d@nny disc@ mc²#[derive(Debug, Snafu)]
#[snafu(display("needed {}, remaining {}", needed, remaining))]
pub struct RemainingError {
pub needed: usize,
pub remaining: usize,
backtrace: Option<Backtrace>,
}
wrote a good amount of code like this for the zstd impl that stopped when i realized zstd was weird and bad
the idea is so basic it's literally:
.tar in two like a magicianthis is my enum for their block type header
#[inline(always)]
fn block_type(&self) -> BlockType {
match (self.0 & 0b110) >> 1 {
0 => BlockType::RawBlock,
1 => BlockType::RLEBlock,
2 => BlockType::CompressedBlock,
3 => BlockType::Reserved,
_ => unreachable!("block type is limited to two bits"),
}
}
two whole bits:
the problem with prefix codes as i understand it now is that they require at least one bit to signal an output symbol (which is typically a byte--i also think this is incredibly important to parameterize). however prefix codes seem to be ideal in every other way (collet mentions performance)
and for extremely biased distributions i.e. ones that would benefit from less than one bit for some highly frequent output symbols, we have (actually a whole field here, but i like) duda's tANS. it's cute. it can be sized to a precise memory region. you follow the path of the blocks wherever they lead
i could cite yann collet on this here but i won't cause he doesn't. and i cited everyone here https://codeberg.org/cosmicexplorer/corporeal/src/branch/main/literature/README.md
the compressed block is where it becomes clear zstd is Just Fucking Deflate Again: https://en.wikipedia.org/wiki/Deflate
BTYPE (next two bits): Block type
00: No compression (sometimes called stored). Any bits up to the next byte boundary are ignored. The rest of the block consists of 16-bit LEN, 16-bit NLEN (one's complement of LEN), and LEN bytes of uncompressed data, i.e. up to 65,535 (216 − 1) bytes. Useful for incompressible data (e.g. high-entropy, random, or already compressed), adding minimal overhead (i.e. ~5 bytes per block).01: A static Huffman compressed block, using a pre-agreed Huffman tree defined in the RFC.10: A dynamic Huffman compressed block, complete with the Huffman table supplied.11: Reserved (error).shameless shit. they even kept the reserved block
more from the DEFLATE page
Most compressible data will end up being encoded using method 10, the dynamic Huffman encoding, which produces an optimized Huffman tree customized for each block of data individually. Instructions to generate the necessary Huffman tree immediately follow the block header. The static Huffman option is used for short messages, where the fixed saving gained by omitting the tree outweighs the percentage compression loss due to using a non-optimal (thus, not technically Huffman) code.
one thing i haven't seen ANYONE fucking talk about is generating compression blocks that understand the boundaries of file contents
some kind of orange shapetarlz archiver does https://www.nongnu.org/lzip/manual/tarlz_manual.html#Amendments-to-pax-format