归为同一组。", - "inputs": "P114输入结构化prefix及typesize=1/4/8；P119输入raw:uint8且typesize固定1。正常按plane合并1至typesize路bytes，memcpy/特殊模式1路body（zero可空）；物理chunk或B2P1。", - "outputs": "P114输入结构化prefix及typesize=1/4/8；P119输入raw:uint8且typesize固定1。正常按plane合并1至typesize路bytes，memcpy/特殊模式1路body（zero可空）；物理chunk或B2P1。", - "example": "raw为4个相同64位字，typesize=8、单线程、BLOSCLZ-5、byte shuffle；C-Blosc2生成chunk，再由对应入口转成B2P1或保留本机块结构。解码恢复32字节。", - "status": "完整或剩余流程入口，包含多个内部机制；不声称已拆成独立原子动作。" - }, - { - "id": "A091", - "name": "Bzip2 Block-sort Remaining Pipeline", - "layer": "composition", - "description": "逆项目Huffman、URL1、MTF、BWT恢复raw，再调用Bzip2 level4。原生编码包含块排序/BWT、MTF与RUNA/RUNB零游程符号，并按组选择多张Huffman表编码。", - "inputs": "Huffman(URL1(MTF(BWT(raw)))) bytes，URL1 symbol_width=1。", - "outputs": "5语义组used_alphabet:uint8、BWT_primary:uint32、code_lengths:uint8、selectors:uint8、coded_MTF_RLE2:bytes；跨块按同角色合并。", - "example": "项目前缀BWT/MTF/RLE/Huffman表示BANANA，入口先恢复6字节，再调用Bzip2 level=4，生成带块排序和自身熵模型的完整流。最终解压应为BANANA。", - "status": "完整或剩余流程入口，包含多个内部机制；不声称已拆成独立原子动作。" - }, - { - "id": "A092", - "name": "Parameterized DEFLATE Core", - "layer": "composition", - "description": "gzip/zlib profile先逆项目前缀得到raw；CORAD直接压CTK2。三者共用parameterized_deflate，执行LZ匹配及stored/fixed/dynamic块选择；压缩块使用literal/length与distance两套Huffman码。", - "inputs": "输入由io_variant选择gzip_prefix/zlib_prefix bytes或CTK2；level:int32∈[0,9]。stored/fixed/dynamic语义流分别1/2/3路，混合最多4角色：raw、literal_length_bits、distance_bits、Huffman_model（bytes）。", - "outputs": "输入由io_variant选择gzip_prefix/zlib_prefix bytes或CTK2；level:int32∈[0,9]。stored/fixed/dynamic语义流分别1/2/3路，混合最多4角色：raw、literal_length_bits、distance_bits、Huffman_model（bytes）。", - "example": "raw=ABCABCABC，DEFLATE核心选择stored/fixed/dynamic块并输出压缩core；raw-DEFLATE、zlib、gzip的窗口和封包配置由调用方指定，恢复使用匹配配置。", - "status": "完整或剩余流程入口，包含多个内部机制；不声称已拆成独立原子动作。" - }, - { - "id": "A093", - "name": "XZ / LZMA2 Remaining Pipeline", - "layer": "composition", - "description": "逆项目Range、parse wrapper和LZ前缀恢复raw，再用preset5、CRC64的LZMA2路径。其literal/匹配/重复距离由二元自适应概率模型和range coder写出，并在chunk边界处理模型或字典重置。", - "inputs": "Range(ZPP1(LZS1(raw))) bytes；压缩chunk", - "outputs": "range_bytes:bytes与final_flush:bytes两角色，raw为1角色，混合3角色；物理为XZ核心包，preset5固定。", - "example": "项目前缀可逆表示ABCABCABC，入口还原raw后运行preset=5、CRC64的XZ/LZMA2路径；输出XZ流，解压得到同一9字节。", - "status": "完整或剩余流程入口，包含多个内部机制；不声称已拆成独立原子动作。" - }, - { - "id": "A094", - "name": "REGER Candidate Search with Compatibility Arithmetic", - "layer": "composition", - "description": "把64位原字按int64处理，在块内比较排列、是否保存time及六种序列模式，选择完整payload较小的候选。拟合用double乘积及Python式fsum，预测用double，再做截断和范围处理。", - "inputs": "words:uint8[8N]，按int64位型解释；配置block_size默认513、segment_size默认16、profile=fast/balanced。", - "outputs": "每time/value序列按模式1/1/3/4/4/5路，可加order；字段类型同A085，物理RGP1。", - "example": "一个int64块[5,5,5,5]会把constant作为候选，与raw/FOR/Delta等完整候选比较；最终payload同时记录实际选定模式，解码不能自行重新运行成本搜索。", - "status": "完整或剩余流程入口，包含多个内部机制；不声称已拆成独立原子动作。" - } - ] - }, - { - "id": "container_frame", - "family_number": "28", - "name": "Container / Framing / 容器与封包", - "group": "Layout and state", - "description": "Wraps a compressed payload with format identifiers, lengths, checksums, member boundaries, or other metadata needed to transport, validate, and decode the stream.", - "reference_ids": [ - "O26" - ], - "implementations": [ - { - "id": "A086", - "name": "Magic and Length Envelope", - "layer": "adapter", - "description": "在已有payload前写4字节magic和8字节payload长度，逆向移除并校验对应字段。五个reference容器使用同一wrap/unwrap；CORAD在CZL2外增加CRD2外壳。", - "inputs": "payload:uint8[L]；CORAD输入CZL2:bytes。", - "outputs": "1条完整container:bytes，主体不变；magic/profile及payload合法性属于格式配置。", - "example": "magic=TEST、payload=01 02 03，输出TEST、长度3的小端8字节字段、01 02 03。解码核对长度后返回3字节。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A095", - "name": "Reference Integer Header and Checksum Slots", - "layer": "adapter", - "description": "在int64值序列前插入[31,139,N低16位]，尾部插入所有值之和的低16位；逆向直接取去掉前三项和最后一项后的原序列。", - "inputs": "values:int64[N]", - "outputs": "1流container:int64[N+4]。无参数，新增4个int64槽位；与byte帧的区别留在I/O和schema中。", - "example": "输入[5,7]，输出int64序列[31,139,2,5,7,12]。当前逆函数去除前3项及末项得到[5,7]；末字段是槽位，不构成经过验证的校验和保障。", - "status": "当前项目参考/代理或封装步骤；按下面实际计算解释，不等同名称暗示的上游算法。" - }, - { - "id": "A096", - "name": "LZ4 Frame Assembly", - "layer": "adapter", - "description": "由LZI1已编码块生成LZ4F：写magic、descriptor及其XXH32派生校验字节，再写块长度与raw标志、块body，最后写零长度终止块。逆向解析回块记录。", - "inputs": "LZI1:bytes（raw_size:uint64及每块标志、长度、body）", - "outputs": "1流LZ4F:bytes。不接受裸值或原始payload，无公开算法参数。", - "example": "已有一个3字节raw块ABC，LZ4F块长字段为0x80000003，后接ABC，再写0终止块；frame descriptor及校验字节另按配置生成。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A097", - "name": "Stored Header / Core / Trailer Assembly", - "layer": "adapter", - "description": "解析已保存的header、core、trailer长度，将三段原样按顺序拼接。gzip/zlib共用assemble；XZ固定两端各12字节，直接移除XZC1的描述头得到原frame。", - "inputs": "DFC1(G/Z)或XZC1:bytes，内含3段uint8[]", - "outputs": "1流完整gzip/zlib/XZ:bytes。profile固定选择格式，长度必须与packet一致。", - "example": "输入header=01 02、core=aa bb cc、trailer=ff及三段长度，输出01 02 aa bb cc ff。逆向用固定或已记录边界恢复三段。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A098", - "name": "Legacy Two-limb Numeric Frame", - "layer": "adapter", - "description": "从LNS1的两条已编码limb取出low/high字节；写low长度后依次放两条body。整数前置I标签，浮点前置F标签及scale；逆向依标签与长度还原字段。", - "inputs": "LNS1 JSON bytes，stage=wires、format=scaled、单块、wires为2个hex串；float另含scale:int32。", - "outputs": "1流legacy numeric wire:bytes。", - "example": "整数profile，两肢body分别为aa bb和cc，输出I类型标记、low长度2、low body与high body；浮点profile还写F和scale。长度字段确定两肢边界。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A099", - "name": "Legacy ALP2 Block Frame", - "layer": "adapter", - "description": "把上游已选择的raw/scaled块写成ALP2；raw块写原字，scaled块写fac、exp、两肢body及异常位置/原字；总长、块长与模式按固定字段宽度序列化。", - "inputs": "LNS1 JSON bytes，stage=wires、format=alp；含count、raw或fac/exp、wires及exceptions。", - "outputs": "1流ALP2:bytes；块长≤1024，无数值搜索参数。", - "example": "ALP2一块mode=raw、count=2、原字对应1.25与2.5，封包写块模式和两原字；scaled块则保存fac/exp、两肢body及异常表，模式决定解码分支。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A100", - "name": "REGER3 Single-column Frame", - "layer": "adapter", - "description": "解析RGP1，写REGER3 magic、总数、固定列数1、块长和segment长；逐块写count、flags、body长度和既有body。逆向重新构造RGP1。", - "inputs": "RGP1:bytes，count:uint64、block/segment_size:uint32、块记录", - "outputs": "1流REGER3:bytes。输出总数限uint32、块及segment尺寸限uint16。", - "example": "单列共3项、一块，RGP1内已有flags及body；REGER3 frame写总数3、列数1、块长、segment长，再写该块count/flags/body_size和body。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A101", - "name": "BUFF Raw64 Header", - "layer": "adapter", - "description": "对已经按BUFF高32位在前布局的words，前置count:uint32、bits_per_value=64、lower_bits=0；空输入输出空包。逆向核对这三个头字段后返回原words。", - "inputs": "words:uint8[8N]，每字高32位在前、各肢内部小端", - "outputs": "1流BUFF raw64 wire:bytes。N≤UINT32_MAX，无公开参数。", - "example": "已有两项重排后的64位word，前置count=2、bits_per_value=64、lower_bits=0。解码核对头后取出16字节word载荷，再由A052恢复字序。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - }, - { - "id": "A102", - "name": "Blosc2 Logical-to-physical Block Frame", - "layer": "adapter", - "description": "解析B2P1的logical_id与block body；按记录的物理顺序拼body，并为各逻辑块重建offset表。direct/memcpy/special-zero模式走专门分支，最后写Blosc2头与扩展字段。", - "inputs": "B2P1:bytes，含header/extension、direct、logical_id:uint32、body:bytes", - "outputs": "1流Blosc2 chunk:bytes。当前schema固定version=5、codec_version=1、typesize=1。", - "example": "两个逻辑块L0、L1在物理区按L1后L0存放，offset表分别指向各块的实际开始地址。解码按逻辑顺序索引块，不把物理拼接次序误当作列顺序。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。" - } - ] - }, - { - "id": "split_columns", - "family_number": "29", - "name": "Split Columns / 列拆分", - "group": "Layout and state", - "description": "Splits fixed records, floating-point fields, numeric bit fields, or serialized fixed-width bitstreams into typed component streams while retaining the metadata required for exact reconstruction. The four current entries are project implementations informed by OpenZL layout nodes; their local transport is not the OpenZL wire format.", - "reference_ids": [ - "O27", - "S01" - ], - "implementations": [ - { - "id": "A049", - "name": "Parameterized Fixed-record Byte Columnization（参数化定宽记录列化）", - "layer": "mechanism", - "description": "把输入解释为 N 个宽度 W 的定长记录；transpose 按记录内字节位置输出 W 条 byte 列。split-struct 则按给定字段宽度数组切连续字节字段，各字段仍保持行序。1/4/8-byte shuffle 是 W 的固定 profile。", - "inputs": "Struct(W) 的 N 条定长记录；兼容 Serial adapter 可接任意 byte[L]，仅完整记录参与列化并保留尾部。参数：W 或字段宽度数组，当前兼容 profile W∈{1,4,8}。", - "outputs": "transpose 为 W 条 bytes[N]；split-struct 为字段数条 Struct(field_width) 流。", - "example": "两条宽3字节记录[01,02,03]和[11,12,13]，transpose输出三列[01,11]、[02,12]、[03,13]。若字段宽[1,2]，输出1字节列和2字节结构列。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。", - "source_url": "https://github.com/facebook/openzl", - "reference_ids": [ - "O27" - ] - }, - { - "id": "A103", - "name": "Float Sign-fraction/Exponent Deconstruction（浮点符号尾数与指数拆分）", - "layer": "mechanism", - "description": "按 fp16、bf16 或 fp32 位格式读取原始位型，将 fraction 左移一位并把 sign 填入最低位，形成 sign_fraction；exponent 单独取出。逆变换再把 sign、fraction、exponent 放回原位，因此保留 NaN、无穷和 signed zero 的位型。", - "inputs": "Numeric fp16/bf16/fp32 原始位型，存储宽度分别 2/2/4 bytes；参数：format。", - "outputs": "2 条流--Struct(ceil((fraction_bits+1)/8)) 的 sign_fraction[N] 与 uint8 exponent[N]。当前实现不接 fp64。", - "example": "fp32的-0原位0x80000000：exponent=0、fraction=0、sign=1，输出sign_fraction=1、exponent=0。fp32的1.0则输出sign_fraction=0、exponent=127。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。", - "source_url": "https://github.com/facebook/openzl", - "reference_ids": [ - "O27" - ] - }, - { - "id": "A104", - "name": "LSB-first Bit-field Split（低位优先位域拆分）", - "layer": "mechanism", - "description": "对每个 Numeric 元素从最低位起，依次按 widths=[w0,w1,…] 掩码并右移，得到若干定宽数值字段。各宽度为 1-64 bit，总宽不超过输入存储位数；若未覆盖高位，所有被省略高位必须为 0 才可逆。", - "inputs": "Numeric 整数或原始浮点位型，storage_bits 由类型确定；参数：widths，或 top8/浮点格式 profile。", - "outputs": "len(widths) 条 Numeric 流，每条为能容纳对应 wi 的最小存储宽度；浮点 profile 输出 fraction、exponent、sign 3 流。", - "example": "uint16值0x1234，widths=[4,4,8]，低位优先输出[4,3,0x12]。逆向4|(3<<4)|(0x12<<8)=0x1234。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。", - "source_url": "https://github.com/facebook/openzl", - "reference_ids": [ - "O27" - ] - }, - { - "id": "A105", - "name": "Fixed-width Bitstream Unpack（定宽位流解包）", - "layer": "mechanism", - "description": "把 Serial byte 流视为 LSB-first 连续位串，每 bits 位读出一个无符号值，共 floor(8L/bits) 个；末尾不足 bits 的 remainder 位值写入 side metadata。逆变换按相同位序重打包并补回 remainder。", - "inputs": "Serial bytes[L]；参数bits∈[1,64]。N=floor(8L/bits)，要求ceil(N*bits/8)=L，即未消费部分不足一个完整字节。", - "outputs": "Numeric unsigned[N]，每元素按storage_width(bits)占用1/2/4/8字节；其存储字节数可大于输入。side metadata保存bits与末尾remainder_value，原输入长度L随框架保存。", - "example": "单字节0xd6按LSB-first、bits=3解包：得到[6,2]，剩余高2位值3作为remainder保存。逆向6|(2<<3)|(3<<6)=0xd6。", - "status": "源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。", - "source_url": "https://github.com/facebook/openzl", - "reference_ids": [ - "O27" - ] - } - ] - }, - { - "id": "prediction_probability", - "family_number": "30", - "name": "Prediction Probability / 概率预测", - "group": "Probability coding", - "description": "Produces causal probabilities before the current event is observed. A106-A109 expose next-bit P(1) from distinct cmix-family models, while A110 exposes a 256-way next-byte distribution from NNCP. These providers remain separate from graph-level entropy integration and benchmark method scheduling.", - "reference_ids": [ - "O28", - "O29" - ], - "implementations": [ - { - "id": "A106", - "name": "cmix PPMd Byte-model Probability（PPMd 字节模型概率）", - "layer": "mechanism", - "description": "PPMd 按最近字节上下文维护自适应符号概率。adapter 在每个 bit 前从模型取 P(1)，观察 8 个 MSB-first bit 后才组装字节并调用 ByteUpdate；字节边界还可把 256 个符号权重归一化为完整下一字节分布。", - "inputs": "状态化命令流；每步 observe 输入 uint8 bit∈{0,1}，批量接口输入 bytes[L] 并按 MSB-first 展开。参数：order 2-64（默认 6）、memory 4-256 MiB（默认 16）。", - "outputs": "每 bit 1 条 float64 P(1) 流；字节边界可选 256×float64 分布。", - "example": "输入字节A=0x41，对应MSB-first事件0,1,0,0,0,0,0,1。每个事件之前读取模型的P(1)，随后观察该bit；第8次观察后执行ByteUpdate。八个实际概率由固定PPMd状态产生，不用常数概率代替。", - "status": "真实来源模型与adapter已完成本轮PPMd/DMC/PAQ共8项有限回归；尚未接入图熵编码边，不能据此声称完整方法测评完成。", - "reference_ids": [ - "O29" - ], - "source_url": "https://github.com/byronknoll/cmix" - }, - { - "id": "A107", - "name": "cmix PAQ dmcModel Count Branch（PAQ DMC 计数分支）", - "layer": "mechanism", - "description": "在 bit 状态图当前节点保存定点 c0、c1；观察 y 后用 x←x·63/64+y·1024 更新计数和有限状态，达到阈值时按访问占比分裂后继节点，随后沿 nx[y] 转移。预测为 (c1+1)·4096/(c0+c1+2)。", - "inputs": "与 A106 相同的因果 uint8 bit 观察流或 MSB-first bytes[L]；当前参数固定 extra_nodes=65536、cloning_threshold_fixed_point=1024。", - "outputs": "1 条 float64 P(1) 流，由原生 12-bit 整数概率除以 4096；无字节分布流。", - "example": "在已给定节点c0=c1=1024时，预测整数概率为(1024+1)×4096/(2048+2)=2048，即P(1)=0.5；接收bit=1后再更新计数并沿nx[1]转移。此例是节点状态计算，不声称为默认初始节点。", - "status": "真实来源模型与adapter已完成本轮PPMd/DMC/PAQ共8项有限回归；尚未接入图熵编码边，不能据此声称完整方法测评完成。", - "reference_ids": [ - "O29" - ], - "source_url": "https://github.com/byronknoll/cmix" - }, - { - "id": "A108", - "name": "cmix PAQ8 Hybrid Final Predictor（PAQ8 混合最终预测器）", - "layer": "mechanism", - "description": "运行 PAQ8 Predictor：contextModel2 先汇集上下文模型，再按内容状态经过 APM/APM1 校准与加权，形成 12-bit 最终 pr；每次 observe 把真实 bit 写入全局 y 并执行完整 update，字节完成时同步缓冲区和上下文。", - "inputs": "因果 uint8 bit 流或 MSB-first bytes[L]；参数：paq_level 0-3（默认 0），用于选择内存级别。", - "outputs": "1 条 float64 P(1) 流，值为原生最终 p()/4096；无压缩字节或额外语义流。", - "example": "对输入bits=[0,1]，先取得最终12-bit pr0，再observe(0)更新整套PAQ状态；取pr1后再observe(1)。pr0/pr1是混合和APM校准后的最终输出，不能用某个子模型值替代。", - "status": "真实来源模型与adapter已完成本轮PPMd/DMC/PAQ共8项有限回归；尚未接入图熵编码边，不能据此声称完整方法测评完成。", - "reference_ids": [ - "O29" - ], - "source_url": "https://github.com/byronknoll/cmix" - }, - { - "id": "A109", - "name": "cmix v21 Full Predictor（cmix v21 完整预测器）", - "layer": "mechanism", - "description": "完整 Predictor 收集 models、byte_models 与 byte_mixers 的输出，送入分层 mixers；最终经 logistic 与 SSE 校准得到 P(1)，observe 时逐一更新模型、mixer、SSE 和上下文管理器。当前构造使用全 256 字节 vocabulary，不做预处理或预训练。", - "inputs": "因果 uint8 bit 流或 MSB-first bytes[L]，vocabulary 固定为全部 256 个字节；参数：allow_large_memory=true 才允许启动，adapter 另有 timeout。", - "outputs": "1 条 float64 P(1) 流，由上游 float 转为传输用 binary64；无压缩数据流。", - "example": "输入字节00时，依次执行8次predict→处理当前0→observe(0)；各层mixer与SSE在观察后更新。解码端从相同256字节词表及模型初始化开始重放相同8次状态转移。", - "status": "来源模型及adapter见源码；可运行性受构建依赖与内存门控约束。本次未执行cmix或验证其图熵编码接入。", - "reference_ids": [ - "O29" - ], - "source_url": "https://github.com/byronknoll/cmix" - }, - { - "id": "A110", - "name": "NNCP Transformer Byte Probability (stream1)", - "layer": "mechanism", - "description": "调用官方NNCP 3.3 Transformer与LibNC得到下一字节的256项softmax概率。先predict，再由匹配的write_sym/read_sym处理当前事件，最后observe以该字节为训练目标执行gradient和model_update；模型状态、初始化及重置在解码端一致。", - "inputs": "逐个 uint8 字节事件；必须先 predict，再对当前字节 encode/decode，最后 observe。每会话最多 4096 字节。", - "outputs": "float32[256]，按符号0..255排列。transformer-stream1-v1固定batch=1、seg_len=1、seed=123、V=256，保留上游default网络结构；每会话最多4096字节。官方LibNC依赖Linux x86-64，本机通过显式Docker验证。独立NNP1压缩路径有24字节固定头和完整算术码体；未接图熵边或86方法调度。", - "example": "以初始previous=0预测首字节t0的256项分布，write_sym/read_sym处理t0后再observe(t0)，调用原模型gradient/update；下一步以已知t0预测t1。重复predict返回同一缓存分布。独立NNCPSymbolSession和NNP1冷解码路径已接入；不把概率向量当作压缩数据。", - "status": "真实NNCP独立字节概率接口与配套NNP1冷解码路径已接入；官方LibNC要求Linux x86-64或显式Docker。与图熵编码边、默认NNCP方法评测分开。", - "reference_ids": [ - "O28" - ], - "source_url": "https://bellard.org/nncp/" - } - ] - } - ], - "references": [ - { - "id": "M01", - "citation": "Panagiotis Liakos, Katia Papakonstantinopoulou, Yannis Kotidis. “Chimp.” Proceedings of the VLDB Endowment, 2022.", - "url": "https://doi.org/10.14778/3551793.3551852", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "web_verified_primary_or_official_2026-10-08" - }, - { - "id": "M02", - "citation": "Yufeng Liu, Yao Shen, Fenghua Zhang, Feiteng Huang. “Pongo: Efficient Lossless Floating Point Compression.” Cloud Computing and Data Science, 2025.", - "url": "https://doi.org/10.37256/ccds.6220257102", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "publisher_blocked_403_not_a_dead_link" - }, - { - "id": "M03", - "citation": "Tuomas Pelkonen, Scott Franklin, Justin Teller, Paul Cavallaro, Qi Huang, Justin Meza, Kaushik Veeraraghavan. “Gorilla.” Proceedings of the VLDB Endowment, 2015.", - "url": "https://doi.org/10.14778/2824032.2824078", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "web_verified_primary_or_official_2026-10-08" - }, - { - "id": "M04", - "citation": "Apache Software Foundation. “Apache TsFile Encoding (TS_2DIFF).” Official documentation or source repository, n.d.", - "url": "https://tsfile.apache.org/", - "kind": "official_documentation_or_project_page", - "source_slide": 2, - "verification_status": "web_verified_primary_or_official_2026-10-08" - }, - { - "id": "M05", - "citation": "Jinzhao Xiao, Zihan Guo, Shaoxu Song. “BOS: Bit-Packing with Outlier Separation.” 2025 IEEE 41st International Conference on Data Engineering (ICDE), 2025.", - "url": "https://doi.org/10.1109/ICDE65448.2025.00107", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "doi_resolved_to_primary_publisher_2026-10-08" - }, - { - "id": "M06", - "citation": "Jinzhao Xiao, Wendi He, Shaoxu Song, Xiangdong Huang, Chen Wang, Jianmin Wang. “REGER: Reordering Time Series Data for Regression Encoding.” 2024 IEEE 40th International Conference on Data Engineering (ICDE), 2024.", - "url": "https://doi.org/10.1109/ICDE60146.2024.00100", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "doi_resolved_to_primary_publisher_2026-10-08" - }, - { - "id": "M07", - "citation": "Daniel Lemire, Nathan Kurz, Christoph Rupp. “Stream VByte : Faster byte-oriented integer compression.” Information Processing Letters, 2018.", - "url": "https://doi.org/10.1016/j.ipl.2017.09.011", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M08", - "citation": "Daniel Lemire, Leonid Boytsov, Nathan Kurz. “SIMD compression and the intersection of sorted integers.” Software: Practice and Experience, 2015.", - "url": "https://doi.org/10.1002/spe.2326", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "publisher_blocked_403_not_a_dead_link" - }, - { - "id": "M09", - "citation": "lemire project contributors. “FastPFor/Simple8b reference implementation.” Official documentation or source repository, n.d.", - "url": "https://github.com/lemire/FastPFor", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "web_verified_primary_or_official_2026-10-08" - }, - { - "id": "M10", - "citation": "Apache Software Foundation. “Run Length Encoding / Bit-Packing Hybrid.” Official documentation or source repository, n.d.", - "url": "https://parquet.apache.org/docs/file-format/data-pages/encodings/#run-length-encoding-bit-packing-hybrid-rle-3", - "kind": "official_documentation_or_project_page", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M11", - "citation": "Julien Spiegel, Patrice Wira, Gilles Hermann. “A Comparative Experimental Study of Lossless Compression Algorithms for Enhancing Energy Efficiency in Smart Meters.” 2018 IEEE 16th International Conference on Industrial Informatics (INDIN), 2018.", - "url": "https://doi.org/10.1109/INDIN.2018.8471921", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "doi_resolved_to_primary_publisher_2026-10-08" - }, - { - "id": "M12", - "citation": "Davis Blalock, Samuel Madden, John Guttag. “Sprintz.” Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 2018.", - "url": "https://doi.org/10.1145/3264903", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "publisher_blocked_403_not_a_dead_link" - }, - { - "id": "M13", - "citation": "Daniel Abadi, Samuel Madden, Miguel Ferreira. “Integrating compression and execution in column-oriented database systems.” Proceedings of the 2006 ACM SIGMOD international conference on Management of data, 2006.", - "url": "https://doi.org/10.1145/1142473.1142548", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "publisher_blocked_403_not_a_dead_link" - }, - { - "id": "M14", - "citation": "Apache Software Foundation. “Apache Parquet Encodings: DELTA_BINARY_PACKED.” Official documentation or source repository, n.d.", - "url": "https://parquet.apache.org/docs/file-format/data-pages/encodings/#delta-encoding-delta_binary_packed-5", - "kind": "official_documentation_or_project_page", - "source_slide": 2, - "verification_status": "web_verified_primary_or_official_2026-10-08" - }, - { - "id": "M15", - "citation": "Azim Afroozeh, Leonardo X. Kuffo, Peter Boncz. “ALP: Adaptive Lossless floating-Point Compression.” Proceedings of the ACM on Management of Data, 2023.", - "url": "https://doi.org/10.1145/3626717", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "publisher_blocked_403_not_a_dead_link" - }, - { - "id": "M16", - "citation": "Yihao Liu, Xinyu Zeng, Huanchen Zhang. “LeCo: Lightweight Compression via Learning Serial Correlations.” Proceedings of the ACM on Management of Data, 2024.", - "url": "https://doi.org/10.1145/3639320", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "publisher_blocked_403_not_a_dead_link" - }, - { - "id": "M17", - "citation": "Yuanyuan Yao, Lu Chen, Ziquan Fang, Yunjun Gao, Christian S. Jensen, Tianyi Li. “Camel: Efficient Compression of Floating-Point Time Series.” Proceedings of the ACM on Management of Data, 2024.", - "url": "https://doi.org/10.1145/3698802", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "publisher_blocked_403_not_a_dead_link" - }, - { - "id": "M18", - "citation": "Ruiyuan Li, Zheng Li, Yi Wu, Chao Chen, Yu Zheng. “Elf: Erasing-Based Lossless Floating-Point Compression.” Proceedings of the VLDB Endowment, 2023.", - "url": "https://doi.org/10.14778/3587136.3587149", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "web_verified_primary_or_official_2026-10-08" - }, - { - "id": "M19", - "citation": "Ruiyuan Li et al. “Erasing-based Lossless Compression Method for Streaming Floating-Point Time Series.” arXiv:2306.16053, 2023.", - "url": "https://doi.org/10.48550/arXiv.2306.16053", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M20", - "citation": "Martin Burtscher, Paruj Ratanaworabhan. “High Throughput Compression of Double-Precision Floating-Point Data.” 2007 Data Compression Conference (DCC'07), 2007.", - "url": "https://doi.org/10.1109/DCC.2007.44", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "doi_resolved_to_primary_publisher_2026-10-08" - }, - { - "id": "M21", - "citation": "Martin Burtscher, Paruj Ratanaworabhan. “pFPC: A Parallel Compressor for Floating-Point Data.” 2009 Data Compression Conference, 2009.", - "url": "https://doi.org/10.1109/DCC.2009.43", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "doi_resolved_to_primary_publisher_2026-10-08" - }, - { - "id": "M22", - "citation": "Steven Claggett, Sahar Azimi, Martin Burtscher. “SPDP: An Automatically Synthesized Lossless Compression Algorithm for Floating-Point Data.” 2018 Data Compression Conference, 2018.", - "url": "https://doi.org/10.1109/DCC.2018.00042", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "doi_resolved_to_primary_publisher_2026-10-08" - }, - { - "id": "M23", - "citation": "Abdelouahab Khelifati, Mourad Khayati, Philippe Cudre-Mauroux. “CORAD: Correlation-Aware Compression of Massive Time Series using Sparse Dictionary Coding.” 