xet_data/deduplication/

chunking.rs

use std::cmp::min;
use std::io::{Read, Seek, SeekFrom};

use bytes::Bytes;
use more_asserts::{debug_assert_ge, debug_assert_le};

use super::Chunk;
use super::constants::{MAXIMUM_CHUNK_MULTIPLIER, MINIMUM_CHUNK_DIVISOR, TARGET_CHUNK_SIZE};

/// Chunk Generator given an input stream. Do not use directly.
/// Use `chunk_target_default`.
pub struct Chunker {
    // configs
    hash: gearhash::Hasher<'static>,
    minimum_chunk: usize,
    maximum_chunk: usize,
    mask: u64,

    // generator state
    chunkbuf: Vec<u8>,
}

impl Default for Chunker {
    fn default() -> Self {
        Self::new(*TARGET_CHUNK_SIZE)
    }
}

impl Chunker {
    pub fn new(target_chunk_size: usize) -> Self {
        assert_eq!(target_chunk_size.count_ones(), 1);

        // Some of the logic only works if the target_chunk_size is greater than the
        // window size of the hash.
        assert!(target_chunk_size > 64);

        // note the strict lesser than. Combined with count_ones() == 1,
        // this limits to 2^31
        assert!(target_chunk_size < u32::MAX as usize);

        let mask = (target_chunk_size - 1) as u64;

        // we will like to shift the mask left by a bunch since the right
        // bits of the gear hash are affected by only a small number of bytes
        // really. we just shift it all the way left.
        let mask = mask << mask.leading_zeros();
        let minimum_chunk = target_chunk_size / *MINIMUM_CHUNK_DIVISOR;
        let maximum_chunk = target_chunk_size * *MAXIMUM_CHUNK_MULTIPLIER;

        assert!(maximum_chunk > minimum_chunk);

        let hash = gearhash::Hasher::default();

        Chunker {
            hash,
            minimum_chunk,
            maximum_chunk,
            mask,
            // generator state init
            chunkbuf: Vec::with_capacity(maximum_chunk),
        }
    }

    /// Create a chunker with custom min chunk sizes.
    /// Only used by the partitioner which has special requirements.
    fn new_with_min(target_chunk_size: usize, min_chunk_size: usize) -> Self {
        let mut chunker = Self::new(target_chunk_size);
        chunker.minimum_chunk = min_chunk_size;
        chunker
    }

    /// Looks for the next chunk boundary in the data.  Assumes that whatever is in the current
    /// state has been prepended to the current data.  If a boundary cannot be found based on the
    /// current amount of data, then None is returned.
    #[inline]
    pub fn next_boundary(&mut self, data: &[u8]) -> Option<usize> {
        const HASH_WINDOW_SIZE: usize = 64;
        let n_bytes = data.len();

        if n_bytes == 0 {
            return None;
        }

        let previous_len = self.chunkbuf.len();
        let mut cur_index = 0;
        let mut create_chunk = false;

        // skip the minimum chunk size
        // and noting that the hash has a window size of 64
        // so we should be careful to skip only minimum_chunk - 64 - 1
        if previous_len + HASH_WINDOW_SIZE < self.minimum_chunk {
            let skip = min(self.minimum_chunk - previous_len - HASH_WINDOW_SIZE - 1, n_bytes);
            cur_index += skip;
        }

        // If we have a lot of data, don't read all the way to the end when we'll stop reading
        // at the maximum chunk boundary.
        let read_end = n_bytes.min(cur_index + self.maximum_chunk - previous_len);

        loop {
            if let Some(next_match) = self.hash.next_match(&data[cur_index..read_end], self.mask) {
                cur_index += next_match;

                // If we trigger a boundary before the end, create a chunk.

