% divANS Module

The divANS crate is meant to be used for generic data compression. The algorithm has been tuned to significantly favor gains in compression ratio over performance, operating at line speeds of 150 Mbit/s.

The name originates from "divided-ANS" since the intermediate representation is divided from the ANS codec

More information at https://blogs.dropbox.com/tech/2018/06/building-better-compression-together-with-divans/

Divans should primarily be considered for cold storage and compression research. The compression algorithm is highly modular and new algorithms only need to be written a single time since generic trait specialization constructs optimized variants of the codec for both compression and decompression at compile time.

extern crate divans;
fn main() {
    use std::io;
    let stdin = &mut io::stdin();
    {
        use std::io::{Read, Write};
        let mut reader = divans::DivansDecompressorReader::new(
            stdin,
            4096, // buffer size
        );
        io::copy(&mut reader, &mut io::stdout()).unwrap();
    }   
}

extern crate divans;
fn main() {
    use std::io;
    let stdout = &mut io::stdout();
    {
        use std::io::{Read, Write};
        let mut writer = divans::DivansBrotliHybridCompressorWriter::new(
            stdout,
            divans::DivansCompressorOptions{
                literal_adaptation:None, // should we override how fast the cdfs converge for literals?
                window_size:Some(22), // log 2 of the window size
                lgblock:None, // should we override how often metablocks are created in brotli
                quality:Some(11), // the quality of brotli commands
                dynamic_context_mixing:Some(2), // if we want to mix together the stride prediction and the context map
                use_brotli:divans::BrotliCompressionSetting::default(), // ignored
                use_context_map:true, // whether we should use the brotli context map in addition to the last 8 bits of each byte as a prior
                force_stride_value: divans::StrideSelection::UseBrotliRec, // if we should use brotli to decide on the stride
            },
            4096, // internal buffer size
        );
        io::copy(&mut io::stdin(), &mut writer).unwrap();
        writer.flush().unwrap();
    }
}

The C api is a standard compression API like the one that zlib provides. Despite being rust code, no allocations are made unless the CAllocator struct is passed in with the custom_malloc field set to NULL. This means that any user of the divans library may provide their own allocation system and all allocations will go through that allocation system. The pointers returned by custom_malloc must be 32-byte aligned.

#include "divans/ffi.h"
// compress to stdout
DivansResult compress(const unsigned char *data, size_t len) {
    unsigned char buf[4096];
    struct CAllocator alloc = {custom_malloc, custom_free, custom_alloc_opaque}; // set all 3 to NULL to use rust allocators
    struct DivansCompressorState *state = divans_new_compressor_with_custom_alloc(alloc);
    divans_set_option(state, DIVANS_OPTION_USE_CONTEXT_MAP, 1);
    divans_set_option(state, DIVANS_OPTION_DYNAMIC_CONTEXT_MIXING, 2);
    divans_set_option(state, DIVANS_OPTION_QUALITY, 11);
    while (len) {
        size_t read_offset = 0;
        size_t buf_offset = 0;
        DivansResult res = divans_encode(state,
                                         data, len, &read_offset,
                                         buf, sizeof(buf), &buf_offset);
        if (res == DIVANS_FAILURE) {
            divans_free_compressor(state);
            return res;
        }
        data += read_offset;
        len -= read_offset;
        fwrite(buf, buf_offset, 1, stdout);
    }
    DivansResult res;
    do {
        size_t buf_offset = 0;
        res = divans_encode_flush(state,
                                  buf, sizeof(buf), &buf_offset);
        if (res == DIVANS_FAILURE) {
            divans_free_compressor(state);
            return res;
        }
        fwrite(buf, buf_offset, 1, stdout);
    } while(res != DIVANS_SUCCESS);
    divans_free_compressor(state);
    return DIVANS_SUCCESS;
}

#include "divans/ffi.h"
//decompress to stdout
DivansResult decompress(const unsigned char *data, size_t len) {
    unsigned char buf[4096];
    struct CAllocator alloc = {custom_malloc, custom_free, custom_alloc_opaque}; // set all 3 to NULL for using rust allocators
    struct DivansDecompressorState *state = divans_new_decompressor_with_custom_alloc(alloc);
    DivansResult res;
    do {
        size_t read_offset = 0;
        size_t buf_offset = 0;
        res = divans_decode(state,
                            data, len, &read_offset,
                            buf, sizeof(buf), &buf_offset);
        if (res == DIVANS_FAILURE || (res == DIVANS_NEEDS_MORE_INPUT && len == 0)) {
            divans_free_decompressor(state);
            return res;
        }
        data += read_offset;
        len -= read_offset;
        fwrite(buf, buf_offset, 1, stdout);
    } while (res != DIVANS_SUCCESS);
    divans_free_decompressor(state);
    return DIVANS_SUCCESS;
}