2019 IEEE International Conference on Big Data (Big Data), 2019.", - "url": "https://doi.org/10.1109/BigData47090.2019.9005580", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "doi_resolved_to_primary_publisher_2026-10-08" - }, - { - "id": "M24", - "citation": "Chunwei Liu, Hao Jiang, John Paparrizos, Aaron J. Elmore. “Decomposed bounded floats for fast compression and queries.” Proceedings of the VLDB Endowment, 2021.", - "url": "https://doi.org/10.14778/3476249.3476305", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "web_verified_primary_or_official_2026-10-08" - }, - { - "id": "M25", - "citation": "Peter Lindstrom, Martin Isenburg. “Fast and Efficient Compression of Floating-Point Data.” IEEE Transactions on Visualization and Computer Graphics, 2006.", - "url": "https://doi.org/10.1109/TVCG.2006.143", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "doi_resolved_to_primary_publisher_2026-10-08" - }, - { - "id": "M26", - "citation": "Noushin Azami, Alex Fallin, Martin Burtscher. “Efficient Lossless Compression of Scientific Floating-Point Data on CPUs and GPUs.” Proceedings of the 30th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 1, 2025.", - "url": "https://doi.org/10.1145/3669940.3707280", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "publisher_blocked_403_not_a_dead_link" - }, - { - "id": "M27", - "citation": "P. Deutsch. “GZIP file format specification version 4.3.” RFC Editor, 1996.", - "url": "https://doi.org/10.17487/RFC1952", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "web_verified_primary_or_official_2026-10-08" - }, - { - "id": "M28", - "citation": "J. Alakuijala, Z. Szabadka. “Brotli Compressed Data Format.” RFC Editor, 2016.", - "url": "https://doi.org/10.17487/RFC7932", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "web_verified_primary_or_official_2026-10-08" - }, - { - "id": "M29", - "citation": "Y. Collet. “Zstandard Compression and the 'application/zstd' Media Type.” RFC Editor, 2021.", - "url": "https://doi.org/10.17487/RFC8878", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "web_verified_primary_or_official_2026-10-08" - }, - { - "id": "M30", - "citation": "lz4 project contributors. “LZ4 Frame Format Specification.” Official documentation or source repository, n.d.", - "url": "https://github.com/lz4/lz4/", - "kind": "standard_or_specification", - "source_slide": 2, - "verification_status": "web_verified_primary_or_official_2026-10-08" - }, - { - "id": "M31", - "citation": "The Tukaani Project. “XZ File Format Specification.” Official documentation or source repository, n.d.", - "url": "https://tukaani.org/xz/format.html", - "kind": "standard_or_specification", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M32", - "citation": "g1mv project contributors. “Density compression library.” Official documentation or source repository, n.d.", - "url": "https://github.com/g1mv/density", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M33", - "citation": "google project contributors. “Gipfeli compression library.” Official documentation or source repository, n.d.", - "url": "https://github.com/google/gipfeli", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M35", - "citation": "MATCODE Software. “LZMAT: Open-source Extremely Fast Data Compression Library, Version 1.0.” Official project documentation (archived), 2008.", - "url": "https://web.archive.org/web/20160306180328/http://www.matcode.com/lzmat.htm", - "kind": "archived_project_page_or_source_mirror", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M36", - "citation": "oberhumer.com. “LZO real-time compression library.” Official documentation or source repository, n.d.", - "url": "http://www.oberhumer.com/opensource/lzo/", - "kind": "official_documentation_or_project_page", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M37", - "citation": "Ross N. Williams. “An Extremely Fast Ziv-Lempel Data Compression Algorithm.” Proceedings of the Data Compression Conference, 1991.", - "url": "https://doi.org/10.1109/DCC.1991.213344", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "doi_resolved_to_primary_publisher_2026-10-08" - }, - { - "id": "M38", - "citation": "conorstokes.github.io. “An LZ Codec Designed for SSE Decompression.” Official documentation or source repository, n.d.", - "url": "http://conorstokes.github.io/compression/2016/02/15/an-LZ-codec-designed-for-SSE-decompression", - "kind": "official_documentation_or_project_page", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M39", - "citation": "lzfse project contributors. “LZFSE and LZVN reference implementation.” Official documentation or source repository, n.d.", - "url": "https://github.com/lzfse/lzfse", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "web_verified_primary_or_official_2026-10-08" - }, - { - "id": "M40", - "citation": "johnezang project contributors. “Pithy compression library.” Official documentation or source repository, n.d.", - "url": "https://github.com/johnezang/pithy", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M41", - "citation": "Compression.ru archive. “SHRINKER compressor description.” Official documentation or source repository, n.d.", - "url": "https://compression.ru/arctest/self/shrinker.htm", - "kind": "archived_project_page_or_source_mirror", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M42", - "citation": "oberhumer.com. “UCL data compression library.” Official documentation or source repository, n.d.", - "url": "http://www.oberhumer.com/opensource/ucl/", - "kind": "official_documentation_or_project_page", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M43", - "citation": "code.google.com. “WFLZ compression library.” Official documentation or source repository, n.d.", - "url": "https://code.google.com/p/wflz/", - "kind": "archived_project_page_or_source_mirror", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M44", - "citation": "ebiggers project contributors. “XPACK experimental compression format.” Official documentation or source repository, n.d.", - "url": "https://github.com/ebiggers/xpack", - "kind": "standard_or_specification", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M45", - "citation": "Ivan Tkatchev. “yalz77: Yet Another LZ77.” Official source repository, 2015.", - "url": "https://github.com/ivan-tkatchev/yalz77", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M46", - "citation": "web.archive.org. “Yappy compression algorithm.” Official documentation or source repository, n.d.", - "url": "https://web.archive.org/web/20110617184153/blog.gamedeff.com/?p=371", - "kind": "archived_project_page_or_source_mirror", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M47", - "citation": "google project contributors. “Snappy compression format.” Official documentation or source repository, n.d.", - "url": "https://github.com/google/snappy", - "kind": "standard_or_specification", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M48", - "citation": "Jarek Duda. “Asymmetric Numeral Systems.” arXiv:0902.0271, 2009.", - "url": "https://arxiv.org/abs/0902.0271", - "kind": "official_documentation_or_project_page", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M49", - "citation": "Cyan4973 project contributors. “FiniteStateEntropy.” Official documentation or source repository, n.d.", - "url": "https://github.com/Cyan4973/FiniteStateEntropy", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "web_verified_primary_or_official_2026-10-08" - }, - { - "id": "M50", - "citation": "David Huffman. “A Method for the Construction of Minimum-Redundancy Codes.” Proceedings of the IRE, 1952.", - "url": "https://doi.org/10.1109/JRPROC.1952.273898", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "doi_resolved_to_primary_publisher_2026-10-08" - }, - { - "id": "M51", - "citation": "Ian H. Witten, Radford M. Neal, John G. Cleary. “Arithmetic coding for data compression.” Communications of the ACM, 1987.", - "url": "https://doi.org/10.1145/214762.214771", - "kind": "research_paper_or_standard", - "source_slide": 2, - "verification_status": "publisher_blocked_403_not_a_dead_link" - }, - { - "id": "M52", - "citation": "Blosc Development Team. “Blosc: A blocking, shuffling and compression filter.” Official documentation or source repository, n.d.", - "url": "https://www.blosc.org/pages/blosc-in-depth/", - "kind": "official_documentation_or_project_page", - "source_slide": 2, - "verification_status": "web_verified_primary_or_official_2026-10-08" - }, - { - "id": "M53", - "citation": "jibsen project contributors. “BriefLZ fast compression library.” Official documentation or source repository, n.d.", - "url": "https://github.com/jibsen/brieflz", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M54", - "citation": "IlyaGrebnov project contributors. “libbsc: Block Sorting Compressor.” Official documentation or source repository, n.d.", - "url": "https://github.com/IlyaGrebnov/libbsc", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "web_verified_primary_or_official_2026-10-08" - }, - { - "id": "M55", - "citation": "sourceware.org. “bzip2 and libbzip2.” Official documentation or source repository, n.d.", - "url": "https://www.sourceware.org/bzip2/", - "kind": "official_documentation_or_project_page", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M56", - "citation": "kspalaiologos project contributors. “bzip3: a better and stronger successor to bzip2.” Official documentation or source repository, n.d.", - "url": "https://github.com/kspalaiologos/bzip3", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "web_verified_primary_or_official_2026-10-08" - }, - { - "id": "M57", - "citation": "sourceforge.net. “CRUSH compression library.” Official documentation or source repository, n.d.", - "url": "https://sourceforge.net/projects/crush/", - "kind": "official_documentation_or_project_page", - "source_slide": 2, - "verification_status": "publisher_blocked_403_not_a_dead_link" - }, - { - "id": "M58", - "citation": "fusiyuan2010 project contributors. “CSC lossless compression algorithm.” Official documentation or source repository, n.d.", - "url": "https://github.com/fusiyuan2010/CSC", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M59", - "citation": "ariya project contributors. “FastLZ compression library.” Official documentation or source repository, n.d.", - "url": "https://github.com/ariya/FastLZ", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M60", - "citation": "conor42 project contributors. “Fast LZMA2 compression library.” Official documentation or source repository, n.d.", - "url": "https://github.com/conor42/fast-lzma2", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M61", - "citation": "flanglet project contributors. “Kanzi: high-performance compression framework.” Official documentation or source repository, n.d.", - "url": "https://github.com/flanglet/kanzi-cpp", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M62", - "citation": "ebiggers project contributors. “libdeflate.” Official documentation or source repository, n.d.", - "url": "https://github.com/ebiggers/libdeflate", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "web_verified_primary_or_official_2026-10-08" - }, - { - "id": "M63", - "citation": "inikep project contributors. “Lizard (LZ5) fast compression.” Official documentation or source repository, n.d.", - "url": "https://github.com/inikep/lizard", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M64", - "citation": "lz4.org. “LZ4 fast compression mode.” Official documentation or source repository, n.d.", - "url": "https://lz4.org/", - "kind": "official_documentation_or_project_page", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M65", - "citation": "avaneev project contributors. “LZAV fast LZ77 compressor.” Official documentation or source repository, n.d.", - "url": "https://github.com/avaneev/lzav", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "web_verified_primary_or_official_2026-10-08" - }, - { - "id": "M66", - "citation": "software.schmorp.de. “LibLZF very fast compression library.” Official documentation or source repository, n.d.", - "url": "http://software.schmorp.de/pkg/liblzf.html", - "kind": "official_documentation_or_project_page", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M67", - "citation": "mbitsnbites project contributors. “LibLZG lightweight compression library.” Official documentation or source repository, n.d.", - "url": "https://github.com/mbitsnbites/liblzg", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M68", - "citation": "richgel999 project contributors. “LZHAM compression codec.” Official documentation or source repository, n.d.", - "url": "https://github.com/richgel999/lzham_codec", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M69", - "citation": "openzfs project contributors. “OpenZFS LZJB reference source.” Official documentation or source repository, n.d.", - "url": "https://github.com/openzfs/zfs/blob/master/module/zfs/lzjb.c", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M70", - "citation": "nongnu.org. “Lzlib data compression library.” Official documentation or source repository, n.d.", - "url": "https://www.nongnu.org/lzip/lzlib.html", - "kind": "official_documentation_or_project_page", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M71", - "citation": "7-zip.org. “PPMd statistical compression (PPMD8).” Official documentation or source repository, n.d.", - "url": "https://www.7-zip.org/", - "kind": "official_documentation_or_project_page", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M72", - "citation": "web.archive.org. “QuickLZ compression library.” Official documentation or source repository, n.d.", - "url": "https://web.archive.org/web/20160110073818/https://quicklz.com/", - "kind": "archived_project_page_or_source_mirror", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M73", - "citation": "libslz.org. “libslz fast gzip/deflate encoder.” Official documentation or source repository, n.d.", - "url": "http://www.libslz.org/", - "kind": "official_documentation_or_project_page", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M74", - "citation": "nemequ project contributors. “Tornado compressor reference source.” Official documentation or source repository, n.d.", - "url": "https://github.com/nemequ/tornado", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M75", - "citation": "zlib.net. “zlib compression library.” Official documentation or source repository, n.d.", - "url": "https://zlib.net/", - "kind": "official_documentation_or_project_page", - "source_slide": 2, - "verification_status": "web_verified_primary_or_official_2026-10-08" - }, - { - "id": "M76", - "citation": "zlib-ng project contributors. “zlib-ng optimized zlib fork.” Official documentation or source repository, n.d.", - "url": "https://github.com/zlib-ng/zlib-ng", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "web_verified_primary_or_official_2026-10-08" - }, - { - "id": "M77", - "citation": "richox project contributors. “libzling compression library.” Official documentation or source repository, n.d.", - "url": "https://github.com/richox/libzling", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M78", - "citation": "hellobertrand project contributors. “ZXC compression library.” Official documentation or source repository, n.d.", - "url": "https://github.com/hellobertrand/zxc", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "link_reachable_2026-10-08_content_from_ppt_notes" - }, - { - "id": "M79", - "citation": "BrianPugh project contributors. “TAMP tiny automatic meta-protocol compressor.” Official documentation or source repository, n.d.", - "url": "https://github.com/BrianPugh/tamp", - "kind": "official_source_repository", - "source_slide": 2, - "verification_status": "web_verified_primary_or_official_2026-10-08" - }, - { - "id": "S01", - "citation": "Apache Software Foundation. “Apache Parquet Encodings: BYTE_STREAM_SPLIT.” Official documentation, n.d.", - "url": "https://parquet.apache.org/docs/file-format/data-pages/encodings/#byte-stream-split-byte_stream_split-9", - "kind": "official_documentation_or_project_page", - "source_slide": null, - "verification_status": "web_verified_primary_or_official_2026-10-08" - }, - { - "id": "O01", - "citation": "Ruiyuan Li, Zheng Li, Yi Wu, Chao Chen, Yu Zheng. “Elf: Erasing-Based Lossless Floating-Point Compression.” Proceedings of the VLDB Endowment, 2023.", - "url": "https://doi.org/10.14778/3587136.3587149", - "kind": "paper", - "source_slide": 3, - "verification_status": "ppt_note_primary_source" - }, - { - "id": "O02", - "citation": "Azim Afroozeh, Leonardo X. Kuffo, Peter Boncz. “ALP: Adaptive Lossless floating-Point Compression.” Proceedings of the ACM on Management of Data, 2023.", - "url": "https://doi.org/10.1145/3626717", - "kind": "paper", - "source_slide": 3, - "verification_status": "ppt_note_primary_source" - }, - { - "id": "O03", - "citation": "Yihao Liu, Xinyu Zeng, Huanchen Zhang. “LeCo: Lightweight Compression via Learning Serial Correlations.” Proceedings of the ACM on Management of Data, 2024.", - "url": "https://doi.org/10.1145/3639320", - "kind": "paper", - "source_slide": 3, - "verification_status": "ppt_note_primary_source" - }, - { - "id": "O04", - "citation": "Jinzhao Xiao, Wendi He, Shaoxu Song, Xiangdong Huang, Chen Wang, Jianmin Wang. “REGER: Reordering Time Series Data for Regression Encoding.” 2024 IEEE 40th International Conference on Data Engineering (ICDE), 2024.", - "url": "https://doi.org/10.1109/ICDE60146.2024.00100", - "kind": "paper", - "source_slide": 3, - "verification_status": "ppt_note_primary_source" - }, - { - "id": "O05", - "citation": "Jinzhao Xiao, Zihan Guo, Shaoxu Song. “BOS: Bit-Packing with Outlier Separation.” 2025 IEEE 41st International Conference on Data Engineering (ICDE), 2025.", - "url": "https://doi.org/10.1109/ICDE65448.2025.00107", - "kind": "paper", - "source_slide": 3, - "verification_status": "ppt_note_primary_source" - }, - { - "id": "O06", - "citation": "Daniel Lemire, Leonid Boytsov, Nathan Kurz. “SIMD compression and the intersection of sorted integers.” Software: Practice and Experience, 2015.", - "url": "https://doi.org/10.1002/spe.2326", - "kind": "paper", - "source_slide": 3, - "verification_status": "ppt_note_primary_source" - }, - { - "id": "O07", - "citation": "Apache Software Foundation. “Apache Parquet DELTA_BINARY_PACKED specification.” Official documentation or source repository, n.d. (accessed 2026-09-11).", - "url": "https://parquet.apache.org/docs/file-format/data-pages/encodings/#delta-encoding-delta_binary_packed-5", - "kind": "official_documentation", - "source_slide": 3, - "verification_status": "ppt_note_official_source" - }, - { - "id": "O08", - "citation": "Tuomas Pelkonen, Scott Franklin, Justin Teller, Paul Cavallaro, Qi Huang, Justin Meza, Kaushik Veeraraghavan. “Gorilla.” Proceedings of the VLDB Endowment, 2015.", - "url": "https://doi.org/10.14778/2824032.2824078", - "kind": "paper", - "source_slide": 3, - "verification_status": "ppt_note_primary_source" - }, - { - "id": "O09", - "citation": "Panagiotis Liakos, Katia Papakonstantinopoulou, Yannis Kotidis. “Chimp.” Proceedings of the VLDB Endowment, 2022.", - "url": "https://doi.org/10.14778/3551793.3551852", - "kind": "paper", - "source_slide": 3, - "verification_status": "ppt_note_primary_source" - }, - { - "id": "O10", - "citation": "Daniel Abadi, Samuel Madden, Miguel Ferreira. “Integrating compression and execution in column-oriented database systems.” Proceedings of the 2006 ACM SIGMOD international conference on Management of data, 2006.", - "url": "https://doi.org/10.1145/1142473.1142548", - "kind": "paper", - "source_slide": 3, - "verification_status": "ppt_note_primary_source" - }, - { - "id": "O11", - "citation": "Protocol Buffers project. “Protocol Buffers encoding: ZigZag.” Official documentation or source repository, n.d. (accessed 2026-09-11).", - "url": "https://protobuf.dev/programming-guides/encoding/#signed-ints", - "kind": "official_documentation", - "source_slide": 3, - "verification_status": "verified_primary" - }, - { - "id": "O12", - "citation": "Michael Burrows, David J. Wheeler. “A Block-sorting Lossless Data Compression Algorithm.” Digital Systems Research Center Research Report 124, 1994.", - "url": "https://www.cl.cam.ac.uk/teaching/2526/Bioinfo/papers/burrows_wheeler.pdf", - "kind": "paper", - "source_slide": 3, - "verification_status": "verified_primary" - }, - { - "id": "O13", - "citation": "Jon Louis Bentley, Daniel D. Sleator, Robert E. Tarjan, Victor K. Wei. “A locally adaptive data compression scheme.” Communications of the ACM, 1986.", - "url": "https://doi.org/10.1145/5684.5688", - "kind": "paper", - "source_slide": 3, - "verification_status": "primary_url_preserved_resolver_unavailable" - }, - { - "id": "O14", - "citation": "Apache Software Foundation. “Run Length Encoding / Bit-Packing Hybrid.” Official documentation or source repository, n.d. (accessed 2026-09-11).", - "url": "https://parquet.apache.org/docs/file-format/data-pages/encodings/#run-length-encoding-bit-packing-hybrid-rle-3", - "kind": "official_documentation", - "source_slide": 3, - "verification_status": "ppt_note_official_source" - }, - { - "id": "O15", - "citation": "J. Ziv, A. Lempel. “A universal algorithm for sequential data compression.” IEEE Transactions on Information Theory, 1977.", - "url": "https://doi.org/10.1109/TIT.1977.1055714", - "kind": "paper", - "source_slide": 3, - "verification_status": "primary_url_preserved_resolver_unavailable" - }, - { - "id": "O16", - "citation": "facebook project contributors. “Zstandard compression format and implementation documentation.” Official documentation or source repository, n.d. (accessed 2026-09-11).", - "url": "https://github.com/facebook/zstd", - "kind": "official_documentation", - "source_slide": 3, - "verification_status": "ppt_note_official_source" - }, - { - "id": "O17", - "citation": "lz4 project contributors. “LZ4 Block Format Description.” Official documentation or source repository, n.d. (accessed 2026-09-11).", - "url": "https://github.com/lz4/lz4/blob/dev/doc/lz4_Block_format.md", - "kind": "official_specification", - "source_slide": 3, - "verification_status": "verified_primary" - }, - { - "id": "O18", - "citation": "Daniel Lemire, Nathan Kurz, Christoph Rupp. “Stream VByte : Faster byte-oriented integer compression.” Information Processing Letters, 2018.", - "url": "https://doi.org/10.1016/j.ipl.2017.09.011", - "kind": "paper", - "source_slide": 3, - "verification_status": "ppt_note_primary_source" - }, - { - "id": "O19", - "citation": "P. Deutsch. “DEFLATE Compressed Data Format Specification version 1.3.” RFC Editor, 1996.", - "url": "https://doi.org/10.17487/RFC1951", - "kind": "official_specification", - "source_slide": 3, - "verification_status": "primary_url_preserved_resolver_unavailable" - }, - { - "id": "O20", - "citation": "David Huffman. “A Method for the Construction of Minimum-Redundancy Codes.” Proceedings of the IRE, 1952.", - "url": "https://doi.org/10.1109/JRPROC.1952.273898", - "kind": "paper", - "source_slide": 3, - "verification_status": "ppt_note_primary_source" - }, - { - "id": "O21", - "citation": "Ian H. Witten, Radford M. Neal, John G. Cleary. “Arithmetic coding for data compression.” Communications of the ACM, 1987.", - "url": "https://doi.org/10.1145/214762.214771", - "kind": "paper", - "source_slide": 3, - "verification_status": "ppt_note_primary_source" - }, - { - "id": "O22", - "citation": "Jarek Duda. “Asymmetric Numeral Systems: Entropy Coding Combining Speed of Huffman Coding with Compression Rate of Arithmetic Coding.” arXiv:1311.2540, 2013.", - "url": "https://arxiv.org/abs/1311.2540", - "kind": "paper", - "source_slide": 3, - "verification_status": "verified_primary" - }, - { - "id": "O23", - "citation": "Steven Claggett, Sahar Azimi, Martin Burtscher. “SPDP: An Automatically Synthesized Lossless Compression Algorithm for Floating-Point Data.” 2018 Data Compression Conference, 2018.", - "url": "https://doi.org/10.1109/DCC.2018.00042", - "kind": "paper", - "source_slide": 3, - "verification_status": "ppt_note_primary_source" - }, - { - "id": "O24", - "citation": "Y. Collet. “Zstandard Compression and the 'application/zstd' Media Type.” RFC Editor, 2021.", - "url": "https://doi.org/10.17487/RFC8878", - "kind": "official_specification", - "source_slide": 3, - "verification_status": "ppt_note_primary_source" - }, - { - "id": "O25", - "citation": "Blosc Development Team. “Blosc in-depth: blocking, shuffling and compression.” Official documentation or source repository, n.d. (accessed 2026-09-11).", - "url": "https://www.blosc.org/pages/blosc-in-depth/", - "kind": "official_documentation", - "source_slide": 3, - "verification_status": "ppt_note_official_source" - }, - { - "id": "O26", - "citation": "P. Deutsch. “GZIP file format specification version 4.3.” RFC Editor, 1996.", - "url": "https://doi.org/10.17487/RFC1952", - "kind": "official_specification", - "source_slide": 3, - "verification_status": "ppt_note_primary_source" - }, - { - "id": "O27", - "citation": "Meta Platforms, Inc. “OpenZL.” Official source repository.", - "url": "https://github.com/facebook/openzl", - "kind": "official_source", - "verification_status": "primary_source_verified" - }, - { - "id": "O28", - "citation": "Fabrice Bellard. “NNCP: Lossless Data Compression with Neural Networks.” Official project page.", - "url": "https://bellard.org/nncp/", - "kind": "official_source", - "verification_status": "primary_source_verified" - }, - { - "id": "O29", - "citation": "Byron Knoll. “cmix version 21.” Official source repository.", - "url": "https://github.com/byronknoll/cmix", - "kind": "official_source", - "verification_status": "primary_source_verified" - } - ], - "source": { - "presentation": "编码压缩基准.pptx", - "presentation_sha256": "4a5bed20f2b9f7ba18bf18569ad50b445e3bc8bfe6fe3020c5f68df424e1888e", - "notes_slides": [ - 2, - 3 - ], - "ppt_method_aliases": [ - { - "index": 1, - "ppt_name": "Chimp", - "reference_ids": [ - "M01" - ], - "family": "CHIMP" - }, - { - "index": 2, - "ppt_name": "Chimp128", - "reference_ids": [ - "M01" - ], - "family": "Chimp128" - }, - { - "index": 3, - "ppt_name": "Pongo", - "reference_ids": [ - "M02" - ], - "family": "Pongo" - }, - { - "index": 4, - "ppt_name": "Gorilla", - "reference_ids": [ - "M03" - ], - "family": "GORILLA" - }, - { - "index": 5, - "ppt_name": "2DIFF", - "reference_ids": [ - "M04" - ], - "family": "TS_2DIFF" - }, - { - "index": 6, - "ppt_name": "BOS", - "reference_ids": [ - "M05" - ], - "family": "BOS" - }, - { - "index": 