                // We must enforce that the next boundary is actually past the minimum chunk size.
                // Because of how the rolling hash is computed, bytes before HASH_WINDOW_SIZE don't affect the hash,
                // so with the above skip we depend on it running for at least HASH_WINDOW_SIZE bytes before triggering
                // a boundary.   However, in rare occurrences, there can be a boundary triggered before HASH_WINDOW_SIZE
                // bytes have been processed, which means the boundary is triggered based on the previous state of the
                // hasher rather than on the current chunk content.  Thus we ensure this can't happen by ensuring that
                // we have processed at least HASH_WINDOW_SIZE bytes.
                if cur_index + previous_len < self.minimum_chunk {
                    continue;
                }

                create_chunk = true;
            } else {
                cur_index = read_end;
            }

            break;
        }

        // if we hit maximum chunk we must create a chunk
        if cur_index + previous_len >= self.maximum_chunk {
            cur_index = self.maximum_chunk - previous_len;
            create_chunk = true;
        }

        if create_chunk {
            self.hash.set_hash(0); // Reset for the next time.
            debug_assert_ge!(cur_index + previous_len, self.minimum_chunk);
            debug_assert_le!(cur_index + previous_len, self.maximum_chunk);
            Some(cur_index)
        } else {
            None
        }
    }

    fn reset_state(&mut self) {
        // Strictly speaking, this is unnecessary, as we should always hash 64 bytes out making the previous state
        // of the hasher irrelevant.  However, this explicitly declares we're resetting things to the
        // initial state.
        self.hash.set_hash(0);
        debug_assert!(self.chunkbuf.is_empty());
    }

    /// Process more data; this is a continuation of any data from before when calls were
    ///
    /// Returns the next chunk, if available, and the amount of data that was digested.
    ///
    /// If is_final is true, then it is assumed that no more data after this block will come,
    /// and any data currently present and at the end will be put into a final chunk.
    pub fn next(&mut self, data: &[u8], is_final: bool) -> (Option<Chunk>, usize) {
        let (chunk_data, consume): (Bytes, usize) = {
            if let Some(next_boundary) = self.next_boundary(data) {
                if self.chunkbuf.is_empty() {
                    (Bytes::copy_from_slice(&data[..next_boundary]), next_boundary)
                } else {
                    self.chunkbuf.extend_from_slice(&data[..next_boundary]);
                    (std::mem::take(&mut self.chunkbuf).into(), next_boundary)
                }
            } else if is_final {
                // Put the rest of the data in the chunkbuf.
                let r = if self.chunkbuf.is_empty() {
                    (Bytes::copy_from_slice(data), data.len())
                } else {
                    self.chunkbuf.extend_from_slice(data);
                    (std::mem::take(&mut self.chunkbuf).into(), data.len())
                };

                if is_final {
                    self.reset_state();
                }

                r
            } else {
                self.chunkbuf.extend_from_slice(data);
                return (None, data.len());
            }
        };

        // Special case this specific case.
        if chunk_data.is_empty() {
            return (None, 0);
        }

        (Some(Chunk::new(chunk_data)), consume)
    }

    /// Keeps chunking until no more chunks can be reliably produced, returning a
    /// vector of the resulting chunks.  
    pub fn next_block(&mut self, data: &[u8], is_final: bool) -> Vec<Chunk> {
        let mut ret = Vec::new();

        let mut pos = 0;
        loop {
            debug_assert!(pos <= data.len());
            if pos == data.len() {
                if is_final {
                    self.reset_state();
                }

                return ret;
            }

            let (maybe_chunk, bytes_consumed) = self.next(&data[pos..], is_final);

            if let Some(chunk) = maybe_chunk {
                ret.push(chunk);
            }

            pos += bytes_consumed;
        }
    }

    /// Keeps chunking until no more chunks can be reliably produced, returning a
    /// vector of the resulting chunks.
    ///
    /// The data is inserted here as a Bytes object, which means that no copying of the data
    /// is performed except at the boundaries.  The resulting chunks then end up each holding
    /// a reference to the original data object, which will not be deallocated until all the
    /// original bytes are gone.
    pub fn next_block_bytes(&mut self, data: &Bytes, is_final: bool) -> Vec<Chunk> {
        let mut ret = Vec::new();

        let mut pos = 0;

        // In this case, we have to perform a single cut using the old method,
        // which would copy the data.
        if !self.chunkbuf.is_empty() {
            let (maybe_chunk, skip_idx) = self.next(data, is_final);
            if let Some(chunk) = maybe_chunk {
                ret.push(chunk);
            }
            pos = skip_idx;
        }

        while pos < data.len() {
            let maybe_next_boundary = self.next_boundary(&data[pos..]);

            if let Some(chunk_size) = maybe_next_boundary {
                let next_pos = pos + chunk_size;
                ret.push(Chunk::new(data.slice(pos..next_pos)));
                pos = next_pos;
            } else {
                // No more chunks in this block.
                if is_final {
                    ret.push(Chunk::new(data.slice(pos..)));
                } else {
                    self.chunkbuf.extend_from_slice(&data[pos..]);
                }
                break;
            }
        }

        if is_final {
            self.reset_state();
        }

        ret
    }

    // Finishes, returning the final chunk if one exists, and resets the hasher to
    pub fn finish(&mut self) -> Option<Chunk> {
        self.next(&[], true).0
    }
}