• a writer::DivansBrotliHybridCompressorWriter instantiates a brotli_ir_gen::BrotliDivansHybridCompressor
• The compressor has both a brotli::BrotliEncoderStateStruct from the brotli crate as well as a codec::DivansCodec<ANSEncoder, EncodeSpecialization>.
• Using brotli_ir_gen::BrotliDivansHybridCompressor::encode, the compressor feeds input data into the brotli::BrotliEncoderStateStruct
  • by calling brotli::BrotliEncoderCompressStream
• brotli::BrotliEncoderCompressStream can trigger a callback into brotli_ir_gen::BrotliDivansHybridCompressor::divans_encode_commands
  • The callback will consist of a slice of brotli::interface::Command items
  • These items are fed into the codec::DivansCodec<ANSEncoder, EncodeSpecialization>::encode_or_decode, which encodes them into divans format.
    • codec::DivansCodec<ANSEncoder, cmd_to_divans::EncoderSpecialization>::encode_or_decode accomplishes this by using the EncoderSpecialization to pull input commands as the source of truth
    • unfortunately brotli can pass as much data as it wishes to the caller, up to the maximum metablock size of 16 megs.
      • this means the caller has to buffer this data in a resizable_buffer::ResizableBuffer
• When all the callbacks have completed, brotli_ir_gen::BrotliDivansHybridCompressor::encode_stream does its best to flush the raw buffer
• Eventually when the user calls brotli_ir_gen::BrotliDivansHybridCompressor::flush a similar procedure is followed but with finish flags set

• a reader::DivansDecompressorReader instantiates a divans_decompressor::DivansDecompressor
• The decompressor is an enum that switches from HeaderParser mode into Decode mode after the 16 byte raw header has been parsed
• divans_decompressor::DivansDecompressor::Decode has a codec::DivansCodec<ANSDecoder, DecoderSpecialization> within.
• Using brotli_ir_gen::DivansDeompressor::decode, the decompressor feeds input data directly into codec::DivansCodec<ANSDecoder, DecoderSpecialization>
  • The codec.encode_or_decode is designed to receive commands as input when encoding, so the divans_to_raw::DecoderSpecialization simply makes placeholder commands for each type of command so that the same codepath can encode and decode commands
• when a final state is reached, a checksum is written and success is returned

codec::DivansCodec has three members, cross_command_state, which tracks the probability models, the state, to track which kind of command is being decoded, and codec_traits, used as a repository of compiler constant values that happen to be set that way during this decode or encode phase based on the header and command data.

The state value is an enumerant that can either carry command-specific information or can mark that the ring buffer must be populated, etc.

• Begin: This state means that the decoder is not in the middle of coding a particular command, so the next step will be to decode what the next command is
• Literal(literal::LiteralState: the coder is in the process of coding raw literals to be injected into the file
• Dict(dict::DictState): the coder is in the process of coding a word that appears in the brotli dictionary
• Copy(copy::CopyState): the coder is in the process of coding a reference to pull data from the ring buffer
• BlockSwitchLiteral(block_type::LiteralBlockTypeState): The coder was instructed to serialize an arbitrary value that will affect how the predictor models future literals
• BlockSwitchCommand(block_type::BlockTypeState): The coder was instructed to serialize an arbitrary value that will affect how the predictor models nothing (TODO)
• BlockSwitchDistance(block_type::BlockTypeState): The coder was instructed to serialize an arbitrary value that will affect how the predictor models distances to copy from and dictionary values.
• PredictionMode(context_map::PredictionModeState): The coder was instructed to serialize out a context map that remaps the BlockSwitchLiteral value plus the last 6 bits into a value in [0, 255] that is used as an index into the array of CDFs to be trained
• PopulateRingBuffer(Command<AllocatedMemoryPrefix<u8, AllocU8>>) When Literal, Dict, or Copy states reach their termination state, those states are moved into the PopulateRingBuffer state.
  • PopulateRingBuffer uses the cmd_to_raw::DivansRecoderState stored in DivanCodec::CrossCommandState to populate the ring buffer
    • If DecoderSpecialization is selected, cmd_to_raw::DivansRecoderState copies the data to the output bytes, returning and requesting NeedBytes until all bytes have been serialized
    • Otherwise the EncoderSpecialization avoids serializing those bytes.
  • After all necessary bytes were serialized and the ring buffer populated, then the last_8_literals are saved to be used as future priors
• WriteChecksum(usize) This state happens if an end command (0xf) is encountered during a decode or a code::DivansCodec::flush happens on encode
  • Currently checksum support is not active, but 8 bytes are simply serialized
• DivansSuccess This state is reached when WriteChecksum is complete on the decoder or when the final command is reached on the encoder
• EncodedShutdownNode | ShutdownCoder | CoderBufferDrain appear only in teh encoder during flush/close after the EOF node type as flushed

Special thanks to Jaroslaw (Jarek) Duda and Fabian Giesen for genius work and their detailed and thoughtful presentation of the ANS algorithm.