7, - "ppt_name": "REGER", - "reference_ids": [ - "M06" - ], - "family": "REGER" - }, - { - "index": 8, - "ppt_name": "SVB", - "reference_ids": [ - "M07" - ], - "family": "Stream VByte" - }, - { - "index": 9, - "ppt_name": "FastPFor", - "reference_ids": [ - "M08" - ], - "family": "FastPFor" - }, - { - "index": 10, - "ppt_name": "Simple8b", - "reference_ids": [ - "M09" - ], - "family": "Simple8b" - }, - { - "index": 11, - "ppt_name": "RLE", - "reference_ids": [ - "M10" - ], - "family": "RLE" - }, - { - "index": 12, - "ppt_name": "RLBE", - "reference_ids": [ - "M11" - ], - "family": "RLBE" - }, - { - "index": 13, - "ppt_name": "Sprintz", - "reference_ids": [ - "M12" - ], - "family": "SPRINTZ" - }, - { - "index": 14, - "ppt_name": "Dict.", - "reference_ids": [ - "M13" - ], - "family": "Dictionary" - }, - { - "index": 15, - "ppt_name": "DBP", - "reference_ids": [ - "M14" - ], - "family": "DBP" - }, - { - "index": 16, - "ppt_name": "ALP", - "reference_ids": [ - "M15" - ], - "family": "ALP" - }, - { - "index": 17, - "ppt_name": "LECO", - "reference_ids": [ - "M16" - ], - "family": "LECO" - }, - { - "index": 18, - "ppt_name": "Camel", - "reference_ids": [ - "M17" - ], - "family": "Camel" - }, - { - "index": 19, - "ppt_name": "Elf", - "reference_ids": [ - "M18" - ], - "family": "Elf" - }, - { - "index": 20, - "ppt_name": "Elf+", - "reference_ids": [ - "M19" - ], - "family": "Elf+" - }, - { - "index": 21, - "ppt_name": "FPC", - "reference_ids": [ - "M20" - ], - "family": "FPC" - }, - { - "index": 22, - "ppt_name": "PFPC", - "reference_ids": [ - "M21" - ], - "family": "PFPC" - }, - { - "index": 23, - "ppt_name": "SPDP", - "reference_ids": [ - "M22" - ], - "family": "SPDP" - }, - { - "index": 24, - "ppt_name": "CORAD", - "reference_ids": [ - "M23" - ], - "family": "CORAD" - }, - { - "index": 25, - "ppt_name": "BUFF", - "reference_ids": [ - "M24" - ], - "family": "BUFF" - }, - { - "index": 26, - "ppt_name": "FPZIP", - "reference_ids": [ - "M25" - ], - "family": "FPZIP" - }, - { - "index": 27, - "ppt_name": "FPCmp-R", - "reference_ids": [ - "M26" - ], - "family": "FPCompress DPratio" - }, - { - "index": 28, - "ppt_name": "FPCmp-S", - "reference_ids": [ - "M26" - ], - "family": "FPCompress DPspeed" - }, - { - "index": 29, - "ppt_name": "gzip", - "reference_ids": [ - "M27" - ], - "family": "GZIP" - }, - { - "index": 30, - "ppt_name": "Brotli", - "reference_ids": [ - "M28" - ], - "family": "Brotli" - }, - { - "index": 31, - "ppt_name": "Zstd", - "reference_ids": [ - "M29" - ], - "family": "ZSTD" - }, - { - "index": 32, - "ppt_name": "LZ4", - "reference_ids": [ - "M30" - ], - "family": "LZ4" - }, - { - "index": 33, - "ppt_name": "XZ", - "reference_ids": [ - "M31" - ], - "family": "XZ" - }, - { - "index": 34, - "ppt_name": "Density", - "reference_ids": [ - "M32" - ], - "family": "Density" - }, - { - "index": 35, - "ppt_name": "Gipfeli", - "reference_ids": [ - "M33" - ], - "family": "Gipfeli" - }, - { - "index": 37, - "ppt_name": "LZMAT", - "reference_ids": [ - "M35" - ], - "family": "LZMAT" - }, - { - "index": 38, - "ppt_name": "LZO1", - "reference_ids": [ - "M36" - ], - "family": "LZO1" - }, - { - "index": 39, - "ppt_name": "LZO2A", - "reference_ids": [ - "M36" - ], - "family": "LZO2A" - }, - { - "index": 40, - "ppt_name": "LZRW", - "reference_ids": [ - "M37" - ], - "family": "LZRW" - }, - { - "index": 41, - "ppt_name": "LZSSE", - "reference_ids": [ - "M38" - ], - "family": "LZSSE" - }, - { - "index": 42, - "ppt_name": "LZVN", - "reference_ids": [ - "M39" - ], - "family": "LZVN" - }, - { - "index": 43, - "ppt_name": "Pithy", - "reference_ids": [ - "M40" - ], - "family": "Pithy" - }, - { - "index": 44, - "ppt_name": "Shrinker", - "reference_ids": [ - "M41" - ], - "family": "Shrinker" - }, - { - "index": 45, - "ppt_name": "NRV", - "reference_ids": [ - "M42" - ], - "family": "UCL NRV" - }, - { - "index": 46, - "ppt_name": "WFLZ", - "reference_ids": [ - "M43" - ], - "family": "WFLZ" - }, - { - "index": 47, - "ppt_name": "XPACK", - "reference_ids": [ - "M44" - ], - "family": "XPACK" - }, - { - "index": 48, - "ppt_name": "YALZ77", - "reference_ids": [ - "M45" - ], - "family": "YALZ77" - }, - { - "index": 49, - "ppt_name": "Yappy", - "reference_ids": [ - "M46" - ], - "family": "Yappy" - }, - { - "index": 50, - "ppt_name": "Snappy", - "reference_ids": [ - "M47" - ], - "family": "SNAPPY" - }, - { - "index": 51, - "ppt_name": "ANS", - "reference_ids": [ - "M48" - ], - "family": "ANS" - }, - { - "index": 52, - "ppt_name": "FSE", - "reference_ids": [ - "M49" - ], - "family": "FSE" - }, - { - "index": 53, - "ppt_name": "Huffman", - "reference_ids": [ - "M50" - ], - "family": "Huffman" - }, - { - "index": 54, - "ppt_name": "Huff0", - "reference_ids": [ - "M49" - ], - "family": "Huff0" - }, - { - "index": 55, - "ppt_name": "Arithmetic", - "reference_ids": [ - "M51" - ], - "family": "Arithmetic" - }, - { - "index": 56, - "ppt_name": "Blosc", - "reference_ids": [ - "M52" - ], - "family": "Blosc" - }, - { - "index": 57, - "ppt_name": "Blosc2", - "reference_ids": [ - "M52" - ], - "family": "Blosc2" - }, - { - "index": 58, - "ppt_name": "BriefLZ", - "reference_ids": [ - "M53" - ], - "family": "BriefLZ" - }, - { - "index": 59, - "ppt_name": "BSC", - "reference_ids": [ - "M54" - ], - "family": "BSC" - }, - { - "index": 60, - "ppt_name": "Bzip2", - "reference_ids": [ - "M55" - ], - "family": "Bzip2" - }, - { - "index": 61, - "ppt_name": "Bzip3", - "reference_ids": [ - "M56" - ], - "family": "Bzip3" - }, - { - "index": 62, - "ppt_name": "Crush", - "reference_ids": [ - "M57" - ], - "family": "Crush" - }, - { - "index": 63, - "ppt_name": "CSC", - "reference_ids": [ - "M58" - ], - "family": "CSC" - }, - { - "index": 64, - "ppt_name": "FastLZ", - "reference_ids": [ - "M59" - ], - "family": "FastLZ" - }, - { - "index": 65, - "ppt_name": "FastLZMA2", - "reference_ids": [ - "M60" - ], - "family": "FastLZMA2" - }, - { - "index": 66, - "ppt_name": "Kanzi", - "reference_ids": [ - "M61" - ], - "family": "Kanzi" - }, - { - "index": 67, - "ppt_name": "Libdeflate", - "reference_ids": [ - "M62" - ], - "family": "Libdeflate" - }, - { - "index": 68, - "ppt_name": "Lizard", - "reference_ids": [ - "M63" - ], - "family": "Lizard" - }, - { - "index": 69, - "ppt_name": "LZ4 Fast", - "reference_ids": [ - "M64" - ], - "family": "LZ4 Fast" - }, - { - "index": 70, - "ppt_name": "LZ4HC", - "reference_ids": [ - "M64" - ], - "family": "LZ4HC" - }, - { - "index": 71, - "ppt_name": "LZAV", - "reference_ids": [ - "M65" - ], - "family": "LZAV" - }, - { - "index": 72, - "ppt_name": "LZF", - "reference_ids": [ - "M66" - ], - "family": "LZF" - }, - { - "index": 73, - "ppt_name": "LZFSE", - "reference_ids": [ - "M39" - ], - "family": "LZFSE" - }, - { - "index": 74, - "ppt_name": "LZG", - "reference_ids": [ - "M67" - ], - "family": "LZG" - }, - { - "index": 75, - "ppt_name": "LZHAM", - "reference_ids": [ - "M68" - ], - "family": "LZHAM" - }, - { - "index": 76, - "ppt_name": "LZJB", - "reference_ids": [ - "M69" - ], - "family": "LZJB" - }, - { - "index": 77, - "ppt_name": "Lzlib", - "reference_ids": [ - "M70" - ], - "family": "Lzlib" - }, - { - "index": 78, - "ppt_name": "PPMD8", - "reference_ids": [ - "M71" - ], - "family": "PPMD8" - }, - { - "index": 79, - "ppt_name": "QuickLZ", - "reference_ids": [ - "M72" - ], - "family": "QuickLZ" - }, - { - "index": 80, - "ppt_name": "SLZ", - "reference_ids": [ - "M73" - ], - "family": "SLZ" - }, - { - "index": 81, - "ppt_name": "Tornado", - "reference_ids": [ - "M74" - ], - "family": "Tornado" - }, - { - "index": 82, - "ppt_name": "Zlib", - "reference_ids": [ - "M75" - ], - "family": "Zlib" - }, - { - "index": 83, - "ppt_name": "Zlib-ng", - "reference_ids": [ - "M76" - ], - "family": "Zlib-ng" - }, - { - "index": 84, - "ppt_name": "Zling", - "reference_ids": [ - "M77" - ], - "family": "Zling" - }, - { - "index": 85, - "ppt_name": "ZXC", - "reference_ids": [ - "M78" - ], - "family": "ZXC" - }, - { - "index": 86, - "ppt_name": "TAMP", - "reference_ids": [ - "M79" - ], - "family": "TAMP" - } - ], - "additional_measured_families": [ - { - "family": "BSS+LZ4", - "affiliation": "Project-specific composition of a Parquet BYTE_STREAM_SPLIT-style transform and LZ4 framing.", - "reference_ids": [ - "S01", - "M30" - ] - }, - { - "family": "LZO1X", - "affiliation": "Variant of PPT LZO1.", - "reference_ids": [ - "M36" - ] - }, - { - "family": "Slz_deflate", - "affiliation": "Wrapper-specific split of PPT SLZ.", - "reference_ids": [ - "M73" - ] - }, - { - "family": "Slz_gzip", - "affiliation": "Wrapper-specific split of PPT SLZ.", - "reference_ids": [ - "M73", - "M27" - ] - }, - { - "family": "Slz_zlib", - "affiliation": "Wrapper-specific split of PPT SLZ.", - "reference_ids": [ - "M73", - "M75" - ] - }, - { - "family": "Sub-column", - "affiliation": "Project-specific composition.", - "reference_ids": [ - "M09", - "M10", - "M13" - ] - }, - { - "family": "UCL NRV2b", - "affiliation": "Variant of PPT NRV.", - "reference_ids": [ - "M42" - ] - } - ], - "description_note": "Method references follow presentation Notes (slides 2–3). Operator descriptions follow the final 30-family implementation catalog, preserving source-backed boundaries and reusing presentation citations with additional official sources.", - "operator_catalog": { - "catalog_version": "operator-implementation-layers-v1", - "files": [ - "docs/operator_102/review_current.json", - "benchmark/reference/actions/operator_implementations_layered_v1.json", - "docs/operator_102/interface_revision_status.json", - "docs/operator_102/source_inputs/comparison_01_05.json", - "docs/operator_102/source_inputs/comparison_06_10.json", - "docs/operator_102/source_inputs/comparison_11_20.json", - "docs/operator_102/source_inputs/comparison_21_30.json", - "docs/operator_102/source_inputs/comparison_27_root.json", - "docs/operator_102/source_inputs/comparison_28_root.json", - "docs/operator_102/source_inputs/comparison_29_30_final.json", - "deploy/huggingface-compressionbenchmark/qa/operator_catalog_content.json", - "deploy/huggingface-compressionbenchmark/qa/method_catalog_content.json" - ], - "sha256": { - "docs/operator_102/review_current.json": "ac5e403edd361c04b4f0f4d33de3b089456a3f9cc958062704b050ddfc16fcaf", - "benchmark/reference/actions/operator_implementations_layered_v1.json": "09e41517b2e8aea04067546aa24f6d0e1dc4e6eef1475339114283a0c9acf252", - "docs/operator_102/interface_revision_status.json": "4075f9783b85767fda48c82cfda125a27d3f649529d6ae5ce989924daddc844f", - "docs/operator_102/source_inputs/comparison_01_05.json": "78f88580fd1030969a4ff012add30a4f8d751dd7bf0ab1539fc87e48746911ae", - "docs/operator_102/source_inputs/comparison_06_10.json": "bfd843823e76fba92a7bfe310ac0b1de64a02642e66bb97b273c74ad0ee5eff9", - "docs/operator_102/source_inputs/comparison_11_20.json": "6ccff1fff27f874261fa1f07c1f344c79692705a398e9a49b417cfc9193b8135", - "docs/operator_102/source_inputs/comparison_21_30.json": "f1c09afcba5f0c0eb7a183dce94439e53be947c00c66a48268d8ceb1c7fcc4cf", - "docs/operator_102/source_inputs/comparison_27_root.json": "6de78e3539cabfd9495cb5d34f5dc3c8df22dcc5a51c8a9007e71857e6d048fd", - "docs/operator_102/source_inputs/comparison_28_root.json": "7ab1ac8902d079907d36b14d951b70faf442010946927ecf10423703d99c2165", - "docs/operator_102/source_inputs/comparison_29_30_final.json": "4913028f99abcc70e5915053c1603ed65f85fa7450f3a056ad409f845feca7c6", - "deploy/huggingface-compressionbenchmark/qa/operator_catalog_content.json": "8ffdeda95b1599cdb3fcb846471cae39a9a51f6fbb4c3c94685b6dfd9bd49b53", - "deploy/huggingface-compressionbenchmark/qa/method_catalog_content.json": "5bc289539e6d3298cb4f2b6006c5503b7973f52707cf10219f878c2cabcc33df" - }, - "boundaries": [ - "A013 is Minimum Frame of Reference only.", - "A048 is pure byte-plane split; minimum, scale, and width selection remain outside this mechanism.", - "STOP, Concat, and continuous Split are special graph operations and are not counted among the 30 families or 102 entries.", - "Complete backends and compatibility adapters are labeled separately from mechanisms.", - "A110 has a verified independent NNCP/NNP1 path but is not integrated into the graph entropy edge or default method schedule.", - "A111 has an independent three-stream C++ interface; the original fused FastPFor path remains separate." - ] - } - }, - "operator_summary": { - "family_count": 30, - "entry_count": 102, - "mechanism_count": 67, - "configuration_count": 7, - "composition_count": 14, - "adapter_count": 14 - } -} +{"methods":[{"description":"ALP adaptively converts decimal-valued floating-point data to integers, separates values that cannot use the selected scale as exceptions, and frame-of-reference packs the integer stream. Decoding restores the scaled values and their exceptions.","display_name":"ALP","family":"ALP","reference_ids":["M15"]},{"description":"Asymmetric Numeral Systems encodes a symbol sequence into an integer state using transitions determined by symbol probabilities. Frequent symbols consume less information, and decoding reverses those state transitions.","display_name":"ANS","family":"ANS","reference_ids":["M48"]},{"description":"Arithmetic coding represents a whole symbol sequence by repeatedly narrowing a numeric interval according to the current probability model. Range-coder implementations emit bytes as the interval state is renormalized.","display_name":"Arithmetic","family":"Arithmetic","reference_ids":["M51"]},{"description":"Bit-Packing with Outlier Separation isolates large residuals so the remaining values can use narrow packed widths. In the latest current-135 profile, BOS2 applies delta, unsigned ZigZag, deterministic exception separation, an independent minimum for each partition, and independent bit packing; positions and all reconstruction metadata are charged. Historical BOS variants remain preserved.","display_name":"BOS","family":"BOS","reference_ids":["M05"]},{"description":"libbsc applies block-sorting transforms, symbol recoding, and a statistical entropy back end to expose and encode repeated contexts. Presets select the preprocessing and coding choices for each block.","display_name":"BSC","family":"BSC","reference_ids":["M54"]},{"description":"This project composition transposes floating-point words into separate byte planes, using a Parquet BYTE_STREAM_SPLIT-style layout so bytes in the same position become adjacent. LZ4 then compresses the transformed stream.","display_name":"Byte-stream split + LZ4","family":"BSS+LZ4","notes":"S01 documents the BYTE_STREAM_SPLIT reference design and PPT M30 documents LZ4; the measured composition is the project's implementation.","reference_ids":["S01","M30"]},{"description":"BUFF converts exactly representable decimal values to scaled integers, subtracts a block or column reference, and lays out the remaining low-width information in byte or bit planes. The project path retains a raw-bit fallback when exact decimal scaling is not available.","display_name":"BUFF","family":"BUFF","reference_ids":["M24"]},{"description":"Blosc1 splits input into blocks and shuffles bytes by their position within each numeric word before invoking an inner codec. The measured configuration uses BloscLZ, and the Blosc container records the information needed to unshuffle and decode the blocks.","display_name":"Blosc","family":"Blosc","reference_ids":["M52"]},{"description":"Blosc2 combines block partitioning, reversible filters, an inner compressor, and container metadata. The measured configuration uses byte shuffle followed by BloscLZ, arranging like-positioned bytes together before LZ matching.","display_name":"Blosc2","family":"Blosc2","reference_ids":["M52"]},{"description":"BriefLZ searches previously decoded bytes for repeated strings and emits Lempel-Ziv distance/length references alongside literals. It packs the resulting symbols with a small universal code rather than a full entropy-coding stage.","display_name":"BriefLZ","family":"BriefLZ","reference_ids":["M53"]},{"description":"Brotli combines LZ77 back-references with a built-in static dictionary and word transforms. Context models select prefix-code tables, and Huffman-style prefix codes encode literals, lengths, distances, and block-switch commands.","display_name":"Brotli","family":"Brotli","reference_ids":["M28"]},{"description":"bzip2 applies run-length preprocessing and the Burrows-Wheeler block transform, then converts the transformed symbols with move-to-front and a second run-length stage. Multiple Huffman tables encode the final symbol stream.","display_name":"Bzip2","family":"Bzip2","reference_ids":["M55"]},{"description":"bzip3 combines an LZP pass based on LZ77 matching and context prediction with run-length coding and a suffix-array Burrows-Wheeler transform. An order-0 context-mixing entropy coder encodes the transformed block.","display_name":"Bzip3","family":"Bzip3","reference_ids":["M56"]},{"description":"Chimp XORs a floating-point bit pattern with a previous reference and chooses compact control forms for unchanged values or nonzero residuals. Quantized leading-zero information and trailing-zero-aware significant-bit fields reduce the XOR metadata cost.","display_name":"Chimp","family":"CHIMP","reference_ids":["M01"]},{"description":"CORAD exploits correlation across time-series signals by representing values with a sparse shared dictionary and compact residual information. The benchmark's corad_tricklet path is a project adaptation that uses exact normalization and atom selection before its outer payload coding.","display_name":"CORAD","family":"CORAD","reference_ids":["M23"]},{"description":"CSC is an LZMA-inspired block compressor that preprocesses byte streams, searches a dictionary for repeated strings, and parses the input into literals and LZ77-style matches. Context-adaptive range coding encodes the resulting symbols.","display_name":"CSC","family":"CSC","reference_ids":["M58"]},{"description":"Camel classifies floating-point values into compact decimal or integer-delta records when possible and otherwise retains a literal/XOR representation. Tags and residual fields are packed into a chunked bitstream that lets the decoder reproduce the chosen representation.","display_name":"Camel","family":"Camel","reference_ids":["M17"]},{"description":"Chimp128 searches a history of 128 recent values for an XOR reference that yields more trailing zeros, instead of always selecting the immediately preceding value. It stores the reference index and the residual in Chimp-style control and significant-bit fields.","display_name":"Chimp128","family":"Chimp128","reference_ids":["M01"]},{"description":"CRUSH is an LZ77 compressor: it searches an earlier window for repeated byte strings and replaces them with distance/length copy tokens, while unmatched bytes remain literals. Its presets vary the match search and parsing effort.","display_name":"Crush","family":"Crush","reference_ids":["M57"]},{"description":"Project kernel that stores the first int64 value and encodes each subsequent delta with ZigZag + ULEB128. The Parquet DELTA_BINARY_PACKED specification is a reference design only; this implementation does not use Parquet blocks, miniblocks, minimum-delta frames, or miniblock bit packing.","display_name":"DBP (Delta + ZigZag + ULEB128)","family":"DBP","notes":"Float benchmark rows preprocess values to integer lanes outside the timed kernel; scale and mode metadata are not part of this DBP payload.","reference_ids":["M14"]},{"description":"The measured Density path uses the Lion algorithm. Lion performs multiple swapped dictionary lookups and prediction steps, adapts its coding decisions to the current block, and rank-entropy-codes the resulting symbols.","display_name":"Density","family":"Density","reference_ids":["M32"]},{"description":"Dictionary encoding stores each distinct column value once and replaces occurrences with compact dictionary identifiers. The decoder reconstructs the column by looking up those identifiers in the stored dictionary.","display_name":"Dictionary encoding","family":"Dictionary","reference_ids":["M13"]},{"description":"Elf conditionally clears predictable mantissa suffix bits using decimal precision information, then XOR-encodes the modified floating-point words. Erasure metadata allows the decoder to recover the original represented values when the erasure rule applies.","display_name":"Elf","family":"Elf","reference_ids":["M18"]},{"description":"Elf+ conditionally erases predictable floating-point mantissa suffix bits and then XORs consecutive words. It reuses an earlier leading/trailing-zero window when the new XOR fits, or transmits a new significant-bit window when it does not.","display_name":"Elf+","family":"Elf+","reference_ids":["M19"]},{"description":"FPC predicts floating-point bit patterns with finite-context and differential finite-context tables, choosing the predictor with the smaller XOR residual. A predictor flag and leading-zero-byte count let it store only the residual bytes that remain significant.","display_name":"FPC","family":"FPC","reference_ids":["M20"]},{"description":"The FPCompress DPratio configuration applies a ratio-oriented sequence of reversible transforms, including dictionary formation, differencing, bit packing, and run-length coding, to expose repeated and low-width structure in floating-point data.","display_name":"FPCompress DPratio","family":"FPCompress DPratio","reference_ids":["M26"]},{"description":"The FPCompress DPspeed configuration uses a shorter data-parallel path centered on differencing and bit packing, reducing values to narrow residual fields while limiting transform overhead.","display_name":"FPCompress DPspeed","family":"FPCompress DPspeed","reference_ids":["M26"]},{"description":"FPZIP predicts each floating-point value from nearby samples, represents the prediction difference compactly, and entropy-codes the resulting symbols. The benchmark uses the project's measured FPZIP path.","display_name":"FPZIP","family":"FPZIP","reference_ids":["M25"]},{"description":"Finite State Entropy is a table-driven Asymmetric Numeral Systems coder. It normalizes symbol frequencies and uses a coding state, transition tables, and packed low-order bits to encode and decode symbols near their modeled entropy.","display_name":"FSE","family":"FSE","reference_ids":["M49"]},{"description":"FastLZ uses byte-aligned LZ77 coding. A hash-based match finder turns repeated substrings into compact offset/length tokens and writes unmatched input as literal runs.","display_name":"FastLZ","family":"FastLZ","reference_ids":["M59"]},{"description":"Fast LZMA2 retains LZMA2's dictionary matches, literal contexts, and range-coded symbol stream. Its encoder uses a buffered radix match finder and block state management to produce ordinary LZMA2 data.","display_name":"FastLZMA2","family":"FastLZMA2","reference_ids":["M60"]},{"description":"FastPFOR packs most integers in a block at a selected narrow bit width and carries high-bit exceptions separately. Frame-of-reference offsets and SIMD-oriented packing reduce the width and support fast bulk decoding.","display_name":"FastPFor","family":"FastPFor","reference_ids":["M08"]},{"description":"Gorilla value coding XORs consecutive floating-point bit patterns, emits a short code for repeated values, and reuses a previous significant-bit window when possible. The measured value kernel stores XOR control metadata and the remaining significant bits; the cited database design also describes timestamp coding.","display_name":"Gorilla","family":"GORILLA","reference_ids":["M03"]},{"description":"Gzip wraps a DEFLATE payload, whose LZ77 matches and literals are Huffman-coded, in a header and a CRC-32/original-size trailer. The wrapper supports concatenated members and stream-level metadata.","display_name":"gzip","family":"GZIP","reference_ids":["M27"]},{"description":"Gipfeli parses the input into LZ77 literal runs and distance/length matches. It separately models and entropy-codes literal data while packing match commands into a compact byte stream.","display_name":"Gipfeli","family":"Gipfeli","reference_ids":["M33"]},{"description":"Huff0 builds a Huffman codebook and writes symbols as variable-length prefix codes. Its implementation divides coding work into multiple streams to expose parallel instruction execution during decoding.","display_name":"Huff0","family":"Huff0","reference_ids":["M49"]},{"description":"Huffman coding assigns shorter prefix codes to more frequent symbols and longer codes to rare symbols. A canonical representation records code lengths so the decoder can reconstruct the codebook compactly.","display_name":"Huffman","family":"Huffman","reference_ids":["M50"]},{"description":"Kanzi is a block-compression framework whose preset selects a transform chain, such