/// Find valid partition points in a file where we can
/// chunk in parallel. Returns the start points of each partition
/// (i.e. file offset 0 is always the first entry, and `file_size`
/// is never in the result).
/// Note that reader position is modified and not restored.
///
/// partition_scan_bytes is the number of bytes to scan at each
/// proposed partition boundary in search of a valid chunk.
///
/// Partition alignment is guaranteed by the hash warmup fix: the
/// chunker feeds `min_chunk - 64 - 1` bytes before scanning for
/// boundaries, ensuring the gear hash window is fully warmed (purely
/// data-dependent) at all accepted trigger positions. This function
/// additionally verifies the absence of hidden triggers by re-chunking
/// with `min_chunk = 0`. See `parallel chunking.lyx` for the proof.
///
/// For finding stable chunk boundaries from existing chunk boundaries (without
/// data access), see [`next_stable_chunk_boundary`].
pub fn find_partitions<R: Read + Seek>(
    reader: &mut R,
    file_size: usize,
    target_chunk_size: usize,
    min_partition_size: usize,
    partition_scan_bytes: usize,
) -> std::io::Result<Vec<usize>> {
    assert!(min_partition_size > 0);
    let mut partitions: Vec<usize> = Vec::new();
    partitions.push(0);
    // minimum chunk must be at least the hash window size.
    // the way the chunker works, the minimum may be up to
    // target_min_chunk_size - 64
    let minimum_chunk = target_chunk_size / *MINIMUM_CHUNK_DIVISOR;
    let maximum_chunk = target_chunk_size * *MAXIMUM_CHUNK_MULTIPLIER;

    assert!(minimum_chunk > 64);

    if maximum_chunk >= min_partition_size {
        return Ok(partitions);
    }
    let mut buf = vec![0u8; partition_scan_bytes];
    let mut curpos: usize = 0;
    // we jump curpos forward by min_partition_size
    // and read *PARALLEL_CHUNKING_PARTITION_SCAN_BYTES bytes
    // and try to find a partition boundary condition.
    //
    // We should also make sure There should also be at least
    // min_partition_size bytes remaining at curpos so that
    // we do not make a teeny tiny partition.
    while curpos < file_size {
        curpos += min_partition_size;
        // there are not enough bytes to make a full partition
        // or not enough bytes to scan for a partition
        if curpos + min_partition_size >= file_size || curpos + partition_scan_bytes >= file_size {
            break;
        }
        // read and chunk the scan bytes
        reader.seek(SeekFrom::Start(curpos as u64))?;
        reader.read_exact(&mut buf)?;
        let mut chunker = Chunker::new_with_min(target_chunk_size, 0);
        // TODO: there is a definite optimization here
        // as we really only need the chunk lengths and not the data
        let chunks = chunker.next_block(&buf, false);
        if chunks.is_empty() {
            continue;
        }
        // skip the first chunk
        let mut offset = chunks[0].data.len();
        offset += chunks[1].data.len();
        for i in 2..chunks.len() {
            let cprev = chunks[i - 1].data.len();
            let c = chunks[i].data.len();
            offset += chunks[i].data.len();
            if cprev > minimum_chunk
                && cprev < maximum_chunk - minimum_chunk
                && c > minimum_chunk
                && c < maximum_chunk - minimum_chunk
            {
                // we have a valid partition at this position
                partitions.push(curpos + offset);
                break;
            }
        }
    }
    Ok(partitions)
}

// Re-exported from xet_core_structures where the canonical implementation lives,
// so that downstream users of xet_data::deduplication::next_stable_chunk_boundary
// continue to work without a source change.
pub use xet_core_structures::xorb_object::constants::next_stable_chunk_boundary;

#[cfg(test)]
mod tests {
    use std::collections::HashSet;
    use std::io::Cursor;

    use rand::rngs::StdRng;
    use rand::{RngExt, SeedableRng};

    use super::*;