as Burrows-Wheeler, text, multimedia, or LZ transforms. The transformed symbols are then passed to a selected entropy coder such as ANS, Huffman, or context modeling.","display_name":"Kanzi","family":"Kanzi","reference_ids":["M61"]},{"description":"LeCo fits a lightweight linear model to each block and stores model parameters plus prediction residuals. The project path ZigZag-maps and bit-packs those residuals, with exceptions carried by its block representation.","display_name":"LECO","family":"LECO","reference_ids":["M16"]},{"description":"LZ4 encodes each block as sequences of literal bytes followed by an offset/length copy from already decoded data. The framed form wraps these LZ4 blocks with stream and block metadata and optional checksums.","display_name":"LZ4","family":"LZ4","reference_ids":["M30"]},{"description":"LZ4 Fast writes the standard LZ4 literal-and-match sequence format. Its acceleration setting makes the greedy hash-table parser skip more candidate positions, changing match selection without changing the decoder format.","display_name":"LZ4 Fast","family":"LZ4 Fast","reference_ids":["M64"]},{"description":"LZ4HC writes the same LZ4 block tokens but spends more search effort on match chains and parsing choices before choosing distance/length copies. Decoding remains the standard LZ4 literal-and-match reconstruction.","display_name":"LZ4HC","family":"LZ4HC","reference_ids":["M64"]},{"description":"LZAV is an in-memory LZ77 design that finds repeated substrings in a dynamic window and represents them with distance/length references interleaved with literal runs. Its stream is token based and does not add a separate entropy-coding stage.","display_name":"LZAV","family":"LZAV","reference_ids":["M65"]},{"description":"libLZF uses a hash-indexed sliding window to find LZ77 matches. The output alternates literal runs with compact length/distance back-references and does not apply a separate entropy coder.","display_name":"LZF","family":"LZF","reference_ids":["M66"]},{"description":"LZFSE parses the input into Lempel-Ziv literals and matches, then entropy-codes the match and literal streams with Finite State Entropy tables. The decoder reconstructs data by decoding those streams and applying the recorded match copies.","display_name":"LZFSE","family":"LZFSE","reference_ids":["M39"]},{"description":"libLZG uses LZ77 string matching and reserves marker bytes for copy commands. Repeated strings become marker, offset, and length forms, while other bytes are emitted as literals with marker escaping when required.","display_name":"LZG","family":"LZG","reference_ids":["M67"]},{"description":"LZHAM uses a large sliding dictionary and LZ parsing to represent repeated strings as length/distance matches. Quasi-adaptive Huffman models encode literals, match classes, lengths, and distances in the bitstream.","display_name":"LZHAM","family":"LZHAM","reference_ids":["M68"]},{"description":"LZJB hashes three-byte prefixes into a 1024-entry match table. A one-byte copy map distinguishes each group of eight literals or matches, and each match uses two bytes to store a 10-bit offset and a 6-bit length code.","display_name":"LZJB","family":"LZJB","reference_ids":["M69"]},{"description":"LZMAT uses a hash-indexed dictionary to locate earlier byte sequences and emits either literals or distance/length copies. Its tag and copy fields are packed at nibble granularity, with variable-width forms for longer lengths and distances.","display_name":"LZMAT","family":"LZMAT","reference_ids":["M35"]},{"description":"The LZO1 variants use LZ77-style dictionary matching and encode the parse as byte-aligned literal runs and distance/length match tokens. They omit a separate entropy-coding pass; the variants mainly differ in match finding and token details.","display_name":"LZO1","family":"LZO1","reference_ids":["M36"]},{"description":"LZO1X uses the LZO byte-oriented LZ77 format, interleaving literal runs with distance/length copy commands. Its presets change compressor search effort and parsing while retaining the LZO1X decoder representation.","display_name":"LZO1X","family":"LZO1X","notes":"This Space family is a measured variant of PPT method LZO1 and reuses M36.","reference_ids":["M36"]},{"description":"LZO2A is an LZO-family LZ77 design that replaces repeated byte sequences with compact match tokens and writes unmatched bytes as literal runs. Its token layout is distinct from the LZO1 variants and does not rely on a separate entropy coder.","display_name":"LZO2A","family":"LZO2A","reference_ids":["M36"]},{"description":"LZRW uses a hash table to find recent LZ77 matches and chooses between literal items and fixed-format copy items. Control words mark which following items are literals and which carry a match offset and length.","display_name":"LZRW","family":"LZRW","reference_ids":["M37"]},{"description":"LZSSE packs literal-run and match-length decisions into four-bit control values and stores match offsets beside them; offsets and literals are XOR-coded against recent state. The LZSSE2, LZSSE4, and LZSSE8 formats use different control thresholds and parsing strategies while arranging decode work for SIMD execution.","display_name":"LZSSE","family":"LZSSE","reference_ids":["M38"]},{"description":"LZVN is a compact Lempel-Ziv format that represents literal runs and distance/length matches with variable-size opcodes. The decoder maintains recent match state and applies copy commands directly to the output buffer.","display_name":"LZVN","family":"LZVN","reference_ids":["M39"]},{"description":"libdeflate implements DEFLATE: LZ77 match finding converts repeated strings into length/distance pairs, and canonical Huffman codes encode those pairs together with literals. The same compressed payload can be emitted raw or inside zlib and gzip wrappers.","display_name":"Libdeflate","family":"Libdeflate","reference_ids":["M62"]},{"description":"Lizard is derived from LZ4 and parses input into literals and LZ-style match commands using either fastLZ4 or the LIZv1 format. Higher method families add Huffman coding to the literal stream, so the measured range includes both token-only and entropy-coded modes.","display_name":"Lizard","family":"Lizard","reference_ids":["M63"]},{"description":"Lzlib writes lzip members containing a simplified LZMA stream. A sliding dictionary produces literals, new matches, and repeated-distance matches; context-dependent probability models and a range encoder code those decisions, followed by CRC and size fields for the member.","display_name":"Lzlib","family":"Lzlib","reference_ids":["M70"]},{"description":"pFPC partitions floating-point data for parallel processing and predicts bit patterns with FCM and differential FCM tables. It XORs each prediction with the actual value and stores the residual using a compact leading-zero-byte representation.","display_name":"PFPC","family":"PFPC","reference_ids":["M21"]},{"description":"PPMd8 predicts each symbol from variable-order contexts, escapes to shorter contexts when needed, and sends the modeled probabilities through a range coder. Its memory setting controls how much context state and history the model can retain.","display_name":"PPMD8","family":"PPMD8","reference_ids":["M71"]},{"description":"Pithy hashes recent byte sequences to find LZ-style repetitions, then represents the stream as literal runs and compact offset-length match tokens. Its search effort controls how many candidates are considered.","display_name":"Pithy","family":"Pithy","reference_ids":["M40"]},{"description":"Pongo applies shortest-decimal-aware floating-point recoding, then XORs adjacent representations and encodes the significant region and zero-boundary metadata. The reconstruction metadata carried by the project path preserves the original represented values.","display_name":"Pongo","family":"Pongo","reference_ids":["M02"]},{"description":"QuickLZ uses hash tables to find recent byte matches and emits literals or compact offset-length commands. Its levels vary the number of stored candidates and the match search performed before choosing a token.","display_name":"QuickLZ","family":"QuickLZ","reference_ids":["M72"]},{"description":"REGER reorders values to strengthen local prediction and packs exact residuals. In the latest current-135 profile, balanced_initial_candidates_no_local_move_v1 dispatches directly to the whole-column backend and omits the four local-move re-encoding rounds; historical REGER profiles remain preserved.","display_name":"REGER","family":"REGER","reference_ids":["M06"]},{"description":"Run-Length and Bit-Level Encoding combines compact bit fields with run coding for repeated patterns in a numeric stream. The measured project row follows the smart-meter encoding reference and retains the metadata needed to reconstruct those fields and runs.","display_name":"RLBE","family":"RLBE","reference_ids":["M11"]},{"description":"Run-length encoding replaces consecutive equal values with a value and repeat count. The measured TsFile-style path combines repeated runs with bit-packed groups for values that do not form long runs.","display_name":"RLE","family":"RLE","reference_ids":["M10"]},{"description":"SLZ is a low-state DEFLATE encoder that finds LZ77 matches and writes fixed-Huffman blocks. This aggregate family covers the same core payload under raw DEFLATE, gzip, and zlib framing.","display_name":"SLZ","family":"SLZ","reference_ids":["M73"]},{"description":"Snappy uses a hash-based LZ77 match finder and emits literal runs or compact offset-length copy commands without a separate entropy coder. Its framed format divides the stream into chunks and adds checksums and chunk metadata.","display_name":"Snappy","family":"SNAPPY","reference_ids":["M47"]},{"description":"SPDP first forms lag-2 modular differences on floating-point words, transposes and differences their byte planes, and then replaces repeated history-conditioned byte sequences with compact repeat counts. These reversible stages target regularity in both numeric words and their byte layout.","display_name":"SPDP","family":"SPDP","reference_ids":["M22"]},{"description":"Sprintz predicts numeric samples, maps signed residuals to small unsigned integers, and bit-packs small groups at widths selected from their residuals. Its design also compresses runs of zero-residual groups to exploit predictable stretches.","display_name":"Sprintz","family":"SPRINTZ","reference_ids":["M12"]},{"description":"Shrinker is a compact LZ-family codec that replaces repeated byte strings with short match commands and leaves unmatched data as literals. The benchmark measures the codec's native token stream as exposed by the archived implementation.","display_name":"Shrinker","family":"Shrinker","reference_ids":["M41"]},{"description":"Simple8b packs multiple nonnegative integers into one 64-bit word, reserving a selector field to choose the count and bit width of the values. Small integers fit more densely, and special selector modes compact long sequences of small repeated values.","display_name":"Simple8b","family":"Simple8b","reference_ids":["M09"]},{"description":"This SLZ variant writes the library's LZ77 matches and literals as fixed-Huffman DEFLATE blocks without an outer gzip or zlib wrapper.","display_name":"SLZ raw DEFLATE","family":"Slz_deflate","notes":"This Space family splits the raw-DEFLATE wrapper from PPT method SLZ.","reference_ids":["M73"]},{"description":"This SLZ variant wraps its fixed-Huffman DEFLATE payload in a gzip member with the gzip header, CRC-32, and original-size trailer.","display_name":"SLZ gzip","family":"Slz_gzip","notes":"This Space family splits the gzip wrapper from PPT method SLZ; M27 documents the wrapper format.","reference_ids":["M73","M27"]},{"description":"This SLZ variant wraps its fixed-Huffman DEFLATE payload in a zlib stream with the zlib header and Adler-32 integrity value.","display_name":"SLZ zlib","family":"Slz_zlib","notes":"This Space family splits the zlib wrapper from PPT method SLZ; M75 documents the zlib implementation family.","reference_ids":["M73","M75"]},{"description":"Stream VByte encodes small integers in one to four data bytes and keeps two-bit length controls in a separate control stream, enabling SIMD decoding. The measured path first takes deltas so smooth integer columns produce smaller unsigned values.","display_name":"Stream VByte","family":"Stream VByte","reference_ids":["M07"]},{"description":"This project kernel scales floating-point inputs when required, partitions integers into subcolumns, and selects bit-packed, run-length, or dictionary bodies for each block. Its block metadata identifies the partitioning and coding choices for reconstruction.","display_name":"Sub-column","family":"Sub-column","notes":"This Space-only project composition uses component ideas documented by M09, M10, and M13; none of those references specifies the complete implementation.","reference_ids":["M09","M10","M13"]},{"description":"TAMP is a small-memory, DEFLATE-inspired LZ codec that searches a configurable plaintext window and emits literal or match tokens in a compact framed stream. The window setting changes the available match history and token cost.","display_name":"TAMP","family":"TAMP","reference_ids":["M79"]},{"description":"The measured TsFile TS_2DIFF kernel stores first differences in blocks, subtracts the minimum difference in each block, and bit-packs the nonnegative offsets. It retains the first value and block metadata; floating-point inputs use the configured decimal scaling before integer encoding.","display_name":"TS_2DIFF","family":"TS_2DIFF","reference_ids":["M04"]},{"description":"Tornado combines LZ77 match finding with greedy, lazy, or optimal parsing and encodes the selected literals and matches with a byte, bit, Huffman, or arithmetic back end. The measured presets select different search structures, window sizes, parsers, and coders.","display_name":"Tornado","family":"Tornado","reference_ids":["M74"]},{"description":"UCL NRV2D and NRV2E find repeated byte strings and replace them with LZ-style distance-length references. Compact bit-coded control, offset, and length fields describe the matches while unmatched bytes remain literals.","display_name":"UCL NRV","family":"UCL NRV","reference_ids":["M42"]},{"description":"UCL NRV2B is a distinct NRV token format that combines LZ-style distance-length matches with bit-coded match fields and literal bytes. Its levels change compressor search effort while using the same NRV2B decoder.","display_name":"UCL NRV2B","family":"UCL NRV2b","notes":"This Space family is a measured variant of PPT method NRV and reuses M42.","reference_ids":["M42"]},{"description":"wfLZ finds repeated byte sequences in a bounded work-memory dictionary and writes literals and offset-length matches in a compact LZ stream. Its chunked form can split input into independently decodable blocks.","display_name":"WFLZ","family":"WFLZ","reference_ids":["M43"]},{"description":"XPACK parses a byte stream into literals and LZ-style matches, selecting among candidates before writing codec-specific match tokens. The benchmark uses the experimental reference implementation cited in the PPT notes.","display_name":"XPACK","family":"XPACK","reference_ids":["M44"]},{"description":"XZ uses LZMA2 dictionary matching with probability-modeled literal, length, and distance symbols encoded by a range coder. The XZ container adds stream framing, checks, and independently controlled LZMA2 chunks.","display_name":"XZ","family":"XZ","reference_ids":["M31"]},{"description":"YALZ77 is a compact LZ77 implementation that searches a sliding history for repeated strings and emits streamable literal and offset-length records. Its search setting trades additional match exploration for a potentially smaller token stream.","display_name":"YALZ77","family":"YALZ77","reference_ids":["M45"]},{"description":"Yappy hashes four-byte prefixes into a bounded recent-history table, searches a level-dependent number of candidates, and emits one-byte literal-run markers or two-byte match commands. Match lengths and offsets are decoded through fixed lookup tables.","display_name":"Yappy","family":"Yappy","reference_ids":["M46"]},{"description":"Zstandard parses data into literals and LZ77 sequences with repeat-offset support. It Huffman-codes literal bytes, FSE-codes match-related symbol streams, and packages the result in independently described frame blocks.","display_name":"Zstandard","family":"ZSTD","reference_ids":["M29"]},{"description":"ZXC performs block-level LZ77 parsing, represents literals and matches with compact integer/token layouts, and can Huffman-code selected sections. Its implementation selects scalar or vectorized paths without changing the lossless format semantics.","display_name":"ZXC","family":"ZXC","reference_ids":["M78"]},{"description":"Zlib implements DEFLATE by replacing repeated strings with LZ77 length-distance pairs and coding literals, lengths, and distances with fixed or dynamic Huffman codes. A zlib wrapper supplies stream metadata and an Adler-32 check.","display_name":"Zlib","family":"Zlib","reference_ids":["M75"]},{"description":"Zlib-ng produces zlib/DEFLATE-compatible streams using optimized match finding, hashing, Huffman construction, and architecture-specific routines. Its compression mechanism remains LZ77 followed by fixed or dynamic Huffman coding.","display_name":"Zlib-ng","family":"Zlib-ng","reference_ids":["M76"]},{"description":"libzling processes data in blocks with a reduced-offset LZ (ROLZ) match transform, then Huffman-codes literals and match-related symbols. Block boundaries bound the state used by the match and entropy stages.","display_name":"Zling","family":"Zling","reference_ids":["M77"]}],"operator_summary":{"adapter_count":14,"composition_count":14,"configuration_count":7,"entry_count":102,"family_count":30,"mechanism_count":67},"operators":[{"description":"Preprocesses floating-point bit patterns by clearing selected mantissa suffix bits and encoding the information needed to reconstruct them. Whether a concrete pipeline is lossless depends on its erasure rule and stored metadata.","family_number":"01","group":"Bit processing","id":"eraser","implementations":[{"description":"对每个int64 x计算high=x>>4和low4=x&15；逆向按64位位型计算(high<<4)|low4。它固定拆4位，不检查浮点指数、尾数或十进制可逆性。","example":"输入 [33,34,-1]，得到 high=[2,2,-1]、low4=[1,2,15]。例如 (2<<4)|1=33；-1 按64位原位恢复。","id":"A001","inputs":"1流values:int64[N]","layer":"configuration","name":"Fixed Low-4-bit Split Proxy","outputs":"2流high:int64[N]、low4:uint8语义[N]；low_bits固定4，可参数化为通用低位拆分。","status":"当前项目参考/代理或封装步骤；按下面实际计算解释，不等同名称暗示的上游算法。"},{"description":"对有限非零double求最短十进制小数，提取前导零数与后续十进制整数；若其位数能放入当前指数留下的尾数字段，则反转低位写入该字段并逐位验证恢复，否则保留原字。","example":"输入原字序列 [+0.0,+0.0]：首字为0，后续零值不做十进制改写，target仍为0，擦除标记表示原字保留。此时后接相邻XOR得到0残差。非零值仅在逆十进制重编码逐位相等时采用改写。","id":"A002","inputs":"1流64位原字bytes:uint8[8N]，逻辑可为int64或float64","layer":"mechanism","name":"Pongo Shortest-decimal Mantissa Recoding","outputs":"3流first:uint64[0/1]、eraser_flags:uint8[N-1]、targets:uint64[N-1]；无外部参数。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"从浮点十进制有效位推导alpha与beta，据指数计算可擦除的低尾数位；仅当掩码后的原字配合beta能逐位恢复且收益达到阈值时输出擦除字，否则发literal，并按块重置beta状态。","example":"输入 [-0.0,+0.0] 走 literal 分支，输出对应 tag/beta 与 erased_bits=[0x8000000000000000,0x0000000000000000]（64位profile）。解码恢复两种符号零；beta没有被用来近似替代原位。","id":"A003","inputs":"float64[N]原位；P003要求可精确窄化为float32，P005处理binary64。","layer":"mechanism","name":"Elf Decimal Significant-digit Mantissa Erasure","outputs":"3流tag:uint8[N]、beta:uint8[N]、erased_bits:uint32或uint64[N]；word_bits、block_size、threshold为profile参数。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"每1000项首值直接保存；其余有限值截取int64整数部分，检测十进制位数并最多保留3位小数及符号，按Java式round重建并逐位核验；失败则保存literal原字。","example":"一块中先存首值100.0；下一值100.25可表示为 integer=100、decimal_count=2、decimal_value=25及正号。解码以100+25/100恢复，并逐位检查；无法精确恢复的记录改用literal。","id":"A004","inputs":"1流float64[N]或可精确转float64的int64[N]","layer":"mechanism","name":"Camel Integer-plus-short-decimal Decomposition","outputs":"6流tag:uint8[N]、raw_bits:uint64[R]、integer:int64[D]、signal:uint8[D]、decimal_count:uint8[D]、decimal_value:uint32[D]；chunk=1000、max_decimal_digits=3。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"Eraser / 擦除","reference_ids":["O01"]},{"description":"Maps decimal-valued floating-point data to integers by multiplying by a power of ten and applying a rounding rule. Exact restoration requires the scale and a conversion rule that preserves every source value used by the pipeline.","family_number":"02","group":"Numeric representation","id":"scaling","implementations":[{"description":"选择阶段从max_point向下取s，使用double乘积与llround生成候选整数，直到相邻差满足int64范围；停止后再逐位检查逆缩放。当前convert在乘积越界时另有退回llround(x)/原位解释的兼容行为，不能笼统说仅“越界拒绝”。","example":"max_point=2，输入 [1.25,1.50,1.75]：候选code=[125,150,175]，相邻差25可装入int64，逆缩放精确，因此输出scale=2与该整数流。","id":"A006","inputs":"1流values:float64[N]及max_point:int32[0…18]","layer":"mechanism","name":"Decimal Scaling with Delta-chain Constraint","outputs":"2流scale:uint32[1]、values:int64[N]；selector=delta_chain_fit、rounding=llround。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"s=max_point直接作为输入；用long double乘积计算llround(x×10^s)，再用double(code)/10^s逐位验证；非有限、越界或不精确拒绝。TS_2DIFF与Simple8b的执行一致。","example":"给定s=2，输入 [1.25,-2.50]，输出 [125,-250]；解码除以100得到原值。s必须作为恢复参数保存。","id":"A007","inputs":"1流values:float64[N]及s=max_point:int32[0…18]","layer":"mechanism","name":"Fixed Exact Decimal Scale","outputs":"1流values:int64[N]，s作为输入参数/packet恢复字段而非新增活动流；rounding=llround。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"从max_point向0搜索首个使所有x×10^s落入int64的s，按llround生成整数；找到首个fit尺度后才做逐位逆检查，若不精确则立即拒绝，不继续尝试更小s。","example":"max_point=2，输入 [1.25,2.50]，最高候选不越界且逆变换精确，输出scale=2、codes=[125,250]；若该候选逆变换不精确，当前profile拒绝输入。","id":"A008","inputs":"1流values:float64[N]及max_point:int32[0…18]","layer":"mechanism","name":"Largest-fitting Decimal Scale Selector","outputs":"2流effective_scale:int32[1]、values:int64[N]；selector=largest_fitting_then_exact_check。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"逐值从0至15找最少可逆小数位，取整组最大值s，再按共同s重新量化并逐位复核。选择阶段乘积为double，量化阶段为long double；P010分1024项，P011整列。失败整块/整列保存raw，含负零也可触发raw。","example":"输入 [1.2,1.23,1.234]，最少所需小数位为[1,2,3]，整组取s=3，得到[1200,1230,1234]。块mode区分scaled与整块raw。","id":"A009","inputs":"values:float64[N]；block_size=1024 或 whole。","layer":"mechanism","name":"Recipe: Required Precision → Decimal Scale","outputs":"scaled 路径输出 precision:uint8 与 scaled_int:int64；raw 路径输出 raw_bits:uint64。P010 可混合 3 流，P011 整列为 scaled 2 流或 raw 1 流。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"float64先从max_point下降选择一组能装入int64的codes，再从max(18,max_point)下降，找可产生同一codes且按wire尺度逐位恢复原输入的s。C++显式实现Python ties-to-even语义；int64直接透传。","example":"整数profile输入 [10,11,12]，仅输出同一code流；浮点profile输入[1.25,2.5]，候选s=2可得到[125,250]，另保存实际恢复scale，解码逐位校验。","id":"A010","inputs":"1流int64[N]或float64[N]原字及kind:string；float需max_point","layer":"mechanism","name":"Code-equivalent Decimal Scale Search","outputs":"effective_scale:int32[1]+codes:int64[N]共2流，int64仅codes 1流；rounding=ties_to_even。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"每1024项枚举exp=10..0、fac=min(exp,9)..0，用估计位宽与异常成本选参数。数值执行x×10^exp→×10^-fac→magic舍入；再按逐位恢复是否相同分离异常，正常code保留，异常另存位置及原位，无可用参数则整块raw。","example":"给定候选exp=2、fac=0，输入[1.25,2.50]，有序乘法及magic舍入得到[125,250]，两个位置都可恢复，无异常。若另有不能恢复的原字，该位置及原字进入异常表。参数选择仍比较全部候选成本。","id":"A011","inputs":"float64 values；selector 产生的 fac/exp:int32[2B]；当前入口仍整体执行，未暴露独立活动端口。","layer":"mechanism","name":"Recipe: ALP Selector → Ordered Scale → Exception Split","outputs":"Scaling 生成 codes:int64 及可逆性结果；异常分离生成带占位正常 codes、exception_index:uint32、exception_bits:uint64。scaled 块最终 4 个语义角色（含 fac/exp），raw 块 1 个角色；mode/count 为恢复信息。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"先用有限值估计mean/sigma，再输出z=(x-mean)/sigma；异常z置0。由z逆预测原字，生成repair=original_bits XOR predicted_bits。解码依靠z、mean/sigma与repair才能逐位恢复。","example":"输入[1.0,2.0,3.0]，均值为2，标准差按实现规则计算为sigma；输出z=[-1/sigma,0,1/sigma]与mean/sigma。每项还输出original_bits XOR inverse_prediction_bits；解码必须应用这条repair流。","id":"A012","inputs":"原始int64/float64位流。参数估计","layer":"mechanism","name":"Recipe: Mean/Sigma Estimate → Normalize → XOR Repair","outputs":"mean_sigma:float64[2]；Normalize输出z:float64[N]；Repair读取原始bits:uint64[N]与逆预测bits:uint64[N]，输出repair:uint64[N]。完整配方共3语义流mean_sigma、z、repair；原类型及N保留用于逆预测。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"Scaling / 数值缩放","reference_ids":["O02"]},{"description":"Stores a block reference value, commonly the minimum or another base, and represents each value as a smaller offset from that reference.","family_number":"03","group":"Numeric representation","id":"frame_of_reference","implementations":[{"description":"求minimum=min(values)，逐项以同宽无符号算术计算offset=unsigned(value)-unsigned(minimum)。逆向将minimum加回并恢复原位。空输入以minimum=0配合N=0表示。","example":"输入int64 [100,103,101]，输出minimum=100、offsets:uint64=[0,3,1]。逆向100+[0,3,1]恢复原序列。","id":"A013","inputs":"values:T[N]，T为int32/uint32/int64/uint64。","layer":"mechanism","name":"Minimum Frame of Reference","outputs":"minimum:T[1]和offsets:unsigned(T)[N]。参数：字宽、块边界。首值、scale、排列、异常字段为旁路恢复元数据；本步不做Bit Packing。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"Frame of Reference (FOR) / 基准帧","reference_ids":["O03"]},{"description":"Reorders values, or timestamp-value pairs, within a block to expose stronger local correlation and reduce later residual widths. A reversible pipeline must preserve enough ordering information to restore the original sequence.","family_number":"04","group":"Reordering and symbol recoding","id":"value_reorder","implementations":[{"description":"建立order=[0…N-1]，按values[order[j]]的signed int64值stable_sort；依order输出sorted_values并保存原位置，逆向把第j个排序值写回original_index[j]。","example":"输入[30,10,30,20]，稳定排序后values=[10,20,30,30]，original_index=[1,3,0,2]。按索引散写恢复[30,10,30,20]。","id":"A014","inputs":"1流values:int64[N]","layer":"mechanism","name":"Stable Sort with Inverse Permutation","outputs":"2流sorted_values:int64[N]、original_index:int64[N]或uint32[N]；参数block_size=whole或64、index_dtype/index_scope为I/O配置。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"Value Reorder / 数值重排","reference_ids":["O04"]},{"description":"Predicts values or timestamp-value pairs from a fitted regression relation, then passes prediction residuals to later integer or bit-level coding stages.","family_number":"05","group":"Residual formation","id":"regression_predictor","implementations":[{"description":"以x_i=value[i]、y_i=value[i+1]对整列做OLS，求theta0、theta1；从首值开始用round(theta0+theta1×previous_value)预测下一值，保存模2^64有符号残差。","example":"输入[10,20,30,40]，相邻样本满足next=10+previous，OLS得到theta0=10、theta1=1；首值10单存，后三个残差全为0。","id":"A015","inputs":"1流values:int64[N]","layer":"mechanism","name":"First-order Autoregression by OLS","outputs":"3流first:int64[0/1]、coefficients:float64[2]、residuals:int64[N-1]；无外部参数，rounding=nearbyint。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"每个512值块以块内位置i为自变量拟合intercept+slope×i；预测经llround并限制在int64域，逐值输出有符号残差。输出有符号残差，ZigZag由A046执行。","example":"一块[10,12,14,16]以位置[0,1,2,3]拟合得到intercept=10、slope=2，signed残差[0,0,0,0]；系数和块长随包保存，残差可再接A046。","id":"A016","inputs":"1流values:int64[N]，参数block_size（当前LECO profile为512）。","layer":"mechanism","name":"Blockwise Linear Regression against Position","outputs":"2流coefficients:float64[blocks,2]、residuals:int64[N]；块长度和原值个数是恢复元数据。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"Regression Predictor / 回归预测","reference_ids":["O04"]},{"description":"Splits exceptional values from a regular block so the common values can use a narrow bit width, while positions and payloads for outliers are encoded separately.","family_number":"06","group":"Numeric representation","id":"outlier_separation","implementations":[{"description":"计算Q1、Q3，令IQR=max(Q3-Q1,1)，边界为floor(Q1-1.5IQR)与ceil(Q3+1.5IQR)。越界值在主体中替换为边界，同时保存原位置与原值，解码再覆盖恢复。","example":"在Q1=10、Q3=12的状态下，IQR=2，裁剪区间[7,15]；输入末项100会在主体变为15，并保存其位置与100。解码覆盖该位置恢复100。","id":"A018","inputs":"values:int64[N]","layer":"mechanism","name":"IQR Clipping（四分位距裁剪）","outputs":"3流clipped_values:int64[N]、outlier_index:int64[E]���outlier_original:int64[E]。倍率1.5和IQR下限1当前固定；无公开参数。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"对上游FOR码排序，枚举宽度b和跨度不超过2^b-1的连续值域，比较K×b+96×E位的代价，选常规区间；按原序标注异常并保留原数值，本步尚不减区间基准。","example":"上游码为[0,1,2,1000]。常规域[0,3]取b=2，成本3×2+96=102位；全域b=10成本40位，因此该小例选择全域、不拆异常。大规模常规值才可能摊薄96位异常开销。","id":"A019","inputs":"BF11 packet:uint8[L]，核心codes:uint64[N-1]","layer":"mechanism","name":"BOS Cost-based Range Split（按成本选区间）","outputs":"7流normal_values:uint64、outlier_values:uint64、tags:uint8、range_base:uint64、bit_width:uint8及透传first:uint64、source_base:int64。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"调用方给定b，对每个uint32值提取low=x & mask(b)，仅当high=x>>b非零时记录pos和high。b=32时low保留全值且异常为空。逆向严格验证位置、流长和有效位范围，按low[pos] OR (high<>2)×delta右移16位得到的校正；以预测误差符号更新accumulator，再更新delta。每9值组重置。","example":"组起点previous=100、delta=0、accumulator=0，下一值103的预测为100、误差为3；观察误差符号更新accumulator，再令delta=3。后续预测使用更新后的状态。","id":"A029","inputs":"int64[N]或float64原位字","layer":"mechanism","name":"Sprintz FIRE Predictor（在线FIRE预测）","outputs":"3流anchors:int64原字[B]、residuals:int64模残差[8B]、raw_tail:int64原字[T]，B=floor(N/9)，T=N%9。当前移位/学习规则固定。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"把double原位映射为可逆的单调uint64码；按固定FPZIP一维浮点表达式从previous产生预测，再做mapped_current-mapped_prediction。初始预测为+0，解码加回并逆映射。","example":"若当前位置的原位与预测原位相同，则两者经同一单调映射得到同一码，mapped residual=0。解码把0加回预测码并逆映射，恢复相同float64位型。","id":"A030","inputs":"float64[N]的8N字节原位，N≤UINT32_MAX","layer":"mechanism","name":"FPZIP Mapped Float Residual（浮点映射残差）","outputs":"1流mapped_residuals:uint64[N]。无公开参数；须保留运算顺序与初始+0。源码未拒绝NaN/Inf，但因预测执行double运算，特殊值位级可逆性需单独round-trip验证。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"Delta / 差分","reference_ids":["O07"]},{"description":"XORs adjacent fixed-width values or floating-point bit patterns so shared bits become zeros and only changing bit regions need further coding.","family_number":"09","group":"Numeric representation","id":"xor","implementations":[{"description":"对每个32/64位原字计算residual=current XOR previous；首值作为anchor保留。逆操作同样是residual XOR previous。状态由调用方按原始记录顺序和chunk边界更新，不做浮点数值转换。","example":"原字[0x3ff0000000000000,0x3ff0000000000001,0x3ff0000000000001]，首字单存，残差[1,0]。逐项与前一恢复字异或，得到完全相同的原位。","id":"A031","inputs":"current:uint32/uint64[N]，previous来自lag=1历史或显式前字。","layer":"mechanism","name":"Adjacent Previous-word XOR","outputs":"anchor与同宽residual流；标签、beta及块计数旁路。Camel的literal路径与decimal路径共用原记录前值状态，不能先分类型过滤再计算相邻关系。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"用当前字低14位查最近同key的位置；若该候选仍在128项历史内且候选XOR尾零数>13，使用该历史槽，否则用前一字。输出所选引用及XOR，再更新环形历史和索引。","example":"历史槽7存A，当前值再次为A且低14位key命中槽7，XOR=0、尾零数64，满足引用条件。输出reference=7和residual=0；解码用槽7原字恢复A，再更新环形历史。","id":"A032","inputs":"int64/float64原字；纯参考","layer":"mechanism","name":"Chimp128 History-reference XOR（历史引用异或）","outputs":"first、references:int64、xor_residuals:uint64共3流；实际Chimp128包还存trailing_zeros:uint8共4流。128槽、14位key、阈值13当前固定。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"FCM按值历史hash预测原字，DFCM按差分历史hash预测stride，再加last得到第二候选；比较两种XOR残差的无符号大小，记录较小者和selector，随后同时更新两个表。","example":"假定本步FCM预测0x1000、DFCM预测0x1010，真实原字0x1011；候选残差为0x11与0x01，选择DFCM并记录selector和0x01。解码先预测，再用selector选预测并XOR残差。","id":"A033","inputs":"非空8字节对齐原字bytes:uint8[8N]","layer":"mechanism","name":"FCM/DFCM Predictor Selection（双预测器选择）","outputs":"2流predictor_selector:uint8[N]、xor_residuals:uint64[N]。公开lg2固定16；pFPC每2^20值重置，底层其他lg2不是已开放图参数。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"XOR / 异或残差","reference_ids":["O08"]},{"description":"Represents an XOR result using its significant bits plus leading/trailing-zero metadata, allowing zero suffixes and other unchanged regions to be omitted from the payload.","family_number":"10","group":"Bit processing","id":"trailing_zero_pack","implementations":[{"description":"消费已有残差。零残差写01；若分桶leading与已保存leading相等且trailing≥已保存trailing，则写00与已保存窗口宽度的载荷；否则写10/11短长窗口头，包含3位leading桶和有效位数，载荷省去已知为1的最低有效位。逆向重建窗口及残差，不隐含XOR。","example":"64位profile，残差1的leading桶=24、trailing=0，significant=40：写长头11、leading码111、长度101000，以及省略最低1后的39位0。随后残差3可复用窗口，写00及40位载荷；零残差写01。","id":"A035","inputs":"residual:uint32/uint64[N]及窗口状态。","layer":"mechanism","name":"Elf+ Residual Window Packing","outputs":"控制/窗口头与有效位组成的bitstream，恢复N/profile/有效长度。32位桶[0,6,10,12,14,16,18,20]，短长长度字段3/5位；64位桶[0,8,12,16,18,20,22,24]，短长字段4/6位。方法适配器单独处理首字、EOS、eraser元数据和A031。","status":"共享C++内核与独立residual_window_pack_cli可用；原方法适配器复用该内核。独立图窗口动作尚未注册。"},{"description":"对所有非零int64取最小尾零数，再与16取小得到shift；每项算术右移shift。空列或全零列shift=0，逆向利用同一个shift恢复各原字。","example":"输入[16,48,80]，共同尾零最小值4，输出shift=4、values=[1,3,5]。逆向同宽左移4位恢复。","id":"A037","inputs":"values:int64[N]","layer":"mechanism","name":"Common Trailing-zero Shift（共同尾零移位）","outputs":"2流values:int64[N]与shift:int64[1]。上限16当前固定，可作参数；并非float数值缩放。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"读取已生成XOR：零残差只写zero标记；非零的leading和trailing都不少于已保存窗口时复用已保存窗口，否则写新leading、有效位宽与有效位，并更新窗口。","example":"已有窗口leading=30、trailing=30，残差0xc0000000的真实leading=32、trailing=30，满足复用条件；只需复用标记及保存窗口宽4位中的0011。残差0只写zero标记。","id":"A038","inputs":"GXI1/GXF1或PXR1 packet；Gorilla","layer":"mechanism","name":"Zero/Reuse/New Window Packing（零值/复用/新窗口）","outputs":"4流first:uint64、window_tags:uint8、window_parameters:uint8、significant_payload:bytes；Pongo多eraser_flags:uint8共5流。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"消费已有XOR与可选reference。四分支：00零残差��01省略尾零并写leading/有效位数，10复用leading，11写新leading。00/01按profile保存引用槽。零值或省尾零分支使leading复用状态失效；逆向读控制字段恢复原残差。","example":"64位、history_index_bits=0、threshold=6，残差1：leading桶24，新头11|111后写40位载荷1；下一残差3复用leading，写10与40位载荷3。零残差写00。索引位数7的profile在需要时另写7位历史槽。","id":"A039","inputs":"residual:uint32/uint64[N]，可选reference:uint8[N]，trailing及leading桶信息。","layer":"mechanism","name":"Parameterized Chimp Residual Window Packing","outputs":"控制位、可选历史索引、窗口字段与有效位流。word_bits=32/64，history_index_bits=0/7，长度字段5/6位，leading桶和尾零阈值固定在profile中。独立CLI用N终止；方法适配器负责首字及EOS。","status":"共享C++内核与独立residual_window_pack_cli可用；原方法适配器复用该内核。独立图窗口动作尚未注册。"},{"description":"按XOR残差落入的字节范围编码为0…7，实际字节数为[0,1,2,3,5,6,7,8]；把此3-bit码与预测器selector合为nibble，每两项合一控制字节，再写低位残差字节。","example":"残差0用byte-count码0；残差0x1234需2字节，用码2。若selector分别为0和1，两个nibble分别为0和10，合成control=0x0a，再写第二项两字节残差（顺序按profile）。","id":"A040","inputs":"FPR1/PFR1预测包，记录(selector:uint8,residual:uint64)","layer":"mechanism","name":"FPC Nibble/Byte Packing（半字控制与整字节残差）","outputs":"2流control_nibbles:bytes、residual_payload:bytes。预测器选择来自上游，本步不重新预测。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"Trailing-zero Packing / 零位与残差打包","reference_ids":["O09"]},{"description":"Builds a set of distinct values or symbols and replaces repeated occurrences with compact dictionary identifiers; the dictionary is retained for decoding.","family_number":"11","group":"Match and repetition","id":"dictionary","implementations":[{"description":"统计int64唯一值，但字典直接保留std::map的数值升序；令dictionary[j]为第j小值，再逐项输出ids[i]=dictionary中values[i]的下标，解码按下标查表。","example":"输入[30,10,30,20]，字典按值升序为[10,20,30]，ID=[2,0,2,1]；查表逐项恢复原值。","id":"A041","inputs":"1流values:int64[N]","layer":"configuration","name":"Value-sorted Integer Dictionary Recoding","outputs":"2流dictionary:int64[D]、ids:int64[N]；无参数。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"把每个8字节原字读成uint64；从左到右遇到新原字即追加到dictionary并记录下标，重复原字复用既有下标，解码用uint64下标恢复原始64位字。","example":"输入原字[A,B,A,C]，首次出现字典[A,B,C]，ID=[0,1,0,2]。即使两个NaN的payload不同，也作为不同原字处理。","id":"A042","inputs":"1流64位原字，可标为int64[N]或float64[N]","layer":"mechanism","name":"First-seen 64-bit-word Dictionary Recoding","outputs":"2流dictionary:同输入原字[D]、ids:uint64[N]；无算法参数。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"先按A042相同的首次出现规则建立float64原位字典和ID；再取b=bit_width(D-1)，把每个ID恰用b位连续写入位流，逆向按相同b读ID并查字典。","example":"输入[A,B,A,C]，字典[A,B,C]、D=3，b=2；ID位串00|01|00|10，连同字典和N保存。解码按2位读ID并查表。","id":"A043","inputs":"1流values:float64[N]原位","layer":"composition","name":"First-seen Dictionary plus Fixed-width ID Packing","outputs":"3流dictionary:float64[D]、id_bit_width:uint8[1]、packed_ids:bytes[ceil(Nb/8)]；b由D生成。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"从左到右扫描bytes，在固定12词表中按表序寻找当前位置可匹配的首个词；命中则写tag=1和词ID并前进词长，否则写tag=0和一个literal字节。","example":"对于不匹配词表的单字节0xff，输出literal tag=0及0xff；若当前位置匹配词表第j项，则输出tag=1和j，解码展开该固定词条。词表及顺序是算法参数的一部分。","id":"A044","inputs":"1流payload:uint8[N]","layer":"mechanism","name":"Fixed 12-word Greedy Substitution","outputs":"3流kind:uint8[T]、literals:uint8[L]、dictionary_ids:uint8[M]；当前词表固定，无需输出。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"将64位字按40-lane块组织，最多取4个40-lane atom作字典；每块对每个atom计算各lane异或残差位宽之和，选代价最小atom，输出selector与逐lane XOR残差。","example":"当前40-lane块与候选atom0���lane完全相同，则40个XOR残差均为0，成本0；选择atom0并保存selector。解码从atom0逐lane异或0恢复该块。","id":"A045","inputs":"CNM2 packet:uint8[L]，含float64归一化字及repair","layer":"mechanism","name":"CORAD Tricklet XOR Atom Selection","outputs":"5流parameters:float64[2]、dictionary:float64[40A]、selectors:uint8[B]、xor_residuals:uint64[N]、repair:uint64[N]；当前width=40、A≤4。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"Dictionary / 字典编码","reference_ids":["O10"]},{"description":"Maps signed integers to unsigned integers by interleaving nonnegative and negative values, keeping small magnitudes small before varint or bit-packed coding.","family_number":"12","group":"Numeric representation","id":"zigzag","implementations":[{"description":"对选中的int64残差逐值计算ZigZag双射，输出uint64码；逆向恢复原signed64位型。reference、Sprintz、DBP、LECO、REGER与BOS64只选择不同位置并旁路各自元数据。","example":"输入signed [-2,-1,0,1,2]，输出unsigned [3,1,0,2,4]；负数映为奇数，非负数映为偶数。INT64_MIN对应UINT64_MAX。","id":"A046","inputs":"输入selected residuals:int64[N]，参数selected positions/I/O profile；A046新生成1流codes:uint64[N]。anchors、coefficients、indices、scale等输入流原样旁路。","layer":"mechanism","name":"Signed64 ZigZag Mapping（有符号到无符号映射）","outputs":"输入selected residuals:int64[N]，参数selected positions/I/O profile；A046新生成1流codes:uint64[N]。anchors、coefficients、indices、scale等输入流原样旁路。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"ZigZag / 之字形映射","reference_ids":["O11"]},{"description":"Applies a reversible block permutation such as the Burrows-Wheeler transform to group similar symbol contexts before recoding and entropy coding.","family_number":"13","group":"Reordering and symbol recoding","id":"block_transform","implementations":[{"description":"建立order=[0…N-1]，按values[order[j]]的signed int64值stable_sort；依order输出sorted_values并保存原位置，逆向把第j个排序值写回original_index[j]。","example":"输入[30,10,30,20]，稳定排序后values=[10,20,30,30]，original_index=[1,3,0,2]。按索引散写恢复[30,10,30,20]。","id":"A014","inputs":"1流values:int64[N]","layer":"mechanism","name":"Stable Sort with Inverse Permutation","outputs":"2流sorted_values:int64[N]、original_index:int64[N]或uint32[N]；参数block_size=whole或64、index_dtype/index_scope为I/O配置。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"把每个64位lag-2残差的第0至7字节依次组成8个平面；按该平面串接顺序维护单个previous状态，输出d=(current-previous) mod256，previous随后更新为current。","example":"两个uint64值[1,2]的小端字节列化后为[1,2,0,0,...]；以previous=0做连续字节Delta，前四字节变为[1,1,254,0]。逆向先前缀加，再逆转置。","id":"A047","inputs":"S2D1 packet:uint8[12+8N]，语义为lag2_deltas:uint64[N]","layer":"composition","name":"SPDP Byte-plane Transpose plus Cross-plane Delta","outputs":"8流plane0_deltas…plane7_deltas:uint8[N]；word_width固定8，previous初值0。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"输入已有uint64值和有效位宽b，令S=ceil(b/8)，将有效位左对齐到S字节，按从高到低字节位置输出S条平面。逆向拼接平面并去除低位padding，仅还原输入值。","example":"输入offsets=[0,3,1]、b=2，S=1，左移6位后输出uint8平面[0x00,0xc0,0x40]。逆向右移6位恢复[0,3,1]；base不参与本步。b=0时没有平面，但N仍保留。","id":"A048","inputs":"values:uint64[N]，b∈[0,64]，每值须可由b位表示。","layer":"mechanism","name":"Byte-plane Split","outputs":"S条uint8[N]平面；N、b以及高低顺序用于恢复，空/零宽情况也保留N。minimum和scale由上游旁路保存，本步骤不计算minimum、不做减法。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"把整个字节串的N个循环旋转按字典序排序，输出每行旋转的末字节last_column，并记录原串所在行primary_index；逆向用字节计数和LF映射恢复原串。","example":"输入ASCII BANANA，排序循环旋转后last_column=NNBAAA、primary_index=3（0起点）。保留两者，LF映射可还原BANANA。","id":"A050","inputs":"1流payload:uint8[N]","layer":"mechanism","name":"Cyclic Burrows-Wheeler Transform","outputs":"2流last_column:uint8[N]、primary_index:uint64[1]；当前整个输入为单块，无算法参数。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"对每个64位code计算low=uint32(code&0xffffffff)、high=uint32(code>>32)，分别按小端写入两条等长limb流；逆向按code=low|(uint64(high)<<32)逐项合成。","example":"输入uint64 0x1122334455667788，low32=0x55667788、high32=0x11223344；两条limb流分别保存，逆向high<<32|low恢复原字。","id":"A051","inputs":"codes:int64[N]或其LNS1块","layer":"configuration","name":"Low/High 32-bit Limb Decomposition","outputs":"2流low32:uint32[N]、high32:uint32[N]；scaled/ALP仅增加scale、factor/exponent、exception等透传流。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"对每个8字节原字交换前4字节与后4字节，半字内部小端顺序不变；该变换自逆，等价于先取得high32、low32，再按high32后low32写回单一路载荷。","example":"输入字节88 77 66 55 44 33 22 11，输出44 33 22 11 88 77 66 55；再执行一次即恢复。","id":"A052","inputs":"1流64位原字bytes:uint8[8N]，逻辑可为int64或float64","layer":"mechanism","name":"Swap 32-bit Halves within Each 64-bit Word","outputs":"1流reordered_words:uint8[8N]；layout=interleaved-high-low，无数值参数。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"Block Transform / 块变换","reference_ids":["O12"]},{"description":"Re-expresses symbols after a transform, for example as move-to-front ranks, so recent or repeated symbols become small values and longer runs.","family_number":"14","group":"Reordering and symbol recoding","id":"symbol_recode","implementations":[{"description":"先统计每个int64符号频数，再按频数降序stable_sort唯一符号；同频保持std::map数值升序，随后令ID为该排名，解码依赖完整dictionary查表。","example":"输入[8,3,8,5,3,8]，频数8:3、3:2、5:1，字典[8,3,5]，ID=[0,1,0,2,1,0]。输出仍是ID流，需要另接熵编码才能得到前缀码。","id":"A053","inputs":"1流symbols:int64[N]","layer":"configuration","name":"Frequency Ordered Dictionary IDs","outputs":"2流dictionary:int64[D]、ids:int64[N]；无熵模型参数，排序规则固定。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"初始化table=[0,1,…,255]；对每个byte输出它在当前table中的rank，再把命中符号移到table[0]并保持其他符号相对次序；解码同步执行同一更新。","example":"初始表[0,1,2,...]，输入[2,2,1]，rank依次为[2,0,2]；第一步表变为[2,0,1,...]，第三步取rank2得到1。","id":"A054","inputs":"1流symbols:uint8[N]","layer":"mechanism","name":"Move-to-Front Recoding","outputs":"1流ranks:uint8[N]；alphabet固定0…255，无输出表和算法参数。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"Symbol Recoding / 符号重编码","reference_ids":["O13"]},{"description":"Replaces consecutive equal values or symbols with a value-and-count representation, reducing long runs while adding overhead to rapidly changing data.","family_number":"15","group":"Match and repetition","id":"run_length","implementations":[{"description":"在raw、constant、全流bitpack与hybrid中按精确编码字节数选最小；hybrid用DP在最长128项literal-bitpack段和任意等值repeat段间切分，同成本优先repeat再优先更早起点。","example":"输入100个7，constant候选只需保存常量及长度；与raw、bitpack、hybrid的完整候选字节数比较，选择最短表示。解码由count=100重复该常量。","id":"A055","inputs":"C++核心输入bytes:uint8[N×symbol_width]，按unsigned固定宽原字比较，symbol_width=1…8；当前公开profile实际用1或8字节。","layer":"mechanism","name":"Size-optimal Hybrid RLE/Bitpack/Raw Selection","outputs":"按模式为1流raw、2流constant、2流bitpack或5流hybrid；scaled传64位raw codes，RLBE先生成位宽序列并旁路delta_bits。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"Run Length / 游程编码","reference_ids":["O14"]},{"description":"Searches previously decoded data for repeated byte strings and emits distance-length copy commands in place of duplicate literals, following the LZ77 family model.","family_number":"16","group":"Match and repetition","id":"lz_match","implementations":[{"description":"从左到右寻找相邻相等int64的最大游程；长度至少4时输出repeat标记、run length和一个值，较短游程输出literal标记、长度和全部原值，不产生向后距离。","example":"输入[5,5,5,5,2]，形成repeat(value=5,length=4)与literal([2],length=1)。解码按长度展开，未出现任何向后距离字段。","id":"A056","inputs":"1流values:int64[N]","layer":"adapter","name":"Threshold-four int64 RLE Compatibility Proxy","outputs":"3流tag:int64[R]、count:int64[R]、values:int64[K]；repeat_threshold固定4，无offset。","status":"当前项目参考/代理或封��步骤；按下面实际计算解释，不等同名称暗示的上游算法。"},{"description":"读取已有NLS1 role=1上下文，验证body仍等于raw_size，仅把role改为2并写入minimum与maximum；载荷逐字节不变，既不建立索引也不查找匹配。","example":"输入NLS1上下文(window=32768,raw_size=8,body=ABCDEFGH)，设min=4、max=130后body仍为ABCDEFGH；输出只增加/更新约束字段，后续A060才查找匹配。","id":"A057","inputs":"NLS1 packet:uint8[25+N]及minimum:uint32、maximum:uint32","layer":"configuration","name":"Attach Native-LZ Minimum/Maximum Configuration","outputs":"仍1流raw_body:uint8[N]并更新头字段，新增算法语义流0。","status":"当前项目参考/代理或封装步骤；按下面实际计算解释，不等同名称暗示的上游算法。"},{"description":"以3字节key维护历史位置链，每处从最近位置反查至多64个候选，选window内最长匹配；若下一位置的最佳长度大于当前长度加1，则放弃当前match并先输出literal。","example":"位置i最佳match长4，而i+1最佳match长7时，7>4+1，lazy规则先发出i处literal；随后再处理更长候选。保存literal/match序列及raw_size以供逆向复制。","id":"A058","inputs":"1流payload:uint8[N]及window/min_match/max_match:uint32","layer":"mechanism","name":"Reference Lazy Hash-chain LZ Search","outputs":"4流tag:uint8[T]、literal:uint8[L]、offset:uint32[M]、length:uint32[M]；search_policy=lazy。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"使用与A058相同的3字节历史链和最多64候选最长匹配，但当前位置一旦找到长度达到min_match的最佳match便立即发出，不计算下一位置候选，也不执行lazy lookahead。","example":"同一候选状态中，位置i已有满足min_match的长4匹配，立即发出(distance,4)，不比较i+1的长7候选；逆向按距离从已恢复前缀复制4字节。","id":"A059","inputs":"1流payload:uint8[N]及window/min/max","layer":"configuration","name":"Reference Greedy Hash-chain LZ Search","outputs":"同A058的4条LZS1语义流；search_policy=greedy。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"以8字节key只保存最近一次位置；match长度先逐字节扩展，再向下取8的倍数，并受有效min/max约束；有match立即贪心发出，否则首位失败时批量发8个literal。","example":"输入ABCDEFGHABCDEFGH（16字节），前8字节作为literal，位置8命中key ABCDEFGH，发出offset=8、length=8。解码复制已恢复的前8字节。","id":"A060","inputs":"P074输入LNS1双肢bytes，核心参数window=32768、min=5、max=175；每肢","layer":"mechanism","name":"Native Aligned-8 Greedy Last-occurrence LZ Search","outputs":"kind:uint8[T]、literal:uint8[L]、offset:uint32[M]、length:uint32[M]，外围scale/exception透传。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"按64KiB块建立乘法hash表，使用skip增量探测寻找4字节候选，命中后逐字节延长；长copy按64或60切片，当前位置附近hash表按Snappy规则更新。","example":"输入中已存在ABCD且当前位置4字节hash命中它，匹配扩展到长度70时，copy按Snappy规则切成多段（每段最多64），再交A068形成tag。此例说明匹配记录，不指定未经运行的完整压缩字节。","id":"A061","inputs":"1流payload:uint8[N]","layer":"mechanism","name":"Snappy snap-1.1.1 Hash-table Match Generation","outputs":"5流tag:uint8[R]、literal_bytes:uint8[L]、literal_length:uint32[J]、copy_offset:uint32[M]、copy_length:uint32[M]；规则固定。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"LZ Match / LZ匹配","reference_ids":["O15"]},{"description":"Chooses among candidate matches and literals using policies such as greedy, lazy, optimal, or repeat-offset parsing, trading search cost against encoded size.","family_number":"17","group":"LZ decisions and format","id":"match_parse_policy","implementations":[{"description":"以8字节key只保存最近一次位置；match长度先逐字节扩展，再向下取8的倍数，并受有效min/max约束；有match立即贪心发出，否则首位失败时批量发8个literal。","example":"输入ABCDEFGHABCDEFGH（16字节），前8字节作为literal，位置8命中key ABCDEFGH，发出offset=8、length=8。解码复制已恢复的前8字节。","id":"A060","inputs":"P074输入LNS1双肢bytes，核心参数window=32768、min=5、max=175；每肢","layer":"mechanism","name":"Native Aligned-8 Greedy Last-occurrence LZ Search","outputs":"kind:uint8[T]、literal:uint8[L]、offset:uint32[M]、length:uint32[M]，外围scale/exception透传。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"对第i个int64原位按模2^64计算y=3×uint64(x)+(i mod251)，再把同一64位型放回int64载体；逆向用3在模2^64下的乘法逆元恢复。","example":"输入[10,10,10]，位置i=0,1,2时输出[30,31,32]；解码减去i mod251，再乘3的模2^64逆元，恢复三个10。","id":"A062","inputs":"1流values:int64[N]","layer":"adapter","name":"Position-dependent Multiply/Add Proxy","outputs":"1流transformed_values:int64[N]；位置i隐式，乘数3和模项251固定。","status":"当前项目参考/代理或封装步骤；按下面实际计算解释，不等同名称暗示的上游算法。"},{"description":"解析并验证输入LZS1的raw_size与literal/match token合法性，然后在外层增加ZPP1 wrapper；逆向去掉wrapper并再次验证，token次序和字段值完全不变。","example":"输入已有LZS1的literal A及match(offset=1,length=3)，验证可恢复AAAA后加ZPP1包头；内部两个token值与顺序保持一致，解码去头并验证。","id":"A063","inputs":"4条已有语义流tag:uint8[T]、literal:uint8[L]、offset:uint32[M]、length:uint32[M]的LZS1 packet","layer":"adapter","name":"Validated Identity Wrapper for Generic LZ Packet","outputs":"同4流透传；无参数。","status":"当前项目参考/代理或封装步骤；按下面实际计算解释，不等同名称暗示的上游算法。"}],"name":"Match Parse Policy / 匹配解析策略","reference_ids":["O16"]},{"description":"Defines how far back matches may refer and how candidates are indexed, bounded, or reset; larger windows can capture distant repetition at additional memory and search cost.","family_number":"18","group":"LZ decisions and format","id":"match_window_model","implementations":[{"description":"对第i个int64的64位原字计算y=uint64(x) XOR (uint64(i)×131)，逆向再次异或同一位置常量；每项独立，不维护任何跨位置历史状态。","example":"输入[10,10,10]，位置常量为[0,131,262]，输出[10,137,268]；再次异或相同常量恢复[10,10,10]。","id":"A064","inputs":"1流values:int64[N]","layer":"adapter","name":"Position-dependent XOR Proxy","outputs":"1流transformed_values:int64[N]；位置i隐式，常数131固定；兼容block_state profile只改变入口名。","status":"当前项目参考/代理或封装步骤；按下面实际计算解释，不等同名称暗示的上游算法。"},{"description":"把原始bytes装入NLS1 role=1 envelope并在头部记录window_bytes与raw_size；body逐字节不变，逆向只核对window及长度后取回body，不建立窗口数据结构。","example":"raw=ABCDEFGH、window_bytes=32768，输出NLS1(role=1,window=32768,raw_size=8,body=ABCDEFGH)。逆向取回8字节body。","id":"A065","inputs":"1流raw_body:uint8[N]及window_bytes:uint32","layer":"configuration","name":"Attach Native-LZ Window Configuration","outputs":"1个NLS1 packet，语义payload仍1流且新增算法流0；min/max可在同一配置接口补齐。","status":"当前项目参考/代理或封装步骤；按下面实际计算解释，不等同名称暗示的上游算法。"}],"name":"Match Window Model / 匹配窗口模型","reference_ids":["O15"]},{"description":"Serializes literal lengths, literal bytes, match offsets, and match lengths into codec-specific commands, such as the nibble-split token and extensions used by an LZ4 block.","family_number":"19","group":"LZ decisions and format","id":"match_token_format","implementations":[{"description":"对每个int64输出high=value>>4；每两个值取低4位，按(low_even<<4)|low_odd组成一个uint8，奇数尾项的低半字节补0；逆向结合high与低4位恢复。","example":"输入[0x21,0x32,0x43]，high=[2,3,4]，低半字节打包为[0x12,0x30]，N=3指明最后0是padding。","id":"A066","inputs":"1流values:int64[N]","layer":"composition","name":"Nibble Split and Pack","outputs":"2流low_pairs:uint8[ceil(N/2)]、highs:int64[N]；low_width固定4，跨类历史入口不增加参数。