    /// A helper to create random test data using a specified `seed` and `len`.
    /// Using a fixed seed ensures tests are reproducible.
    fn make_test_data(seed: u64, len: usize) -> Vec<u8> {
        let mut rng = StdRng::seed_from_u64(seed);
        let mut data = vec![0; len];
        rng.fill(&mut data[..]);
        data
    }

    fn check_chunks_equal(chunks: &[Chunk], data: &[u8]) {
        // Validate all the chunks are exact.
        let mut new_vec = Vec::with_capacity(10000);
        for c in chunks.iter() {
            new_vec.extend_from_slice(&c.data[..]);
        }

        assert!(new_vec == data);
    }

    // A chunker that wraps two versions of the Chunker class,
    // exposing next_block but then internally testing next_bytes_block and next_block and
    // verifying the output is identical.
    #[derive(Default)]
    struct ChunkerTestWrapper {
        chunker_chunks: Chunker,
        chunker_bytes: Chunker,
    }

    impl ChunkerTestWrapper {
        fn new(target_chunk_size: usize) -> Self {
            ChunkerTestWrapper {
                chunker_chunks: Chunker::new(target_chunk_size),
                chunker_bytes: Chunker::new(target_chunk_size),
            }
        }

        fn next_block(&mut self, data: &[u8], is_final: bool) -> Vec<Chunk> {
            let chunks = self.chunker_chunks.next_block(data, is_final);
            let bytes_chunks = self.chunker_bytes.next_block_bytes(&Bytes::copy_from_slice(data), is_final);

            // Check that the two match.
            assert_eq!(chunks.len(), bytes_chunks.len());
            for (c1, c2) in chunks.iter().zip(bytes_chunks.iter()) {
                assert_eq!(c1.data, c2.data);
            }

            chunks
        }

        fn next_chunk(&mut self, data: &[u8], is_final: bool) -> (Option<Chunk>, usize) {
            let (chunk, consumed) = self.chunker_chunks.next(data, is_final);
            let (bytes_chunk, bytes_consumed) = self.chunker_bytes.next(&Bytes::copy_from_slice(data), is_final);

            // Check that the two match.
            if let Some(c) = &chunk {
                assert_eq!(c.data, bytes_chunk.unwrap().data);
            } else {
                assert!(bytes_chunk.is_none());
            }

            (chunk, consumed.max(bytes_consumed))
        }
    }

    #[test]
    fn test_empty_data_no_chunk_until_final() {
        let mut chunker = ChunkerTestWrapper::new(128);

        // Passing empty slice without final => no chunk
        let (chunk, consumed) = chunker.next_chunk(&[], false);
        assert!(chunk.is_none());
        assert_eq!(consumed, 0);

        // Passing empty slice again with is_final = true => no leftover data, so no chunk
        let (chunk, consumed) = chunker.next_chunk(&[], true);
        assert!(chunk.is_none());
        assert_eq!(consumed, 0);
    }

    #[test]
    fn test_data_smaller_than_minimum_no_boundary() {
        let mut chunker = ChunkerTestWrapper::new(128);

        // Create a small random data buffer. For example, length=3.
        let data = make_test_data(0, 63);

        // We expect no chunk until we finalize, because there's not enough data
        // to trigger a boundary, nor to reach the maximum chunk size.
        let (chunk, consumed) = chunker.next_chunk(&data, false);
        assert!(chunk.is_none());
        assert_eq!(consumed, data.len());

        // Now finalize: we expect a chunk with the leftover data
        let (chunk, consumed) = chunker.next_chunk(&[], true);
        assert!(chunk.is_some());
        assert_eq!(consumed, 0);

        let chunk = chunk.unwrap();
        assert_eq!(chunk.data.len(), 63);
        assert_eq!(&chunk.data[..], &data[..], "Chunk should contain exactly what was passed in");
    }

    #[test]
    fn test_multiple_chunks_produced() {
        let mut chunker = ChunkerTestWrapper::new(128);

        // Produce 100 bytes of random data
        let data = make_test_data(42, 10000);