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"每8个有序token先留一个control byte，第b位为1表示match；literal写1字节，match写offset:uint16和(length-bias):uint8，逆向按control逐token读取并同步展开校验。","example":"token序列为literal A、match(offset=1,length=4)，bias=4：control最低两位为10（二进制），写literal 0x41、offset16=1、length8=0。其余control位对应不存在的token，不增加数据。","id":"A067","inputs":"literal/match token；参数offset_bits=16、length_bits=8、length_bias∈{3,4,5}","layer":"mechanism","name":"Native ctrl8 Literal/Offset16/Length8 Packing","outputs":"4流control:uint8[ceil(T/8)]、literal:uint8[L]、offset:uint16[M]、length_bias:uint8[M]。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"按Snappy语法为literal长度选择1至5字节头，为copy按offset和length选择COPY_1、COPY_2或COPY_4 tag；部分长度与offset位嵌入tag，长copy已由matcher切片。","example":"literal ABC长度3，用tag=(3-1)<<2=0x08，后接41 42 43；COPY_2(offset=3,length=4)用tag=((4-1)<<2)|2=0x0e，后接小端offset 03 00。","id":"A068","inputs":"5条Snappy语义流tag、literal_bytes、literal_length、copy_offset、copy_length","layer":"mechanism","name":"Snappy Literal/COPY Tag Formation","outputs":"物理tag body:bytes，文档仍按5个语义角色展示；规则固定。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"每个sequence以一字节token保存literal_length的高nibble和match_length-4的低nibble；达到15时追加扩展长度，随后写literal bytes、offset:uint16及match扩展。","example":"一个sequence有3字节literal ABC、match长度8和offset=3；token高半字节3、低半字节8-4=4，即0x34，后接ABC及offset小端03 00。","id":"A069","inputs":"LZ4 sequence","layer":"mechanism","name":"LZ4 Token Nibbles and Offset Formation","outputs":"7个语义角色token:uint8[S]、literal_extension:uint32[S]、literal_length:uint32[S]、literal_bytes:uint8[L]、has_match:bool[S]、offset:uint16[M]、match_extension:uint32[M]；match bias=4。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"同样每8个token写一个control byte，第b位区分literal和match；literal写uint8，match依次写未偏置offset:uint32和length:uint32，逆向依control恢复原token序列。","example":"literal A、match(offset=1,length=4)同样用control最低位0、下一位1，但match字段为offset32=1与length32=4，均不使用length-bias。","id":"A070","inputs":"literal/match token；参数offset_bits=32、length_bits=32、length_bias=0","layer":"mechanism","name":"Generic ctrl8 Literal/Offset32/Length32 Packing","outputs":"4流control:uint8[ceil(T/8)]、literal:uint8[L]、offset:uint32[M]、length:uint32[M]。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"Match Token Format / 匹配令牌格式","reference_ids":["O17"]},{"description":"Uses variable-length or byte-oriented codes for nonnegative integers such as residuals, lengths, offsets, and metadata, giving small integers shorter representations.","family_number":"20","group":"Bit processing","id":"integer_code","implementations":[{"description":"对每个uint32/uint64反复取低7位；非末字节置continuation bit=1，末字节置0。核心可编码整列、DBP已ZigZag残差位置或Snappy的单个raw_size。","example":"输入[0,127,128,300]，ULEB128字节依次为00、7f、80 01、ac 02。解码去掉continuation位并按7位组拼接；signed输入先接A046。","id":"A023","inputs":"核心输入values:uint32[N]或uint64[N]，参数word_bits=32/64与I/O profile；A023新生成1流encoded:bytes。DBP adapter完整","layer":"mechanism","name":"Unsigned ULEB128（无符号Base-128变长码）","outputs":"2种语义：first:int64[0/1]与encoded residuals:bytes；scaled profile另带scale:int32旁路恢复字段。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"解析SVD1中的low_deltas与high_deltas两条uint32流；每4值用一个control byte的四个2-bit字段记录1至4字节长度，低位数据紧密写入各自data区。","example":"四个uint32 [1,255,256,65536]需字节数[1,1,2,3]，2-bit长度码[0,0,1,2]合成control=0x90，数据小端为01 ff 00 01 00 00 01。","id":"A072","inputs":"1个SVD1 packet:uint8[12+8N]，字段scale:int32、count:uint32、low_deltas/high_deltas:uint32[N]","layer":"mechanism","name":"Stream VByte Coding of uint32 Limbs","outputs":"4流low/high_control:uint8[ceil(N/4)]与low/high_data:bytes，scaled另透传scale共5流；无新增参数。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"解析PFS1中已经编码好的low/high pfor_words与raw_tail；每肢原样复制pfor_words，只把raw_tail:uint32[T]交给VariableByte，再把两段拼成公开word序列。","example":"已有pfor_words=[W0,W1]与raw_tail=[1,128]：保留W0/W1不变，仅把尾部交VariableByte编码并追加；两肢各自记录长度，逆向分别恢复前缀和两个尾值。","id":"A073","inputs":"1个PFS1 packet，含kind:uint8、scale:int32、count:uint64及每肢pfor_words:uint32[P]、raw_tail:uint32[T]；完整","layer":"mechanism","name":"VariableByte Tail Coding after FastPFor Prefix","outputs":"每肢4条PFor前缀语义流加1条tail_bytes，整数源至多10流，scaled再加scale；无新增参数。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"解析P086产生的LZT1 sequence；literal或match的长度nibble达到15时写对应extension，包括extra=0时必需的零终止字节。extension反复写255并减255，最后写0..254余数；token、literal字节和offset保持原语义。","example":"LZ4扩展长度为510，输出ff ff 00；长度17输出11。255重复字节后必须有一个小于255的终止余数。","id":"A075","inputs":"LZT1 packet，含token、literal_extra/literal_length、literal_bytes、has_match、offset、match_extra；完整","layer":"mechanism","name":"LZ4 Extension-length Base-255 Coding","outputs":"压缩块5流tokens、literal_extension_bytes、literal_bytes、offset、match_extension_bytes，raw块1流raw_bytes，混合共6流；无参数。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"Integer Code / 整数编码","reference_ids":["O18"]},{"description":"Writes flags, codewords, and variable-width fields into a contiguous bitstream so sub-byte fields do not consume whole bytes; DEFLATE is a canonical bit-packed format.","family_number":"21","group":"Bit processing","id":"bitstream_pack","implementations":[{"description":"对每个int64输出high=value>>4；每两个值取低4位，按(low_even<<4)|low_odd组成一个uint8，奇数尾项的低半字节补0；逆向结合high与低4位恢复。","example":"输入[0x21,0x32,0x43]，high=[2,3,4]，低半字节打包为[0x12,0x30]，N=3指明最后0是padding。","id":"A066","inputs":"1流values:int64[N]","layer":"composition","name":"Nibble Split and Pack","outputs":"2流low_pairs:uint8[ceil(N/2)]、highs:int64[N]；low_width固定4，跨类历史入口不增加参数。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"以首个float64原始位型开块；后续记录写literal标记+64位值，或写decimal标记、整数差分分级码、十进制位数和值；每1000条形成带count/body_size的块。","example":"输入first原位100.0，后续decimal记录integer_delta=0、decimal_count=2、decimal_value=25，写记录类型、分级整数码和小数字段。解码得到100.25；literal记录则携带其XOR残差的64位载荷。","id":"A076","inputs":"1路Camel XOR packet:bytes（记录含float64位型/decimal字段）","layer":"mechanism","name":"Camel decimal/XOR record bitstream writer","outputs":"1路Camel public wire:bytes。chunk_size固定1000；整数差分仅支持既定分级范围。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"Bitstream Write / 位流写入","reference_ids":["O19"]},{"description":"Builds a prefix code in which more frequent symbols receive shorter integral-bit codewords and every symbol remains independently decodable from the bitstream.","family_number":"22","group":"Probability coding","id":"entropy_huffman","implementations":[{"description":"先统计每个int64符号频数，再按频数降序stable_sort唯一符号；同频保持std::map数值升序，随后令ID为该排名，解码依赖完整dictionary查表。","example":"输入[8,3,8,5,3,8]，频数8:3、3:2、5:1，字典[8,3,5]，ID=[0,1,0,2,1,0]。输出仍是ID流，需要另接熵编码才能得到前缀码。","id":"A053","inputs":"1流symbols:int64[N]","layer":"configuration","name":"Frequency Ordered Dictionary IDs","outputs":"2流dictionary:int64[D]、ids:int64[N]；无熵模型参数，排序规则固定。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"统计256个byte频率，按项目固定建树规则生成静态Huffman码；包内保存原长、频率/树恢复信息和有效bit数，解码逐bit沿左右子树直到叶节点。","example":"字节AAAA BBB C（频数4、3、1）建立静态Huffman树；高频A获得更短码。若示意树取A=0、B=10、C=11，则ABAC为010011；实际码字由固定建树平局规则决定，模型与有效bit数必须入包。","id":"A077","inputs":"1路raw:uint8[L]","layer":"mechanism","name":"Project Byte Huffman","outputs":"1路Huffman packet:bytes（头部模型/原长/bit数与payload同包）。当前无公开树策略参数。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"Huffman Coding / 霍夫曼编码","reference_ids":["O20"]},{"description":"Encodes a symbol sequence by successively narrowing a numeric interval according to model probabilities, approaching fractional-bit entropy costs through arithmetic state updates.","family_number":"23","group":"Probability coding","id":"entropy_range","implementations":[{"description":"先统计每个int64符号频数，再按频数降序stable_sort唯一符号；同频保持std::map数值升序，随后令ID为该排名，解码依赖完整dictionary查表。","example":"输入[8,3,8,5,3,8]，频数8:3、3:2、5:1，字典[8,3,5]，ID=[0,1,0,2,1,0]。输出仍是ID流，需要另接熵编码才能得到前缀码。","id":"A053","inputs":"1流symbols:int64[N]","layer":"configuration","name":"Frequency Ordered Dictionary IDs","outputs":"2流dictionary:int64[D]、ids:int64[N]；无熵模型参数，排序规则固定。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"先把256符号原始频数归一到总频率2^15并构造cum；编码每���byte时令temp=range/2^15，更新low+=cum[s]*temp、range=freq[s]*temp并归一化输出。","example":"示意归一化模型总频率32768，freq(A)=24576、freq(B)=8192；初始range=2^32时temp=131072。编码B令low增加24576×131072，range变为8192×131072；解码使用相同cum/freq。","id":"A078","inputs":"1路uint8[L]","layer":"mechanism","name":"Static Byte Range Coder","outputs":"1路bytes包，内含origLen:uint32、256×uint32原频数、有效bit数和range payload。总频率2^15固定。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"以上一float64重建值做1D预测；把预测与当前映射整数的差分符号和最高位宽映射到129符号，余下低bits直接入range流；RCqsmodel在每个符号后更新。","example":"连续相同值的mapped residual=0，选择表示零差分的模型符号并更新自适应频数；解码先从range流读该符号，再按同一预测得到原位。实际符号编号由FPZIP映射固定。","id":"A079","inputs":"1路FPZIP residual packet:bytes（count:uint32+uint64残差）","layer":"mechanism","name":"FPZIP Adaptive 129-symbol Range Model","outputs":"1路range wire:bytes。预测器固定lossless 1D double，符号数129。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"Range Coding / 区间编码","reference_ids":["O21"]},{"description":"Uses an Asymmetric Numeral Systems state, often table-normalized as tANS/FSE, to combine near-arithmetic-coding compression with table-driven encode/decode transitions.","family_number":"24","group":"Probability coding","id":"entropy_fse","implementations":[{"description":"先统计每个int64符号频数，再按频数降序stable_sort唯一符号；同频保持std::map数值升序，随后令ID为该排名，解码依赖完整dictionary查表。","example":"输入[8,3,8,5,3,8]，频数8:3、3:2、5:1，字典[8,3,5]，ID=[0,1,0,2,1,0]。输出仍是ID流，需要另接熵编码才能得到前缀码。","id":"A053","inputs":"1流symbols:int64[N]","layer":"configuration","name":"Frequency Ordered Dictionary IDs","outputs":"2流dictionary:int64[D]、ids:int64[N]；无熵模型参数，排序规则固定。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"把256频率归一到2^15，按输入倒序更新state=(state/fs)*M+(state%fs)+cum；state过大时低字节renorm输出，末尾写32位state。解码按state%M查符号。","example":"示意模型M=32768、freq(A)=16384、cum(A)=0，state=100000时编码A的新状态=(100000 div16384)×32768+(100000 mod16384)=198304；归一化字节与末状态也必须保存。","id":"A080","inputs":"1路uint8[L]","layer":"mechanism","name":"Project Static Byte rANS","outputs":"1路rANS packet:bytes，含origLen:uint32、256×uint16频率表、state/body。prob_bits=15与初态固定。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"ANS / FSE / ANS／FSE编码","reference_ids":["O22"]},{"description":"Represents codec-specific entropy or statistical back ends that do not fit the shared Huffman, range, or FSE/ANS categories; behavior is defined by the selected method.","family_number":"25","group":"Probability coding","id":"entropy_special","implementations":[{"description":"先统计每个int64符号频数，再按频数降序stable_sort唯一符号；同频保持std::map数值升序，随后令ID为该排名，解码依赖完整dictionary查表。","example":"输入[8,3,8,5,3,8]，频数8:3、3:2、5:1，字典[8,3,5]，ID=[0,1,0,2,1,0]。输出仍是ID流，需要另接熵编码才能得到前缀码。","id":"A053","inputs":"1流symbols:int64[N]","layer":"configuration","name":"Frequency Ordered Dictionary IDs","outputs":"2流dictionary:int64[D]、ids:int64[N]；无熵模型参数，排序规则固定。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"以6字节历史匹配作为上下文，hash状态索引last位置；若历史相同则输出最多255的复制count，再输出literal，并以(history<<2)^value更新状态。解码同步重放复制。","example":"历史6字节上下文相同且已输出前缀包含ABCDEFG时，可发出复制count再发literal；count受255上限限制。解码使用相同hash/last状态，从已恢复位置复制，随后观察literal并更新历史。","id":"A081","inputs":"1路SPDP block packet:bytes","layer":"mechanism","name":"SPDP History-context Repeat Encoder","outputs":"1路repeat/literal bytes。level:uint32∈[0,9]控制表大小，默认5；逆向还需original_size:uint64。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"Special / Literal Coding / 特殊值／字面量编码","reference_ids":["O23"]},{"description":"Codes literal bytes through a separate probability model or codebook while match lengths, offsets, and other sequence symbols use their own representation.","family_number":"26","group":"LZ decisions and format","id":"literal_entropy","implementations":[{"description":"先统计每个int64符号频数，再按频数降序stable_sort唯一符号；同频保持std::map数值升序，随后令ID为该排名，解码依赖完整dictionary查表。","example":"输入[8,3,8,5,3,8]，频数8:3、3:2、5:1，字典[8,3,5]，ID=[0,1,0,2,1,0]。输出仍是ID流，需要另接熵编码才能得到前缀码。","id":"A053","inputs":"1流symbols:int64[N]","layer":"configuration","name":"Frequency Ordered Dictionary IDs","outputs":"2流dictionary:int64[D]、ids:int64[N]；无熵模型参数，排序规则固定。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"统计256个byte频率，按项目固定建树规则生成静态Huffman码；包内保存原长、频率/树恢复信息和有效bit数，解码逐bit沿左右子树直到叶节点。","example":"字节AAAA BBB C（频数4、3、1）建立静态Huffman树；高频A获得更短码。若示意树取A=0、B=10、C=11，则ABAC为010011；实际码字由固定建树平局规则决定，模型与有效bit数必须入包。","id":"A077","inputs":"1路raw:uint8[L]","layer":"mechanism","name":"Project Byte Huffman","outputs":"1路Huffman packet:bytes（头部模型/原长/bit数与payload同包）。当前无公开树策略参数。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"Literal Entropy / 字面量熵编码","reference_ids":["O24"]},{"description":"Controls block boundaries, state reuse or reset, streaming flushes, and per-block mode choices that determine which history and models remain available to later data.","family_number":"27","group":"Layout and state","id":"block_state_control","implementations":[{"description":"对raw调用当前Snappy RawCompress，再解析其varint原长和literal/copy tag为SNM1记录。逆向按literal或向后复制恢复字节；它实际上保留了完整原生matcher的决策。","example":"输入ABCABCABC，由Snappy库产生实际压缩结果，再解析为literal/copy记录，例如literal ABC与向后复制的语义。SNM1保存raw_size及库选定的实际记录，不假定每次都选择同一手工切分。","id":"A082","inputs":"raw:uint8[L]，L≤UINT32_MAX","layer":"mechanism","name":"Native Snappy Match Records","outputs":"5语义流tag:uint8、literals:uint8、literal_length:uint32、copy_offset:uint32、copy_length:uint32。物理为SNM1，当前无公开参数。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"先按64KiB linked-block、level=0生成LZ4F，再解析literal run、offset及match length为LZM1。原生LZ4查hash候选，命中后可向前回扩到anchor，再延长匹配。","example":"输入重复的64KiB字节块，先由LZ4 linked-block编码，再把生成块解析为literal序列及(offset,length)。linked模式的第二块可引用第一块保留的字典，解码必须维持该跨块状态。","id":"A083","inputs":"raw:uint8[L]","layer":"mechanism","name":"Native LZ4 Linked-block Match Records","outputs":"5语义流literals:uint8、literal_length:uint32、has_match:uint8、offset:uint16、match_length:uint32。物理为LZM1；块长及level当前固定。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"先逆转ALP1前缀中的bitpack、FOR、scaling恢复float64，再调用原生ALP：每1024项选择factor/exponent、标记不能精确恢复的异常，以两条32位limb做FastPFOR并写ALP2。","example":"输入一个表示[1.25,1.50,1.75]的ALP1前缀包；入口先恢复这三个float64，再运行完整ALP生成ALP2。输出参数、两肢压缩块与异常信息由ALP重新选定，不能把ALP1前缀载荷直接当作原生ALP主体。","id":"A084","inputs":"ALP1:bytes。","layer":"composition","name":"ALP Native Remaining Pipeline","outputs":"raw块1流uint64；scaled块5流factor_exponent:int8、low_words:uint32、high_words:uint32、exception_positions:uint16、exception_bits:uint64；混合6角色。物理ALP2封包。","status":"完整或剩余流程入口，包含多个内部机制；不声称已拆成独立原子动作。"},{"description":"逆转RGR1的位打包、FOR、相邻预测与排列后，调用当前REGER。对候选顺序和是否存time比较完整块成本；每序列再选raw、constant、FOR、Delta、TimeLinear或PrevLinear并分段打包。","example":"输入可逆RGR1前缀表示[10,12,14,16]；入口先恢复数值，再由REGER比较常量、FOR、Delta及回归等候选的完整块大小，输出选择标志和对应payload。","id":"A085","inputs":"输入RGR1:bytes。time/value每序列随模式为1/1/3/4/4/5语义流；字段为int64主体/first/minimum、float32[2]模型、uint8位宽、bytes残差，可另有order。物理为完整REGER wire。","layer":"composition","name":"Current REGER Search and Payload","outputs":"输入RGR1:bytes。time/value每序列随模式为1/1/3/4/4/5语义流；字段为int64主体/first/minimum、float32[2]模型、uint8位宽、bytes残差，可另有order。物理为完整REGER wire。","status":"完整或剩余流程入口，包含多个内部机制；不声称已拆成独立原子动作。"},{"description":"先逆项目Huffman、LZ和固定字典前缀恢复raw，再调用Brotli quality=6。内部将literal、command、distance分开建模，literal/distance可按context map选Huffman表，块切换另有控制码。","example":"项目前缀可逆表示ASCII ABCABCABC；入口先还原这9字节，再由Brotli quality=6生成压缩流B。用Brotli解码B回到同一9字节；不把项目前缀中的token当作Brotli内部command。","id":"A087","inputs":"输入Huffman(LZS1(BRD1(raw)))结构化bytes。压缩块4语义角色：entropy_models、literal_codebits、command_codebits、distance_codebits（均bytes），有块切换则加1；raw块1角色。quality/window当前固定。","layer":"composition","name":"Brotli Quality-6 Remaining Pipeline","outputs":"输入Huffman(LZS1(BRD1(raw)))结构化bytes。压缩块4语义角色：entropy_models、literal_codebits、command_codebits、distance_codebits（均bytes），有块切换则加1；raw块1角色。quality/window当前固定。","status":"完整或剩余流程入口，包含多个内部机制；不声称已拆成独立原子动作。"},{"description":"逆项目rANS、Huffman、parse wrapper及LZ前缀得到raw，交给Zstd level=9。原生路径对literal使用Huffman或raw/RLE，对literal length、match length、offset等序列符号使用FSE及附加位。","example":"项目前缀表示ABCABCABC，入口恢复raw后用Zstd level=9生成Zstd frame；其literal和sequence熵模型由Zstd内部选择。解码frame，再逆构造前缀所需表示。","id":"A088","inputs":"输入项目rANS(Huffman(ZPP1(LZS1(raw)))) bytes。压缩schema为L+H+F+S+I(sequence)：L=1/4路literal bytes，H=0/1模型，F=0..3模型，S=0..3个uint32状态及sequence_bits；raw/RLE各1角色。","layer":"composition","name":"Zstd Level-9 Remaining Pipeline","outputs":"输入项目rANS(Huffman(ZPP1(LZS1(raw)))) bytes。压缩schema为L+H+F+S+I(sequence)：L=1/4路literal bytes，H=0/1模型，F=0..3模型，S=0..3个uint32状态及sequence_bits；raw/RLE各1角色。","status":"完整或剩余流程入口，包含多个内部机制；不声称已拆成独立原子动作。"},{"description":"逆项目token、LZ及SHF1 shuffle前缀后，以typesize执行Blosc1 shuffle与BLOSCLZ-5。BLOSCLZ先对末1/4采样，probe使用12-bit hash；根据split标志或预测压缩性选择minlen/ipshift为3或4。","example":"输入4个相同的64位字，typesize=8；恢复原始32字节后，Blosc1按字节shuffle再运行BLOSCLZ-5，输出完整Blosc块。解码和逆shuffle恢复32字节。","id":"A089","inputs":"输入LZT1(LZS1(SHF1)) bytes，typesize∈{1,4,8}。正常chunk按plane编号合并为1至typesize流encoded_plane:bytes；memcpy为1流raw:bytes。level5/shuffle/单线程固定。","layer":"composition","name":"C-Blosc1 BLOSCLZ with Shuffle","outputs":"输入LZT1(LZS1(SHF1)) bytes，typesize∈{1,4,8}。正常chunk按plane编号合并为1至typesize流encoded_plane:bytes；memcpy为1流raw:bytes。level5/shuffle/单线程固定。","status":"完整或剩余流程入口，包含多个内部机制；不声称已拆成独立原子动作。"},{"description":"执行C-Blosc2的BLOSCLZ-5、byte shuffle、单线程和自动块划分。P114先逆项目prefix；P119直接接raw，再把输出chunk转成B2P1。两者来源版本的该活动压缩路径已归为同一组。","example":"raw为4个相同64位字，typesize=8、单线程、BLOSCLZ-5、byte shuffle；C-Blosc2生成chunk，再由对应入口转成B2P1或保留本机块结构。解码恢复32字节。","id":"A090","inputs":"P114输入结构化prefix及typesize=1/4/8；P119输入raw:uint8且typesize固定1。正常按plane合并1至typesize路bytes，memcpy/特殊模式1路body（zero可空）；物理chunk或B2P1。","layer":"composition","name":"C-Blosc2 BLOSCLZ Shared Pipeline","outputs":"P114输入结构化prefix及typesize=1/4/8；P119输入raw:uint8且typesize固定1。正常按plane合并1至typesize路bytes，memcpy/特殊模式1路body（zero可空）；物理chunk或B2P1。","status":"完整或剩余流程入口，包含多个内部机制；不声称已拆成独立原子动作。"},{"description":"逆项目Huffman、URL1、MTF、BWT恢复raw，再调用Bzip2 level4。原生编码包含块排序/BWT、MTF与RUNA/RUNB零游程符号，并按组选择多张Huffman表编码。","example":"项目前缀BWT/MTF/RLE/Huffman表示BANANA，入口先恢复6字节，再调用Bzip2 level=4，生成带块排序和自身熵模型的完整流。最终解压应为BANANA。","id":"A091","inputs":"Huffman(URL1(MTF(BWT(raw)))) bytes，URL1 symbol_width=1。","layer":"composition","name":"Bzip2 Block-sort Remaining Pipeline","outputs":"5语义组used_alphabet:uint8、BWT_primary:uint32、code_lengths:uint8、selectors:uint8、coded_MTF_RLE2:bytes；跨块按同角色合并。","status":"完整或剩余流程入口，包含多个内部机制；不声称已拆成独立原子动作。"},{"description":"gzip/zlib profile先逆项目前缀得到raw；CORAD直接压CTK2。三者共用parameterized_deflate，执行LZ匹配及stored/fixed/dynamic块选择；压缩块使用literal/length与distance两套Huffman码。","example":"raw=ABCABCABC，DEFLATE核心选择stored/fixed/dynamic块并输出压缩core；raw-DEFLATE、zlib、gzip的窗口和封包配置由调用方指定，恢复使用匹配配置。","id":"A092","inputs":"输入由io_variant选择gzip_prefix/zlib_prefix bytes或CTK2；level:int32∈[0,9]。stored/fixed/dynamic语义流分别1/2/3路，混合最多4角色：raw、literal_length_bits、distance_bits、Huffman_model（bytes）。","layer":"composition","name":"Parameterized DEFLATE Core","outputs":"输入由io_variant选择gzip_prefix/zlib_prefix bytes或CTK2；level:int32∈[0,9]。stored/fixed/dynamic语义流分别1/2/3路，混合最多4角色：raw、literal_length_bits、distance_bits、Huffman_model（bytes）。","status":"完整或剩余流程入口，包含多个内部机制；不声称已拆成独立原子动作。"},{"description":"逆项目Range、parse wrapper和LZ前缀恢复raw，再用preset5、CRC64的LZMA2路径。其literal/匹配/重复距离由二元自适应概率模型和range coder写出，并在chunk边界处理模型或字典重置。","example":"项目前缀可逆表示ABCABCABC，入口还原raw后运行preset=5、CRC64的XZ/LZMA2路径；输出XZ流，解压得到同一9字节。","id":"A093","inputs":"Range(ZPP1(LZS1(raw))) bytes；压缩chunk","layer":"composition","name":"XZ / LZMA2 Remaining Pipeline","outputs":"range_bytes:bytes与final_flush:bytes两角色，raw为1角色，混合3角色；物理为XZ核心包，preset5固定。","status":"完整或剩余流程入口，包含多个内部机制；不声称已拆成独立原子动作。"},{"description":"把64位原字按int64处理，在块内比较排列、是否保存time及六种序列模式，选择完整payload较小的候选。拟合用double乘积及Python式fsum，预测用double，再做截断和范围处理。","example":"一个int64块[5,5,5,5]会把constant作为候选，与raw/FOR/Delta等完整候选比较；最终payload同时记录实际选定模式，解码不能自行重新运行成本搜索。","id":"A094","inputs":"words:uint8[8N]，按int64位型解释；配置block_size默认513、segment_size默认16、profile=fast/balanced。","layer":"composition","name":"REGER Candidate Search with Compatibility Arithmetic","outputs":"每time/value序列按模式1/1/3/4/4/5路，可加order；字段类型同A085，物理RGP1。","status":"完整或剩余流程入口，包含多个内部机制；不声称已拆成独立原子动作。"}],"name":"Block State Control / 块状态控制","reference_ids":["O25"]},{"description":"Wraps a compressed payload with format identifiers, lengths, checksums, member boundaries, or other metadata needed to transport, validate, and decode the stream.","family_number":"28","group":"Layout and state","id":"container_frame","implementations":[{"description":"在已有payload前写4字节magic和8字节payload长度，逆向移除并校验对应字段。五个reference容器使用同一wrap/unwrap；CORAD在CZL2外增加CRD2外壳。","example":"magic=TEST、payload=01 02 03，输出TEST、长度3的小端8字节字段、01 02 03。解码核对长度后返回3字节。","id":"A086","inputs":"payload:uint8[L]；CORAD输入CZL2:bytes。","layer":"adapter","name":"Magic and Length Envelope","outputs":"1条完整container:bytes，主体不变；magic/profile及payload合法性属于格式配置。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"在int64值序列前插入[31,139,N低16位]，尾部插入所有值之和的低16位；逆向直接取去掉前三项和最后一项后的原序列。","example":"输入[5,7]，输出int64序列[31,139,2,5,7,12]。当前逆函数去除前3项及末项得到[5,7]；末字段是槽位，不构成经过验证的校验和保障。","id":"A095","inputs":"values:int64[N]","layer":"adapter","name":"Reference Integer Header and Checksum Slots","outputs":"1流container:int64[N+4]。无参数，新增4个int64槽位；与byte帧的区别留在I/O和schema中。","status":"当前项目参考/代理或封装步骤；按下面实际计算解释，不等同名称暗示的上游算法。"},{"description":"由LZI1已编码块生成LZ4F：写magic、descriptor及其XXH32派生校验字节，再写块长度与raw标志、块body，最后写零长度终止块。逆向解析回块记录。","example":"已有一个3字节raw块ABC，LZ4F块长字段为0x80000003，后接ABC，再写0终止块；frame descriptor及校验字节另按配置生成。","id":"A096","inputs":"LZI1:bytes（raw_size:uint64及每块标志、长度、body）","layer":"adapter","name":"LZ4 Frame Assembly","outputs":"1流LZ4F:bytes。不接受裸值或原始payload，无公开算法参数。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"解析已保存的header、core、trailer长度，将三段原样按顺序拼接。gzip/zlib共用assemble；XZ固定两端各12字节，直接移除XZC1的描述头得到原frame。","example":"输入header=01 02、core=aa bb cc、trailer=ff及三段长度，输出01 02 aa bb cc ff。逆向用固定或已记录边界恢复三段。","id":"A097","inputs":"DFC1(G/Z)或XZC1:bytes，内含3段uint8[]","layer":"adapter","name":"Stored Header / Core / Trailer Assembly","outputs":"1流完整gzip/zlib/XZ:bytes。profile固定选择格式，长度必须与packet一致。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"从LNS1的两条已编码limb取出low/high字节；写low长度后依次放两条body。整数前置I标签，浮点前置F标签及scale；逆向依标签与长度还原字段。","example":"整数profile，两肢body分别为aa bb和cc，输出I类型标记、low长度2、low body与high body；浮点profile还写F和scale。长度字段确定两肢边界。","id":"A098","inputs":"LNS1 JSON bytes，stage=wires、format=scaled、单块、wires为2个hex串；float另含scale:int32。","layer":"adapter","name":"Legacy Two-limb Numeric Frame","outputs":"1流legacy numeric wire:bytes。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"把上游已选择的raw/scaled块写成ALP2；raw块写原字，scaled块写fac、exp、两肢body及异常位置/原字；总长、块长与模式按固定字段宽度序列化。","example":"ALP2一块mode=raw、count=2、原字对应1.25与2.5，封包写块模式和两原字；scaled块则保存fac/exp、两肢body及异常表，模式决定解码分支。","id":"A099","inputs":"LNS1 JSON bytes，stage=wires、format=alp；含count、raw或fac/exp、wires及exceptions。","layer":"adapter","name":"Legacy ALP2 Block Frame","outputs":"1流ALP2:bytes；块长≤1024，无数值搜索参数。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"解析RGP1，写REGER3 magic、总数、固定列数1、块长和segment长；逐块写count、flags、body长度和既有body。逆向重新构造RGP1。","example":"单列共3项、一块，RGP1内已有flags及body；REGER3 frame写总数3、列数1、块长、segment长，再写该块count/flags/body_size和body。","id":"A100","inputs":"RGP1:bytes，count:uint64、block/segment_size:uint32、块记录","layer":"adapter","name":"REGER3 Single-column Frame","outputs":"1流REGER3:bytes。输出总数限uint32、块及segment尺寸限uint16。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"对已经按BUFF高32位在前布局的words，前置count:uint32、bits_per_value=64、lower_bits=0；空输入输出空包。逆向核对这三个头字段后返回原words。","example":"已有两项重排后的64位word，前置count=2、bits_per_value=64、lower_bits=0。解码核对头后取出16字节word载荷，再由A052恢复字序。","id":"A101","inputs":"words:uint8[8N]，每字高32位在前、各肢内部小端","layer":"adapter","name":"BUFF Raw64 Header","outputs":"1流BUFF raw64 wire:bytes。N≤UINT32_MAX，无公开参数。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"解析B2P1的logical_id与block body；按记录的物理顺序拼body，并为各逻辑块重建offset表。direct/memcpy/special-zero模式走专门分支，最后写Blosc2头与扩展字段。","example":"两个逻辑块L0、L1在物理区按L1后L0存放，offset表分别指向各块的实际开始地址。解码按逻辑顺序索引块，不把物理拼接次序误当作列顺序。","id":"A102","inputs":"B2P1:bytes，含header/extension、direct、logical_id:uint32、body:bytes","layer":"adapter","name":"Blosc2 Logical-to-physical Block Frame","outputs":"1流Blosc2 chunk:bytes。当前schema固定version=5、codec_version=1、typesize=1。","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"Container / Framing / 容器与封包","reference_ids":["O26"]},{"description":"Splits fixed records, floating-point fields, numeric bit fields, or serialized fixed-width bitstreams into typed component streams while retaining the metadata required for exact reconstruction. The four current entries are project implementations informed by OpenZL layout nodes; their local transport is not the OpenZL wire format.","family_number":"29","group":"Layout and state","id":"split_columns","implementations":[{"description":"把输入解释为 N 个宽度 W 的定长记录；transpose 按记录内字节位置输出 W 条 byte 列。split-struct 则按给定字段宽度数组切连续字节字段，各字段仍保持行序。1/4/8-byte shuffle 是 W 的固定 profile。","example":"两条宽3字节记录[01,02,03]和[11,12,13]，transpose输出三列[01,11]、[02,12]、[03,13]。