        // Pass everything at once, final = true
        let chunks = chunker.next_block(&data, true);
        assert!(!chunks.is_empty());

        check_chunks_equal(&chunks, &data);
    }

    #[test]
    fn test_repeated_calls_partial_consumption() {
        // We'll feed in two pieces of data to ensure partial consumption

        let data = make_test_data(42, 10000);

        let mut chunks_1 = Vec::new();

        let mut pos = 0;
        let mut chunker = ChunkerTestWrapper::new(128);

        while pos < data.len() {
            for i in 0..16 {
                let next_pos = (pos + i).min(data.len());
                chunks_1.append(&mut chunker.next_block(&data[pos..next_pos], next_pos == data.len()));
                pos = next_pos;
            }
        }

        check_chunks_equal(&chunks_1, &data);

        // Now, rechunk with all at once and make sure it's equal.
        let chunks_2 = ChunkerTestWrapper::new(128).next_block(&data, true);

        assert_eq!(chunks_1, chunks_2);
    }

    #[test]
    fn test_exact_maximum_chunk() {
        // If the data hits the maximum chunk size exactly, we should force a boundary.
        // For target_chunk_size = 128, if MAXIMUM_CHUNK_MULTIPLIER = 2, then max = 256.
        // Adjust if your constants differ.
        let mut chunker = ChunkerTestWrapper::new(512);

        // Use constant data
        let data = vec![0; 8 * *MAXIMUM_CHUNK_MULTIPLIER * 512];

        let chunks = chunker.next_block(&data, true);

        assert_eq!(chunks.len(), 8);

        for c in chunks.iter() {
            assert_eq!(c.data.len(), *MAXIMUM_CHUNK_MULTIPLIER * 512);
        }
    }

    #[test]
    fn test_partition() {
        for _i in 1..5 {
            let data = make_test_data(42, 1000000);
            let mut chunker = Chunker::new(1024);
            let chunks = chunker.next_block(&data, true);
            let mut chunk_offsets = HashSet::new();
            let mut offset = 0;
            eprintln!("{:?}", chunker.minimum_chunk);
            for i in 0..chunks.len() {
                chunk_offsets.insert(offset);
                offset += chunks[i].data.len();
            }

            let partitions =
                find_partitions(&mut Cursor::new(&mut data.as_slice()), data.len(), 1024, 100000, 10000).unwrap();
            assert!(partitions.len() > 1);
            for i in 0..partitions.len() {
                assert!(chunk_offsets.contains(&partitions[i]));
            }
        }
    }

    /// Simple SplitMix64-based deterministic random number generator.
    /// Portable to C, Python, etc. (see https://prng.di.unimi.it/splitmix64.c)
    fn splitmix64_next(state: &mut u64) -> u64 {
        *state = state.wrapping_add(0x9E3779B97F4A7C15);
        let mut z = *state;
        z = (z ^ (z >> 30)).wrapping_mul(0xBF58476D1CE4E5B9);
        z = (z ^ (z >> 27)).wrapping_mul(0x94D049BB133111EB);
        z ^ (z >> 31)
    }

    fn create_random_data(n: usize, seed: u64) -> Vec<u8> {
        // This test will actually need to be run in different environments, so to generate
        // the table below, create random data using a simple SplitMix rng that can be ported here
        // as is without depending on other po
        let mut ret = Vec::with_capacity(n + 7);

        let mut state = seed;

        while ret.len() < n {
            let next_u64 = splitmix64_next(&mut state);
            ret.extend_from_slice(&next_u64.to_le_bytes());
        }

        // Has extra bits on there since we're adding in blocks of 8.
        ret.resize(n, 0);

        ret
    }

    fn get_chunk_boundaries(chunks: &[Chunk]) -> Vec<usize> {
        chunks
            .iter()
            .scan(0, |state, chunk| {
                *state += chunk.data.len();
                Some(*state)
            })
            .collect()
    }

    #[test]
    fn test_chunk_boundaries() {
        let data = create_random_data(256000, 1);

        // Now, run the chunks through the default chunker.
        let chunks = ChunkerTestWrapper::default().next_block(&data, true);

        // Get the boundaries indices as determined by the size of the chunks above.
        let ref_chunk_boundaries: Vec<usize> = get_chunk_boundaries(&chunks);

        // Test that it's correct across different chunk varieties.
        for add_size in [1, 37, 255] {
            let mut chunker = Chunker::default();