若字段宽[1,2]，输出1字节列和2字节结构列。","id":"A049","inputs":"Struct(W) 的 N 条定长记录；兼容 Serial adapter 可接任意 byte[L]，仅完整记录参与列化并保留尾部。参数：W 或字段宽度数组，当前兼容 profile W∈{1,4,8}。","layer":"mechanism","name":"Parameterized Fixed-record Byte Columnization（参数化定宽记录列化）","outputs":"transpose 为 W 条 bytes[N]；split-struct 为字段数条 Struct(field_width) 流。","reference_ids":["O27"],"source_url":"https://github.com/facebook/openzl","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"按 fp16、bf16 或 fp32 位格式读取原始位型，将 fraction 左移一位并把 sign 填入最低位，形成 sign_fraction；exponent 单独取出。逆变换再把 sign、fraction、exponent 放回原位，因此保留 NaN、无穷和 signed zero 的位型。","example":"fp32的-0原位0x80000000：exponent=0、fraction=0、sign=1，输出sign_fraction=1、exponent=0。fp32的1.0则输出sign_fraction=0、exponent=127。","id":"A103","inputs":"Numeric fp16/bf16/fp32 原始位型，存储宽度分别 2/2/4 bytes；参数：format。","layer":"mechanism","name":"Float Sign-fraction/Exponent Deconstruction（浮点符号尾数与指数拆分）","outputs":"2 条流--Struct(ceil((fraction_bits+1)/8)) 的 sign_fraction[N] 与 uint8 exponent[N]。当前实现不接 fp64。","reference_ids":["O27"],"source_url":"https://github.com/facebook/openzl","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"对每个 Numeric 元素从最低位起，依次按 widths=[w0,w1,…] 掩码并右移，得到若干定宽数值字段。各宽度为 1-64 bit，总宽不超过输入存储位数；若未覆盖高位，所有被省略高位必须为 0 才可逆。","example":"uint16值0x1234，widths=[4,4,8]，低位优先输出[4,3,0x12]。逆向4|(3<<4)|(0x12<<8)=0x1234。","id":"A104","inputs":"Numeric 整数或原始浮点位型，storage_bits 由类型确定；参数：widths，或 top8/浮点格式 profile。","layer":"mechanism","name":"LSB-first Bit-field Split（低位优先位域拆分）","outputs":"len(widths) 条 Numeric 流，每条为能容纳对应 wi 的最小存储宽度；浮点 profile 输出 fraction、exponent、sign 3 流。","reference_ids":["O27"],"source_url":"https://github.com/facebook/openzl","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"},{"description":"把 Serial byte 流视为 LSB-first 连续位串，每 bits 位读出一个无符号值，共 floor(8L/bits) 个；末尾不足 bits 的 remainder 位值写入 side metadata。逆变换按相同位序重打包并补回 remainder。","example":"单字节0xd6按LSB-first、bits=3解包：得到[6,2]，剩余高2位值3作为remainder保存。逆向6|(2<<3)|(3<<6)=0xd6。","id":"A105","inputs":"Serial bytes[L]；参数bits∈[1,64]。N=floor(8L/bits)，要求ceil(N*bits/8)=L，即未消费部分不足一个完整字节。","layer":"mechanism","name":"Fixed-width Bitstream Unpack（定宽位流解包）","outputs":"Numeric unsigned[N]，每元素按storage_width(bits)占用1/2/4/8字节；其存储字节数可大于输入。side metadata保存bits与末尾remainder_value，原输入长度L随框架保存。","reference_ids":["O27"],"source_url":"https://github.com/facebook/openzl","status":"源码机制与调用入口已定位；本手册不把目录项数作为可执行内核数。"}],"name":"Split Columns / 列拆分","reference_ids":["O27","S01"]},{"description":"Produces causal probabilities before the current event is observed. A106-A109 expose next-bit P(1) from distinct cmix-family models, while A110 exposes a 256-way next-byte distribution from NNCP. These providers remain separate from graph-level entropy integration and benchmark method scheduling.","family_number":"30","group":"Probability coding","id":"prediction_probability","implementations":[{"description":"PPMd 按最近字节上下文维护自适应符号概率。adapter 在每个 bit 前从模型取 P(1)，观察 8 个 MSB-first bit 后才组装字节并调用 ByteUpdate；字节边界还可把 256 个符号权重归一化为完整下一字节分布。","example":"输入字节A=0x41，对应MSB-first事件0,1,0,0,0,0,0,1。每个事件之前读取模型的P(1)，随后观察该bit；第8次观察后执行ByteUpdate。八个实际概率由固定PPMd状态产生，不用常数概率代替。","id":"A106","inputs":"状态化命令流；每步 observe 输入 uint8 bit∈{0,1}，批量接口输入 bytes[L] 并按 MSB-first 展开。参数：order 2-64（默认 6）、memory 4-256 MiB（默认 16）。","layer":"mechanism","name":"cmix PPMd Byte-model Probability（PPMd 字节模型概率）","outputs":"每 bit 1 �� float64 P(1) 流；字节边界可选 256×float64 分布。","reference_ids":["O29"],"source_url":"https://github.com/byronknoll/cmix","status":"真实来源模型与adapter已完成本轮PPMd/DMC/PAQ共8项有限回归；尚未接入图熵编码边，不能据此声称完整方法测评完成。"},{"description":"在 bit 状态图当前节点保存定点 c0、c1；观察 y 后用 x←x·63/64+y·1024 更新计数和有限状态，达到阈值时按访问占比分裂后继节点，随后沿 nx[y] 转移。预测为 (c1+1)·4096/(c0+c1+2)。","example":"在已给定节点c0=c1=1024时，预测整数概率为(1024+1)×4096/(2048+2)=2048，即P(1)=0.5；接收bit=1后再更新计数并沿nx[1]转移。此例是节点状态计算，不声称为默认初始节点。","id":"A107","inputs":"与 A106 相同的因果 uint8 bit 观察流或 MSB-first bytes[L]；当前参数固定 extra_nodes=65536、cloning_threshold_fixed_point=1024。","layer":"mechanism","name":"cmix PAQ dmcModel Count Branch（PAQ DMC 计数分支）","outputs":"1 条 float64 P(1) 流，由原生 12-bit 整数概率除以 4096；无字节分布流。","reference_ids":["O29"],"source_url":"https://github.com/byronknoll/cmix","status":"真实来源模型与adapter已完成本轮PPMd/DMC/PAQ共8项有限回归；尚未接入图熵编码边，不能据此声称完整方法测评完成。"},{"description":"运行 PAQ8 Predictor：contextModel2 先汇集上下文模型，再按内容状态经过 APM/APM1 校准与加权，形成 12-bit 最终 pr；每次 observe 把真实 bit 写入全局 y 并执行完整 update，字节完成时同步缓冲区和上下文。","example":"对输入bits=[0,1]，先取得最终12-bit pr0，再observe(0)更新整套PAQ状态；取pr1后再observe(1)。pr0/pr1是混合和APM校准后的最终输出，不能用某个子模型值替代。","id":"A108","inputs":"因果 uint8 bit 流或 MSB-first bytes[L]；参数：paq_level 0-3（默认 0），用于选择内存级别。","layer":"mechanism","name":"cmix PAQ8 Hybrid Final Predictor（PAQ8 混合最终预测器）","outputs":"1 条 float64 P(1) 流，值为原生最终 p()/4096；无压缩字节或额外语义流。","reference_ids":["O29"],"source_url":"https://github.com/byronknoll/cmix","status":"真实来源模型与adapter已完成本轮PPMd/DMC/PAQ共8项有限回归；尚未接入图熵编码边，不能据此声称完整方法测评完成。"},{"description":"完整 Predictor 收集 models、byte_models 与 byte_mixers 的输出，送入分层 mixers；最终经 logistic 与 SSE 校准得到 P(1)，observe 时逐一更新模型、mixer、SSE 和上下文管理器。当前构造使用全 256 字节 vocabulary，不做预处理或预训练。","example":"输入字节00时，依次执行8次predict→处理当前0→observe(0)；各层mixer与SSE在观察后更新。解码端从相同256字节词表及模型初始化开始重放相同8次状态转移。","id":"A109","inputs":"因果 uint8 bit 流或 MSB-first bytes[L]，vocabulary 固定为全部 256 个字节；参数：allow_large_memory=true 才允许启动，adapter 另有 timeout。","layer":"mechanism","name":"cmix v21 Full Predictor（cmix v21 完整预测器）","outputs":"1 条 float64 P(1) 流，由上游 float 转为传输用 binary64；无压缩数据流。","reference_ids":["O29"],"source_url":"https://github.com/byronknoll/cmix","status":"来源模型及adapter见源码；可运行性受构建依赖与内存门控约束。本次未执行cmix或验证其图熵编码接入。"},{"description":"调用官方NNCP 3.3 Transformer与LibNC得到下一字节的256项softmax概率。先predict，再由匹配的write_sym/read_sym处理当前事件，最后observe以该字节为训练目标执行gradient和model_update；模型状态、初始化及重置在解码端一致。","example":"以初始previous=0预测首字节t0的256项分布，write_sym/read_sym处理t0后再observe(t0)，调用原模型gradient/update；下一步以已知t0预测t1。重复predict返回同一缓存分布。独立NNCPSymbolSession和NNP1冷解码路径已接入；不把概率向量当作压缩数据。","id":"A110","inputs":"逐个 uint8 字节事件；必须先 predict，再对当前字节 encode/decode，最后 observe。每会话最多 4096 字节。","layer":"mechanism","name":"NNCP Transformer Byte Probability (stream1)","outputs":"float32[256]，按符号0..255排列。transformer-stream1-v1固定batch=1、seg_len=1、seed=123、V=256，保留上游default网络结构；每会话最多4096字节。官方LibNC依赖Linux x86-64，本机通过显式Docker验证。独立NNP1压缩路径有24字节固定头和完整算术码体；未接图熵边或86方法调度。","reference_ids":["O28"],"source_url":"https://bellard.org/nncp/","status":"真实NNCP独立字节概率接口与配套NNP1冷解码路径已接入；官方LibNC要求Linux x86-64或显式Docker。与图熵编码边、默认NNCP方法评测分开。"}],"name":"Prediction Probability / 概率预测","reference_ids":["O28","O29"]}],"references":[{"citation":"Panagiotis Liakos, Katia Papakonstantinopoulou, Yannis Kotidis. “Chimp.” Proceedings of the VLDB Endowment, 2022.","id":"M01","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.14778/3551793.3551852","verification_status":"web_verified_primary_or_official_2026-10-08"},{"citation":"Yufeng Liu, Yao Shen, Fenghua Zhang, Feiteng Huang. “Pongo: Efficient Lossless Floating Point Compression.” Cloud Computing and Data Science, 2025.","id":"M02","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.37256/ccds.6220257102","verification_status":"publisher_blocked_403_not_a_dead_link"},{"citation":"Tuomas Pelkonen, Scott Franklin, Justin Teller, Paul Cavallaro, Qi Huang, Justin Meza, Kaushik Veeraraghavan. “Gorilla.” Proceedings of the VLDB Endowment, 2015.","id":"M03","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.14778/2824032.2824078","verification_status":"web_verified_primary_or_official_2026-10-08"},{"citation":"Apache Software Foundation. “Apache TsFile Encoding (TS_2DIFF).” Official documentation or source repository, n.d.","id":"M04","kind":"official_documentation_or_project_page","source_slide":2,"url":"https://tsfile.apache.org/","verification_status":"web_verified_primary_or_official_2026-10-08"},{"citation":"Jinzhao Xiao, Zihan Guo, Shaoxu Song. “BOS: Bit-Packing with Outlier Separation.” 2025 IEEE 41st International Conference on Data Engineering (ICDE), 2025.","id":"M05","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.1109/ICDE65448.2025.00107","verification_status":"doi_resolved_to_primary_publisher_2026-10-08"},{"citation":"Jinzhao Xiao, Wendi He, Shaoxu Song, Xiangdong Huang, Chen Wang, Jianmin Wang. “REGER: Reordering Time Series Data for Regression Encoding.” 2024 IEEE 40th International Conference on Data Engineering (ICDE), 2024.","id":"M06","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.1109/ICDE60146.2024.00100","verification_status":"doi_resolved_to_primary_publisher_2026-10-08"},{"citation":"Daniel Lemire, Nathan Kurz, Christoph Rupp. “Stream VByte : Faster byte-oriented integer compression.” Information Processing Letters, 2018.","id":"M07","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.1016/j.ipl.2017.09.011","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"Daniel Lemire, Leonid Boytsov, Nathan Kurz. “SIMD compression and the intersection of sorted integers.” Software: Practice and Experience, 2015.","id":"M08","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.1002/spe.2326","verification_status":"publisher_blocked_403_not_a_dead_link"},{"citation":"lemire project contributors. “FastPFor/Simple8b reference implementation.” Official documentation or source repository, n.d.","id":"M09","kind":"official_source_repository","source_slide":2,"url":"https://github.com/lemire/FastPFor","verification_status":"web_verified_primary_or_official_2026-10-08"},{"citation":"Apache Software Foundation. “Run Length Encoding / Bit-Packing Hybrid.” Official documentation or source repository, n.d.","id":"M10","kind":"official_documentation_or_project_page","source_slide":2,"url":"https://parquet.apache.org/docs/file-format/data-pages/encodings/#run-length-encoding-bit-packing-hybrid-rle-3","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"Julien Spiegel, Patrice Wira, Gilles Hermann. “A Comparative Experimental Study of Lossless Compression Algorithms for Enhancing Energy Efficiency in Smart Meters.” 2018 IEEE 16th International Conference on Industrial Informatics (INDIN), 2018.","id":"M11","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.1109/INDIN.2018.8471921","verification_status":"doi_resolved_to_primary_publisher_2026-10-08"},{"citation":"Davis Blalock, Samuel Madden, John Guttag. “Sprintz.” Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 2018.","id":"M12","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.1145/3264903","verification_status":"publisher_blocked_403_not_a_dead_link"},{"citation":"Daniel Abadi, Samuel Madden, Miguel Ferreira. “Integrating compression and execution in column-oriented database systems.” Proceedings of the 2006 ACM SIGMOD international conference on Management of data, 2006.","id":"M13","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.1145/1142473.1142548","verification_status":"publisher_blocked_403_not_a_dead_link"},{"citation":"Apache Software Foundation. “Apache Parquet Encodings: DELTA_BINARY_PACKED.” Official documentation or source repository, n.d.","id":"M14","kind":"official_documentation_or_project_page","source_slide":2,"url":"https://parquet.apache.org/docs/file-format/data-pages/encodings/#delta-encoding-delta_binary_packed-5","verification_status":"web_verified_primary_or_official_2026-10-08"},{"citation":"Azim Afroozeh, Leonardo X. Kuffo, Peter Boncz. “ALP: Adaptive Lossless floating-Point Compression.” Proceedings of the ACM on Management of Data, 2023.","id":"M15","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.1145/3626717","verification_status":"publisher_blocked_403_not_a_dead_link"},{"citation":"Yihao Liu, Xinyu Zeng, Huanchen Zhang. “LeCo: Lightweight Compression via Learning Serial Correlations.” Proceedings of the ACM on Management of Data, 2024.","id":"M16","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.1145/3639320","verification_status":"publisher_blocked_403_not_a_dead_link"},{"citation":"Yuanyuan Yao, Lu Chen, Ziquan Fang, Yunjun Gao, Christian S. Jensen, Tianyi Li. “Camel: Efficient Compression of Floating-Point Time Series.” Proceedings of the ACM on Management of Data, 2024.","id":"M17","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.1145/3698802","verification_status":"publisher_blocked_403_not_a_dead_link"},{"citation":"Ruiyuan Li, Zheng Li, Yi Wu, Chao Chen, Yu Zheng. “Elf: Erasing-Based Lossless Floating-Point Compression.” Proceedings of the VLDB Endowment, 2023.","id":"M18","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.14778/3587136.3587149","verification_status":"web_verified_primary_or_official_2026-10-08"},{"citation":"Ruiyuan Li et al. “Erasing-based Lossless Compression Method for Streaming Floating-Point Time Series.” arXiv:2306.16053, 2023.","id":"M19","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.48550/arXiv.2306.16053","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"Martin Burtscher, Paruj Ratanaworabhan. “High Throughput Compression of Double-Precision Floating-Point Data.” 2007 Data Compression Conference (DCC'07), 2007.","id":"M20","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.1109/DCC.2007.44","verification_status":"doi_resolved_to_primary_publisher_2026-10-08"},{"citation":"Martin Burtscher, Paruj Ratanaworabhan. “pFPC: A Parallel Compressor for Floating-Point Data.” 2009 Data Compression Conference, 2009.","id":"M21","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.1109/DCC.2009.43","verification_status":"doi_resolved_to_primary_publisher_2026-10-08"},{"citation":"Steven Claggett, Sahar Azimi, Martin Burtscher. “SPDP: An Automatically Synthesized Lossless Compression Algorithm for Floating-Point Data.” 2018 Data Compression Conference, 2018.","id":"M22","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.1109/DCC.2018.00042","verification_status":"doi_resolved_to_primary_publisher_2026-10-08"},{"citation":"Abdelouahab Khelifati, Mourad Khayati, Philippe Cudre-Mauroux. “CORAD: Correlation-Aware Compression of Massive Time Series using Sparse Dictionary Coding.” 2019 IEEE International Conference on Big Data (Big Data), 2019.","id":"M23","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.1109/BigData47090.2019.9005580","verification_status":"doi_resolved_to_primary_publisher_2026-10-08"},{"citation":"Chunwei Liu, Hao Jiang, John Paparrizos, Aaron J. Elmore. “Decomposed bounded floats for fast compression and queries.” Proceedings of the VLDB Endowment, 2021.","id":"M24","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.14778/3476249.3476305","verification_status":"web_verified_primary_or_official_2026-10-08"},{"citation":"Peter Lindstrom, Martin Isenburg. “Fast and Efficient Compression of Floating-Point Data.” IEEE Transactions on Visualization and Computer Graphics, 2006.","id":"M25","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.1109/TVCG.2006.143","verification_status":"doi_resolved_to_primary_publisher_2026-10-08"},{"citation":"Noushin Azami, Alex Fallin, Martin Burtscher. “Efficient Lossless Compression of Scientific Floating-Point Data on CPUs and GPUs.” Proceedings of the 30th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 1, 2025.","id":"M26","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.1145/3669940.3707280","verification_status":"publisher_blocked_403_not_a_dead_link"},{"citation":"P. Deutsch. “GZIP file format specification version 4.3.” RFC Editor, 1996.","id":"M27","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.17487/RFC1952","verification_status":"web_verified_primary_or_official_2026-10-08"},{"citation":"J. Alakuijala, Z. Szabadka. “Brotli Compressed Data Format.” RFC Editor, 2016.","id":"M28","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.17487/RFC7932","verification_status":"web_verified_primary_or_official_2026-10-08"},{"citation":"Y. Collet. “Zstandard Compression and the 'application/zstd' Media Type.” RFC Editor, 2021.","id":"M29","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.17487/RFC8878","verification_status":"web_verified_primary_or_official_2026-10-08"},{"citation":"lz4 project contributors. “LZ4 Frame Format Specification.” Official documentation or source repository, n.d.","id":"M30","kind":"standard_or_specification","source_slide":2,"url":"https://github.com/lz4/lz4/","verification_status":"web_verified_primary_or_official_2026-10-08"},{"citation":"The Tukaani Project. “XZ File Format Specification.” Official documentation or source repository, n.d.","id":"M31","kind":"standard_or_specification","source_slide":2,"url":"https://tukaani.org/xz/format.html","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"g1mv project contributors. “Density compression library.” Official documentation or source repository, n.d.","id":"M32","kind":"official_source_repository","source_slide":2,"url":"https://github.com/g1mv/density","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"google project contributors. “Gipfeli compression library.” Official documentation or source repository, n.d.","id":"M33","kind":"official_source_repository","source_slide":2,"url":"https://github.com/google/gipfeli","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"MATCODE Software. “LZMAT: Open-source Extremely Fast Data Compression Library, Version 1.0.” Official project documentation (archived), 2008.","id":"M35","kind":"archived_project_page_or_source_mirror","source_slide":2,"url":"https://web.archive.org/web/20160306180328/http://www.matcode.com/lzmat.htm","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"oberhumer.com. “LZO real-time compression library.” Official documentation or source repository, n.d.","id":"M36","kind":"official_documentation_or_project_page","source_slide":2,"url":"http://www.oberhumer.com/opensource/lzo/","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"Ross N. Williams. “An Extremely Fast Ziv-Lempel Data Compression Algorithm.” Proceedings of the Data Compression Conference, 1991.","id":"M37","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.1109/DCC.1991.213344","verification_status":"doi_resolved_to_primary_publisher_2026-10-08"},{"citation":"conorstokes.github.io. “An LZ Codec Designed for SSE Decompression.” Official documentation or source repository, n.d.","id":"M38","kind":"official_documentation_or_project_page","source_slide":2,"url":"http://conorstokes.github.io/compression/2016/02/15/an-LZ-codec-designed-for-SSE-decompression","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"lzfse project contributors. “LZFSE and LZVN reference implementation.” Official documentation or source repository, n.d.","id":"M39","kind":"official_source_repository","source_slide":2,"url":"https://github.com/lzfse/lzfse","verification_status":"web_verified_primary_or_official_2026-10-08"},{"citation":"johnezang project contributors. “Pithy compression library.” Official documentation or source repository, n.d.","id":"M40","kind":"official_source_repository","source_slide":2,"url":"https://github.com/johnezang/pithy","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"Compression.ru archive. “SHRINKER compressor description.” Official documentation or source repository, n.d.","id":"M41","kind":"archived_project_page_or_source_mirror","source_slide":2,"url":"https://compression.ru/arctest/self/shrinker.htm","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"oberhumer.com. “UCL data compression library.” Official documentation or source repository, n.d.","id":"M42","kind":"official_documentation_or_project_page","source_slide":2,"url":"http://www.oberhumer.com/opensource/ucl/","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"code.google.com. “WFLZ compression library.” Official documentation or source repository, n.d.","id":"M43","kind":"archived_project_page_or_source_mirror","source_slide":2,"url":"https://code.google.com/p/wflz/","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"ebiggers project contributors. “XPACK experimental compression format.” Official documentation or source repository, n.d.","id":"M44","kind":"standard_or_specification","source_slide":2,"url":"https://github.com/ebiggers/xpack","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"Ivan Tkatchev. “yalz77: Yet Another LZ77.” Official source repository, 2015.","id":"M45","kind":"official_source_repository","source_slide":2,"url":"https://github.com/ivan-tkatchev/yalz77","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"web.archive.org. “Yappy compression algorithm.” Official documentation or source repository, n.d.","id":"M46","kind":"archived_project_page_or_source_mirror","source_slide":2,"url":"https://web.archive.org/web/20110617184153/blog.gamedeff.com/?p=371","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"google project contributors. “Snappy compression format.” Official documentation or source repository, n.d.","id":"M47","kind":"standard_or_specification","source_slide":2,"url":"https://github.com/google/snappy","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"Jarek Duda. “Asymmetric Numeral Systems.” arXiv:0902.0271, 2009.","id":"M48","kind":"official_documentation_or_project_page","source_slide":2,"url":"https://arxiv.org/abs/0902.0271","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"Cyan4973 project contributors. “FiniteStateEntropy.” Official documentation or source repository, n.d.","id":"M49","kind":"official_source_repository","source_slide":2,"url":"https://github.com/Cyan4973/FiniteStateEntropy","verification_status":"web_verified_primary_or_official_2026-10-08"},{"citation":"David Huffman. “A Method for the Construction of Minimum-Redundancy Codes.” Proceedings of the IRE, 1952.","id":"M50","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.1109/JRPROC.1952.273898","verification_status":"doi_resolved_to_primary_publisher_2026-10-08"},{"citation":"Ian H. Witten, Radford M. Neal, John G. Cleary. “Arithmetic coding for data compression.” Communications of the ACM, 1987.","id":"M51","kind":"research_paper_or_standard","source_slide":2,"url":"https://doi.org/10.1145/214762.214771","verification_status":"publisher_blocked_403_not_a_dead_link"},{"citation":"Blosc Development Team. “Blosc: A blocking, shuffling and compression filter.” Official documentation or source repository, n.d.","id":"M52","kind":"official_documentation_or_project_page","source_slide":2,"url":"https://www.blosc.org/pages/blosc-in-depth/","verification_status":"web_verified_primary_or_official_2026-10-08"},{"citation":"jibsen project contributors. “BriefLZ fast compression library.” Official documentation or source repository, n.d.","id":"M53","kind":"official_source_repository","source_slide":2,"url":"https://github.com/jibsen/brieflz","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"IlyaGrebnov project contributors. “libbsc: Block Sorting Compressor.” Official documentation or source repository, n.d.","id":"M54","kind":"official_source_repository","source_slide":2,"url":"https://github.com/IlyaGrebnov/libbsc","verification_status":"web_verified_primary_or_official_2026-10-08"},{"citation":"sourceware.org. “bzip2 and libbzip2.” Official documentation or source repository, n.d.","id":"M55","kind":"official_documentation_or_project_page","source_slide":2,"url":"https://www.sourceware.org/bzip2/","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"kspalaiologos project contributors. “bzip3: a better and stronger successor to bzip2.” Official documentation or source repository, n.d.","id":"M56","kind":"official_source_repository","source_slide":2,"url":"https://github.com/kspalaiologos/bzip3","verification_status":"web_verified_primary_or_official_2026-10-08"},{"citation":"sourceforge.net. “CRUSH compression library.” Official documentation or source repository, n.d.","id":"M57","kind":"official_documentation_or_project_page","source_slide":2,"url":"https://sourceforge.net/projects/crush/","verification_status":"publisher_blocked_403_not_a_dead_link"},{"citation":"fusiyuan2010 project contributors. “CSC lossless compression algorithm.” Official documentation or source repository, n.d.","id":"M58","kind":"official_source_repository","source_slide":2,"url":"https://github.com/fusiyuan2010/CSC","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"ariya project contributors. “FastLZ compression library.” Official documentation or source repository, n.d.","id":"M59","kind":"official_source_repository","source_slide":2,"url":"https://github.com/ariya/FastLZ","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"conor42 project contributors. “Fast LZMA2 compression library.” Official documentation or source repository, n.d.","id":"M60","kind":"official_source_repository","source_slide":2,"url":"https://github.com/conor42/fast-lzma2","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"flanglet project contributors. “Kanzi: high-performance compression framework.” Official documentation or source