            // Add repeatedly in blocks of add_size, appending to alt_chunks
            let mut alt_chunks = Vec::with_capacity(chunks.len());

            let mut pos = 0;
            while pos < data.len() {
                let next_pos = (pos + add_size).min(data.len());
                let next_chunk = chunker.next_block(&data[pos..next_pos], next_pos == data.len());
                alt_chunks.extend(next_chunk);
                pos = next_pos;
            }

            let alt_boundaries = get_chunk_boundaries(&alt_chunks);

            assert_eq!(alt_boundaries, ref_chunk_boundaries);
        }
    }

    #[test]
    fn test_correctness_1mb_random_data() {
        // Test this data.
        let data = create_random_data(1000000, 0);

        // Uncomment these to create the lines below:
        // eprintln!("(data[0], {});", data[0] as usize);
        // eprintln!("(data[127], {});", data[127] as usize);
        // eprintln!("(data[111111], {});", data[111111] as usize);

        assert_eq!(data[0], 175);
        assert_eq!(data[127], 132);
        assert_eq!(data[111111], 118);

        // Now, run the chunks through the default chunker.
        let chunks = ChunkerTestWrapper::default().next_block(&data, true);

        // Get the boundaries indices as determined by the size of the chunks above.
        let chunk_boundaries: Vec<usize> = get_chunk_boundaries(&chunks);

        // Uncomment this to create the line below.
        // eprintln!("assert_eq!(chunk_boundaries, vec!{chunk_boundaries:?})");
        assert_eq!(
            chunk_boundaries,
            vec![
                84493, 134421, 144853, 243318, 271793, 336457, 467529, 494581, 582000, 596735, 616815, 653164, 678202,
                724510, 815591, 827760, 958832, 991092, 1000000
            ]
        );
    }

    #[test]
    fn test_correctness_1mb_const_data() {
        // Test this data.
        let data = vec![59u8; 1000000];

        // Now, run the chunks through the default chunker.
        let chunks = ChunkerTestWrapper::default().next_block(&data, true);

        // Get the boundaries indices as determined by the size of the chunks above.
        let chunk_boundaries: Vec<usize> = get_chunk_boundaries(&chunks);

        // Uncomment this to create the line below.
        // eprintln!("assert_eq!(chunk_boundaries, vec!{chunk_boundaries:?})");
        assert_eq!(chunk_boundaries, vec![131072, 262144, 393216, 524288, 655360, 786432, 917504, 1000000])
    }

    fn get_triggering_base_data(n: usize, padding: usize) -> Vec<u8> {
        // This pattern is known to trigger the boundary detection in the chunker, so repeat it to test the
        // correctness of the minimum chunk size processing.
        let mut data = vec![
            154, 52, 42, 34, 159, 75, 126, 224, 70, 236, 12, 196, 79, 236, 178, 124, 127, 50, 99, 178, 44, 176, 174,
            126, 250, 235, 205, 174, 252, 122, 35, 10, 20, 101, 214, 69, 193, 8, 115, 105, 158, 228, 120, 111, 136,
            162, 198, 251, 211, 183, 253, 252, 164, 147, 63, 16, 186, 162, 117, 23, 170, 36, 205, 187, 174, 76, 210,
            174, 211, 175, 12, 173, 145, 59, 2, 70, 222, 181, 159, 227, 182, 156, 189, 51, 226, 106, 24, 50, 183, 157,
            140, 10, 8, 23, 212, 70, 10, 234, 23, 33, 219, 254, 39, 236, 70, 49, 191, 116, 9, 115, 15, 101, 26, 159,
            165, 220, 15, 170, 56, 125, 92, 163, 94, 235, 38, 40, 49, 81,
        ];

        // Add padding so we can comprehensively test the nuances of boundaries.
        data.resize(data.len() + padding, 0u8);

        // Repeat the above pattern until we've filled out n bytes.
        while data.len() < n {
            let n_take = (n - data.len()).min(data.len());
            data.extend_from_within(0..n_take);
        }

        data
    }

    #[test]
    fn test_correctness_100kb_hitting_data() {
        // To ensure we've checked all the nuances of dealing with minimum chunk boundaries,
        // and with the correct chunks as well, run through all the different options with the padding,
        // checking each one.  With this, then, we have a pattern that hits once per pattern with varying
        // bits between the widths.