repository, n.d.","id":"M61","kind":"official_source_repository","source_slide":2,"url":"https://github.com/flanglet/kanzi-cpp","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"ebiggers project contributors. “libdeflate.” Official documentation or source repository, n.d.","id":"M62","kind":"official_source_repository","source_slide":2,"url":"https://github.com/ebiggers/libdeflate","verification_status":"web_verified_primary_or_official_2026-10-08"},{"citation":"inikep project contributors. “Lizard (LZ5) fast compression.” Official documentation or source repository, n.d.","id":"M63","kind":"official_source_repository","source_slide":2,"url":"https://github.com/inikep/lizard","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"lz4.org. “LZ4 fast compression mode.” Official documentation or source repository, n.d.","id":"M64","kind":"official_documentation_or_project_page","source_slide":2,"url":"https://lz4.org/","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"avaneev project contributors. “LZAV fast LZ77 compressor.” Official documentation or source repository, n.d.","id":"M65","kind":"official_source_repository","source_slide":2,"url":"https://github.com/avaneev/lzav","verification_status":"web_verified_primary_or_official_2026-10-08"},{"citation":"software.schmorp.de. “LibLZF very fast compression library.” Official documentation or source repository, n.d.","id":"M66","kind":"official_documentation_or_project_page","source_slide":2,"url":"http://software.schmorp.de/pkg/liblzf.html","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"mbitsnbites project contributors. “LibLZG lightweight compression library.” Official documentation or source repository, n.d.","id":"M67","kind":"official_source_repository","source_slide":2,"url":"https://github.com/mbitsnbites/liblzg","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"richgel999 project contributors. “LZHAM compression codec.” Official documentation or source repository, n.d.","id":"M68","kind":"official_source_repository","source_slide":2,"url":"https://github.com/richgel999/lzham_codec","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"openzfs project contributors. “OpenZFS LZJB reference source.” Official documentation or source repository, n.d.","id":"M69","kind":"official_source_repository","source_slide":2,"url":"https://github.com/openzfs/zfs/blob/master/module/zfs/lzjb.c","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"nongnu.org. “Lzlib data compression library.” Official documentation or source repository, n.d.","id":"M70","kind":"official_documentation_or_project_page","source_slide":2,"url":"https://www.nongnu.org/lzip/lzlib.html","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"7-zip.org. “PPMd statistical compression (PPMD8).” Official documentation or source repository, n.d.","id":"M71","kind":"official_documentation_or_project_page","source_slide":2,"url":"https://www.7-zip.org/","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"web.archive.org. “QuickLZ compression library.” Official documentation or source repository, n.d.","id":"M72","kind":"archived_project_page_or_source_mirror","source_slide":2,"url":"https://web.archive.org/web/20160110073818/https://quicklz.com/","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"libslz.org. “libslz fast gzip/deflate encoder.” Official documentation or source repository, n.d.","id":"M73","kind":"official_documentation_or_project_page","source_slide":2,"url":"http://www.libslz.org/","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"nemequ project contributors. “Tornado compressor reference source.” Official documentation or source repository, n.d.","id":"M74","kind":"official_source_repository","source_slide":2,"url":"https://github.com/nemequ/tornado","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"zlib.net. “zlib compression library.” Official documentation or source repository, n.d.","id":"M75","kind":"official_documentation_or_project_page","source_slide":2,"url":"https://zlib.net/","verification_status":"web_verified_primary_or_official_2026-10-08"},{"citation":"zlib-ng project contributors. “zlib-ng optimized zlib fork.” Official documentation or source repository, n.d.","id":"M76","kind":"official_source_repository","source_slide":2,"url":"https://github.com/zlib-ng/zlib-ng","verification_status":"web_verified_primary_or_official_2026-10-08"},{"citation":"richox project contributors. “libzling compression library.” Official documentation or source repository, n.d.","id":"M77","kind":"official_source_repository","source_slide":2,"url":"https://github.com/richox/libzling","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"hellobertrand project contributors. “ZXC compression library.” Official documentation or source repository, n.d.","id":"M78","kind":"official_source_repository","source_slide":2,"url":"https://github.com/hellobertrand/zxc","verification_status":"link_reachable_2026-10-08_content_from_ppt_notes"},{"citation":"BrianPugh project contributors. “TAMP tiny automatic meta-protocol compressor.” Official documentation or source repository, n.d.","id":"M79","kind":"official_source_repository","source_slide":2,"url":"https://github.com/BrianPugh/tamp","verification_status":"web_verified_primary_or_official_2026-10-08"},{"citation":"Apache Software Foundation. “Apache Parquet Encodings: BYTE_STREAM_SPLIT.” Official documentation, n.d.","id":"S01","kind":"official_documentation_or_project_page","source_slide":null,"url":"https://parquet.apache.org/docs/file-format/data-pages/encodings/#byte-stream-split-byte_stream_split-9","verification_status":"web_verified_primary_or_official_2026-10-08"},{"citation":"Ruiyuan Li, Zheng Li, Yi Wu, Chao Chen, Yu Zheng. “Elf: Erasing-Based Lossless Floating-Point Compression.” Proceedings of the VLDB Endowment, 2023.","id":"O01","kind":"paper","source_slide":3,"url":"https://doi.org/10.14778/3587136.3587149","verification_status":"ppt_note_primary_source"},{"citation":"Azim Afroozeh, Leonardo X. Kuffo, Peter Boncz. “ALP: Adaptive Lossless floating-Point Compression.” Proceedings of the ACM on Management of Data, 2023.","id":"O02","kind":"paper","source_slide":3,"url":"https://doi.org/10.1145/3626717","verification_status":"ppt_note_primary_source"},{"citation":"Yihao Liu, Xinyu Zeng, Huanchen Zhang. “LeCo: Lightweight Compression via Learning Serial Correlations.” Proceedings of the ACM on Management of Data, 2024.","id":"O03","kind":"paper","source_slide":3,"url":"https://doi.org/10.1145/3639320","verification_status":"ppt_note_primary_source"},{"citation":"Jinzhao Xiao, Wendi He, Shaoxu Song, Xiangdong Huang, Chen Wang, Jianmin Wang. “REGER: Reordering Time Series Data for Regression Encoding.” 2024 IEEE 40th International Conference on Data Engineering (ICDE), 2024.","id":"O04","kind":"paper","source_slide":3,"url":"https://doi.org/10.1109/ICDE60146.2024.00100","verification_status":"ppt_note_primary_source"},{"citation":"Jinzhao Xiao, Zihan Guo, Shaoxu Song. “BOS: Bit-Packing with Outlier Separation.” 2025 IEEE 41st International Conference on Data Engineering (ICDE), 2025.","id":"O05","kind":"paper","source_slide":3,"url":"https://doi.org/10.1109/ICDE65448.2025.00107","verification_status":"ppt_note_primary_source"},{"citation":"Daniel Lemire, Leonid Boytsov, Nathan Kurz. “SIMD compression and the intersection of sorted integers.” Software: Practice and Experience, 2015.","id":"O06","kind":"paper","source_slide":3,"url":"https://doi.org/10.1002/spe.2326","verification_status":"ppt_note_primary_source"},{"citation":"Apache Software Foundation. “Apache Parquet DELTA_BINARY_PACKED specification.” Official documentation or source repository, n.d. (accessed 2026-09-11).","id":"O07","kind":"official_documentation","source_slide":3,"url":"https://parquet.apache.org/docs/file-format/data-pages/encodings/#delta-encoding-delta_binary_packed-5","verification_status":"ppt_note_official_source"},{"citation":"Tuomas Pelkonen, Scott Franklin, Justin Teller, Paul Cavallaro, Qi Huang, Justin Meza, Kaushik Veeraraghavan. “Gorilla.” Proceedings of the VLDB Endowment, 2015.","id":"O08","kind":"paper","source_slide":3,"url":"https://doi.org/10.14778/2824032.2824078","verification_status":"ppt_note_primary_source"},{"citation":"Panagiotis Liakos, Katia Papakonstantinopoulou, Yannis Kotidis. “Chimp.” Proceedings of the VLDB Endowment, 2022.","id":"O09","kind":"paper","source_slide":3,"url":"https://doi.org/10.14778/3551793.3551852","verification_status":"ppt_note_primary_source"},{"citation":"Daniel Abadi, Samuel Madden, Miguel Ferreira. “Integrating compression and execution in column-oriented database systems.” Proceedings of the 2006 ACM SIGMOD international conference on Management of data, 2006.","id":"O10","kind":"paper","source_slide":3,"url":"https://doi.org/10.1145/1142473.1142548","verification_status":"ppt_note_primary_source"},{"citation":"Protocol Buffers project. “Protocol Buffers encoding: ZigZag.” Official documentation or source repository, n.d. (accessed 2026-09-11).","id":"O11","kind":"official_documentation","source_slide":3,"url":"https://protobuf.dev/programming-guides/encoding/#signed-ints","verification_status":"verified_primary"},{"citation":"Michael Burrows, David J. Wheeler. “A Block-sorting Lossless Data Compression Algorithm.” Digital Systems Research Center Research Report 124, 1994.","id":"O12","kind":"paper","source_slide":3,"url":"https://www.cl.cam.ac.uk/teaching/2526/Bioinfo/papers/burrows_wheeler.pdf","verification_status":"verified_primary"},{"citation":"Jon Louis Bentley, Daniel D. Sleator, Robert E. Tarjan, Victor K. Wei. “A locally adaptive data compression scheme.” Communications of the ACM, 1986.","id":"O13","kind":"paper","source_slide":3,"url":"https://doi.org/10.1145/5684.5688","verification_status":"primary_url_preserved_resolver_unavailable"},{"citation":"Apache Software Foundation. “Run Length Encoding / Bit-Packing Hybrid.” Official documentation or source repository, n.d. (accessed 2026-09-11).","id":"O14","kind":"official_documentation","source_slide":3,"url":"https://parquet.apache.org/docs/file-format/data-pages/encodings/#run-length-encoding-bit-packing-hybrid-rle-3","verification_status":"ppt_note_official_source"},{"citation":"J. Ziv, A. Lempel. “A universal algorithm for sequential data compression.” IEEE Transactions on Information Theory, 1977.","id":"O15","kind":"paper","source_slide":3,"url":"https://doi.org/10.1109/TIT.1977.1055714","verification_status":"primary_url_preserved_resolver_unavailable"},{"citation":"facebook project contributors. “Zstandard compression format and implementation documentation.” Official documentation or source repository, n.d. (accessed 2026-09-11).","id":"O16","kind":"official_documentation","source_slide":3,"url":"https://github.com/facebook/zstd","verification_status":"ppt_note_official_source"},{"citation":"lz4 project contributors. “LZ4 Block Format Description.” Official documentation or source repository, n.d. (accessed 2026-09-11).","id":"O17","kind":"official_specification","source_slide":3,"url":"https://github.com/lz4/lz4/blob/dev/doc/lz4_Block_format.md","verification_status":"verified_primary"},{"citation":"Daniel Lemire, Nathan Kurz, Christoph Rupp. “Stream VByte : Faster byte-oriented integer compression.” Information Processing Letters, 2018.","id":"O18","kind":"paper","source_slide":3,"url":"https://doi.org/10.1016/j.ipl.2017.09.011","verification_status":"ppt_note_primary_source"},{"citation":"P. Deutsch. “DEFLATE Compressed Data Format Specification version 1.3.” RFC Editor, 1996.","id":"O19","kind":"official_specification","source_slide":3,"url":"https://doi.org/10.17487/RFC1951","verification_status":"primary_url_preserved_resolver_unavailable"},{"citation":"David Huffman. “A Method for the Construction of Minimum-Redundancy Codes.” Proceedings of the IRE, 1952.","id":"O20","kind":"paper","source_slide":3,"url":"https://doi.org/10.1109/JRPROC.1952.273898","verification_status":"ppt_note_primary_source"},{"citation":"Ian H. Witten, Radford M. Neal, John G. Cleary. “Arithmetic coding for data compression.” Communications of the ACM, 1987.","id":"O21","kind":"paper","source_slide":3,"url":"https://doi.org/10.1145/214762.214771","verification_status":"ppt_note_primary_source"},{"citation":"Jarek Duda. “Asymmetric Numeral Systems: Entropy Coding Combining Speed of Huffman Coding with Compression Rate of Arithmetic Coding.” arXiv:1311.2540, 2013.","id":"O22","kind":"paper","source_slide":3,"url":"https://arxiv.org/abs/1311.2540","verification_status":"verified_primary"},{"citation":"Steven Claggett, Sahar Azimi, Martin Burtscher. “SPDP: An Automatically Synthesized Lossless Compression Algorithm for Floating-Point Data.” 2018 Data Compression Conference, 2018.","id":"O23","kind":"paper","source_slide":3,"url":"https://doi.org/10.1109/DCC.2018.00042","verification_status":"ppt_note_primary_source"},{"citation":"Y. Collet. “Zstandard Compression and the 'application/zstd' Media Type.” RFC Editor, 2021.","id":"O24","kind":"official_specification","source_slide":3,"url":"https://doi.org/10.17487/RFC8878","verification_status":"ppt_note_primary_source"},{"citation":"Blosc Development Team. “Blosc in-depth: blocking, shuffling and compression.” Official documentation or source repository, n.d. (accessed 2026-09-11).","id":"O25","kind":"official_documentation","source_slide":3,"url":"https://www.blosc.org/pages/blosc-in-depth/","verification_status":"ppt_note_official_source"},{"citation":"P. Deutsch. “GZIP file format specification version 4.3.” RFC Editor, 1996.","id":"O26","kind":"official_specification","source_slide":3,"url":"https://doi.org/10.17487/RFC1952","verification_status":"ppt_note_primary_source"},{"citation":"Meta Platforms, Inc. “OpenZL.” Official source repository.","id":"O27","kind":"official_source","url":"https://github.com/facebook/openzl","verification_status":"primary_source_verified"},{"citation":"Fabrice Bellard. “NNCP: Lossless Data Compression with Neural Networks.” Official project page.","id":"O28","kind":"official_source","url":"https://bellard.org/nncp/","verification_status":"primary_source_verified"},{"citation":"Byron Knoll. “cmix version 21.” Official source repository.","id":"O29","kind":"official_source","url":"https://github.com/byronknoll/cmix","verification_status":"primary_source_verified"}],"schema_version":1,"source":{"additional_measured_families":[{"affiliation":"Project-specific composition of a Parquet BYTE_STREAM_SPLIT-style transform and LZ4 framing.","family":"BSS+LZ4","reference_ids":["S01","M30"]},{"affiliation":"Variant of PPT LZO1.","family":"LZO1X","reference_ids":["M36"]},{"affiliation":"Wrapper-specific split of PPT SLZ.","family":"Slz_deflate","reference_ids":["M73"]},{"affiliation":"Wrapper-specific split of PPT SLZ.","family":"Slz_gzip","reference_ids":["M73","M27"]},{"affiliation":"Wrapper-specific split of PPT SLZ.","family":"Slz_zlib","reference_ids":["M73","M75"]},{"affiliation":"Project-specific composition.","family":"Sub-column","reference_ids":["M09","M10","M13"]},{"affiliation":"Variant of PPT NRV.","family":"UCL NRV2b","reference_ids":["M42"]}],"description_note":"Method references follow presentation Notes (slides 2–3). Operator descriptions follow the final 30-family implementation catalog, preserving source-backed boundaries and reusing presentation citations with additional official sources.","notes_slides":[2,3],"operator_catalog":{"boundaries":["A013 is Minimum Frame of Reference only.","A048 is pure byte-plane split; minimum, scale, and width selection remain outside this mechanism.","STOP, Concat, and continuous Split are special graph operations and are not counted among the 30 families or 102 entries.","Complete backends and compatibility adapters are labeled separately from mechanisms.","A110 has a verified independent NNCP/NNP1 path but is not integrated into the graph entropy edge or default method schedule.","A111 has an independent three-stream C++ interface; the original fused FastPFor path remains separate."],"catalog_version":"operator-implementation-layers-v1","files":["docs/operator_102/review_current.json","benchmark/reference/actions/operator_implementations_layered_v1.json","docs/operator_102/interface_revision_status.json","docs/operator_102/source_inputs/comparison_01_05.json","docs/operator_102/source_inputs/comparison_06_10.json","docs/operator_102/source_inputs/comparison_11_20.json","docs/operator_102/source_inputs/comparison_21_30.json","docs/operator_102/source_inputs/comparison_27_root.json","docs/operator_102/source_inputs/comparison_28_root.json","docs/operator_102/source_inputs/comparison_29_30_final.json","deploy/huggingface-compressionbenchmark/qa/operator_catalog_content.json","deploy/huggingface-compressionbenchmark/qa/method_catalog_content.json"],"sha256":{"benchmark/reference/actions/operator_implementations_layered_v1.json":"09e41517b2e8aea04067546aa24f6d0e1dc4e6eef1475339114283a0c9acf252","deploy/huggingface-compressionbenchmark/qa/method_catalog_content.json":"5bc289539e6d3298cb4f2b6006c5503b7973f52707cf10219f878c2cabcc33df","deploy/huggingface-compressionbenchmark/qa/operator_catalog_content.json":"8ffdeda95b1599cdb3fcb846471cae39a9a51f6fbb4c3c94685b6dfd9bd49b53","docs/operator_102/interface_revision_status.json":"4075f9783b85767fda48c82cfda125a27d3f649529d6ae5ce989924daddc844f","docs/operator_102/review_current.json":"ac5e403edd361c04b4f0f4d33de3b089456a3f9cc958062704b050ddfc16fcaf","docs/operator_102/source_inputs/comparison_01_05.json":"78f88580fd1030969a4ff012add30a4f8d751dd7bf0ab1539fc87e48746911ae","docs/operator_102/source_inputs/comparison_06_10.json":"bfd843823e76fba92a7bfe310ac0b1de64a02642e66bb97b273c74ad0ee5eff9","docs/operator_102/source_inputs/comparison_11_20.json":"6ccff1fff27f874261fa1f07c1f344c79692705a398e9a49b417cfc9193b8135","docs/operator_102/source_inputs/comparison_21_30.json":"f1c09afcba5f0c0eb7a183dce94439e53be947c00c66a48268d8ceb1c7fcc4cf","docs/operator_102/source_inputs/comparison_27_root.json":"6de78e3539cabfd9495cb5d34f5dc3c8df22dcc5a51c8a9007e71857e6d048fd","docs/operator_102/source_inputs/comparison_28_root.json":"7ab1ac8902d079907d36b14d951b70faf442010946927ecf10423703d99c2165","docs/operator_102/source_inputs/comparison_29_30_final.json":"4913028f99abcc70e5915053c1603ed65f85fa7450f3a056ad409f845feca7c6"}},"ppt_method_aliases":[{"family":"CHIMP","index":1,"ppt_name":"Chimp","reference_ids":["M01"]},{"family":"Chimp128","index":2,"ppt_name":"Chimp128","reference_ids":["M01"]},{"family":"Pongo","index":3,"ppt_name":"Pongo","reference_ids":["M02"]},{"family":"GORILLA","index":4,"ppt_name":"Gorilla","reference_ids":["M03"]},{"family":"TS_2DIFF","index":5,"ppt_name":"2DIFF","reference_ids":["M04"]},{"family":"BOS","index":6,"ppt_name":"BOS","reference_ids":["M05"]},{"family":"REGER","index":7,"ppt_name":"REGER","reference_ids":["M06"]},{"family":"Stream VByte","index":8,"ppt_name":"SVB","reference_ids":["M07"]},{"family":"FastPFor","index":9,"ppt_name":"FastPFor","reference_ids":["M08"]},{"family":"Simple8b","index":10,"ppt_name":"Simple8b","reference_ids":["M09"]},{"family":"RLE","index":11,"ppt_name":"RLE","reference_ids":["M10"]},{"family":"RLBE","index":12,"ppt_name":"RLBE","reference_ids":["M11"]},{"family":"SPRINTZ","index":13,"ppt_name":"Sprintz","reference_ids":["M12"]},{"family":"Dictionary","index":14,"ppt_name":"Dict.","reference_ids":["M13"]},{"family":"DBP","index":15,"ppt_name":"DBP","reference_ids":["M14"]},{"family":"ALP","index":16,"ppt_name":"ALP","reference_ids":["M15"]},{"family":"LECO","index":17,"ppt_name":"LECO","reference_ids":["M16"]},{"family":"Camel","index":18,"ppt_name":"Camel","reference_ids":["M17"]},{"family":"Elf","index":19,"ppt_name":"Elf","reference_ids":["M18"]},{"family":"Elf+","index":20,"ppt_name":"Elf+","reference_ids":["M19"]},{"family":"FPC","index":21,"ppt_name":"FPC","reference_ids":["M20"]},{"family":"PFPC","index":22,"ppt_name":"PFPC","reference_ids":["M21"]},{"family":"SPDP","index":23,"ppt_name":"SPDP","reference_ids":["M22"]},{"family":"CORAD","index":24,"ppt_name":"CORAD","reference_ids":["M23"]},{"family":"BUFF","index":25,"ppt_name":"BUFF","reference_ids":["M24"]},{"family":"FPZIP","index":26,"ppt_name":"FPZIP","reference_ids":["M25"]},{"family":"FPCompress DPratio","index":27,"ppt_name":"FPCmp-R","reference_ids":["M26"]},{"family":"FPCompress DPspeed","index":28,"ppt_name":"FPCmp-S","reference_ids":["M26"]},{"family":"GZIP","index":29,"ppt_name":"gzip","reference_ids":["M27"]},{"family":"Brotli","index":30,"ppt_name":"Brotli","reference_ids":["M28"]},{"family":"ZSTD","index":31,"ppt_name":"Zstd","reference_ids":["M29"]},{"family":"LZ4","index":32,"ppt_name":"LZ4","reference_ids":["M30"]},{"family":"XZ","index":33,"ppt_name":"XZ","reference_ids":["M31"]},{"family":"Density","index":34,"ppt_name":"Density","reference_ids":["M32"]},{"family":"Gipfeli","index":35,"ppt_name":"Gipfeli","reference_ids":["M33"]},{"family":"LZMAT","index":37,"ppt_name":"LZMAT","reference_ids":["M35"]},{"family":"LZO1","index":38,"ppt_name":"LZO1","reference_ids":["M36"]},{"family":"LZO2A","index":39,"ppt_name":"LZO2A","reference_ids":["M36"]},{"family":"LZRW","index":40,"ppt_name":"LZRW","reference_ids":["M37"]},{"family":"LZSSE","index":41,"ppt_name":"LZSSE","reference_ids":["M38"]},{"family":"LZVN","index":42,"ppt_name":"LZVN","reference_ids":["M39"]},{"family":"Pithy","index":43,"ppt_name":"Pithy","reference_ids":["M40"]},{"family":"Shrinker","index":44,"ppt_name":"Shrinker","reference_ids":["M41"]},{"family":"UCL NRV","index":45,"ppt_name":"NRV","reference_ids":["M42"]},{"family":"WFLZ","index":46,"ppt_name":"WFLZ","reference_ids":["M43"]},{"family":"XPACK","index":47,"ppt_name":"XPACK","reference_ids":["M44"]},{"family":"YALZ77","index":48,"ppt_name":"YALZ77","reference_ids":["M45"]},{"family":"Yappy","index":49,"ppt_name":"Yappy","reference_ids":["M46"]},{"family":"SNAPPY","index":50,"ppt_name":"Snappy","reference_ids":["M47"]},{"family":"ANS","index":51,"ppt_name":"ANS","reference_ids":["M48"]},{"family":"FSE","index":52,"ppt_name":"FSE","reference_ids":["M49"]},{"family":"Huffman","index":53,"ppt_name":"Huffman","reference_ids":["M50"]},{"family":"Huff0","index":54,"ppt_name":"Huff0","reference_ids":["M49"]},{"family":"Arithmetic","index":55,"ppt_name":"Arithmetic","reference_ids":["M51"]},{"family":"Blosc","index":56,"ppt_name":"Blosc","reference_ids":["M52"]},{"family":"Blosc2","index":57,"ppt_name":"Blosc2","reference_ids":["M52"]},{"family":"BriefLZ","index":58,"ppt_name":"BriefLZ","reference_ids":["M53"]},{"family":"BSC","index":59,"ppt_name":"BSC","reference_ids":["M54"]},{"family":"Bzip2","index":60,"ppt_name":"Bzip2","reference_ids":["M55"]},{"family":"Bzip3","index":61,"ppt_name":"Bzip3","reference_ids":["M56"]},{"family":"Crush","index":62,"ppt_name":"Crush","reference_ids":["M57"]},{"family":"CSC","index":63,"ppt_name":"CSC","reference_ids":["M58"]},{"family":"FastLZ","index":64,"ppt_name":"FastLZ","reference_ids":["M59"]},{"family":"FastLZMA2","index":65,"ppt_name":"FastLZMA2","reference_ids":["M60"]},{"family":"Kanzi","index":66,"ppt_name":"Kanzi","reference_ids":["M61"]},{"family":"Libdeflate","index":67,"ppt_name":"Libdeflate","reference_ids":["M62"]},{"family":"Lizard","index":68,"ppt_name":"Lizard","reference_ids":["M63"]},{"family":"LZ4 Fast","index":69,"ppt_name":"LZ4 Fast","reference_ids":["M64"]},{"family":"LZ4HC","index":70,"ppt_name":"LZ4HC","reference_ids":["M64"]},{"family":"LZAV","index":71,"ppt_name":"LZAV","reference_ids":["M65"]},{"family":"LZF","index":72,"ppt_name":"LZF","reference_ids":["M66"]},{"family":"LZFSE","index":73,"ppt_name":"LZFSE","reference_ids":["M39"]},{"family":"LZG","index":74,"ppt_name":"LZG","reference_ids":["M67"]},{"family":"LZHAM","index":75,"ppt_name":"LZHAM","reference_ids":["M68"]},{"family":"LZJB","index":76,"ppt_name":"LZJB","reference_ids":["M69"]},{"family":"Lzlib","index":77,"ppt_name":"Lzlib","reference_ids":["M70"]},{"family":"PPMD8","index":78,"ppt_name":"PPMD8","reference_ids":["M71"]},{"family":"QuickLZ","index":79,"ppt_name":"QuickLZ","reference_ids":["M72"]},{"family":"SLZ","index":80,"ppt_name":"SLZ","reference_ids":["M73"]},{"family":"Tornado","index":81,"ppt_name":"Tornado","reference_ids":["M74"]},{"family":"Zlib","index":82,"ppt_name":"Zlib","reference_ids":["M75"]},{"family":"Zlib-ng","index":83,"ppt_name":"Zlib-ng","reference_ids":["M76"]},{"family":"Zling","index":84,"ppt_name":"Zling","reference_ids":["M77"]},{"family":"ZXC","index":85,"ppt_name":"ZXC","reference_ids":["M78"]},{"family":"TAMP","index":86,"ppt_name":"TAMP","reference_ids":["M79"]}],"presentation":"编码压缩基准.pptx","presentation_sha256":"4a5bed20f2b9f7ba18bf18569ad50b445e3bc8bfe6fe3020c5f68df424e1888e"},"updated_at":"2026-10-08T16:01:55.749706+00:00"}