        let mut data_sample_at_11111 = [0u8; 128];
        let mut ref_cb = vec![Vec::new(); 128];

        data_sample_at_11111[0] = 236;
        ref_cb[0] = vec![8256, 16448, 24640, 32832, 41024, 49216, 57408, 65536];
        data_sample_at_11111[1] = 50;
        ref_cb[1] = vec![8320, 16576, 24832, 33088, 41344, 49600, 57856, 65536];
        data_sample_at_11111[2] = 36;
        ref_cb[2] = vec![8254, 16574, 24894, 33214, 41534, 49854, 58174, 65536];
        data_sample_at_11111[3] = 116;
        ref_cb[3] = vec![8317, 16570, 24823, 33076, 41329, 49582, 57835, 65536];
        data_sample_at_11111[4] = 126;
        ref_cb[4] = vec![8248, 16564, 24880, 33196, 41512, 49828, 58144, 65536];
        data_sample_at_11111[5] = 145;
        ref_cb[5] = vec![8310, 16556, 24802, 33048, 41294, 49540, 57786, 65536];
        data_sample_at_11111[6] = 235;
        ref_cb[6] = vec![8238, 16546, 24854, 33162, 41470, 49778, 58086, 65536];
        data_sample_at_11111[7] = 228;
        ref_cb[7] = vec![8299, 16534, 24769, 33004, 41239, 49474, 57709, 65536];
        data_sample_at_11111[8] = 70;
        ref_cb[8] = vec![8224, 16520, 24816, 33112, 41408, 49704, 58000, 65536];
        data_sample_at_11111[9] = 178;
        ref_cb[9] = vec![8284, 16504, 24724, 32944, 41164, 49384, 57604, 65536];
        data_sample_at_11111[10] = 173;
        ref_cb[10] = vec![8206, 16486, 24766, 33046, 41326, 49606, 57886, 65536];
        data_sample_at_11111[11] = 0;
        ref_cb[11] = vec![8265, 16466, 24667, 32868, 41069, 49270, 57471, 65536];
        data_sample_at_11111[12] = 252;
        ref_cb[12] = vec![8324, 16584, 24844, 33104, 41364, 49624, 57884, 65536];
        data_sample_at_11111[13] = 159;
        ref_cb[13] = vec![8242, 16561, 24880, 33199, 41518, 49837, 58156, 65536];
        data_sample_at_11111[14] = 69;
        ref_cb[14] = vec![8300, 16536, 24772, 33008, 41244, 49480, 57716, 65536];
        data_sample_at_11111[15] = 219;
        ref_cb[15] = vec![8215, 16509, 24803, 33097, 41391, 49685, 57979, 65536];
        data_sample_at_11111[16] = 126;
        ref_cb[16] = vec![8272, 16480, 24688, 32896, 41104, 49312, 57520, 65536];
        data_sample_at_11111[17] = 10;
        ref_cb[17] = vec![8329, 16594, 24859, 33124, 41389, 49654, 57919, 65536];
        data_sample_at_11111[18] = 124;
        ref_cb[18] = vec![8240, 16562, 24884, 33206, 41528, 49850, 58172, 65536];
        data_sample_at_11111[19] = 24;
        ref_cb[19] = vec![8296, 16528, 24760, 32992, 41224, 49456, 57688, 65536];
        data_sample_at_11111[20] = 196;
        ref_cb[20] = vec![8204, 16492, 24780, 33068, 41356, 49644, 57932, 65536];
        data_sample_at_11111[21] = 106;
        ref_cb[21] = vec![8259, 16454, 24649, 32844, 41039, 49234, 57429, 65536];
        data_sample_at_11111[22] = 196;
        ref_cb[22] = vec![8314, 16564, 24814, 33064, 41314, 49564, 57814, 65536];
        data_sample_at_11111[23] = 183;
        ref_cb[23] = vec![8218, 16523, 24828, 33133, 41438, 49743, 58048, 65536];
        data_sample_at_11111[24] = 124;
        ref_cb[24] = vec![8272, 16480, 24688, 32896, 41104, 49312, 57520, 65536];
        data_sample_at_11111[25] = 70;
        ref_cb[25] = vec![8326, 16588, 24850, 33112, 41374, 49636, 57898, 65536];
        data_sample_at_11111[26] = 126;
        ref_cb[26] = vec![8226, 16542, 24858, 33174, 41490, 49806, 58122, 65536];
        data_sample_at_11111[27] = 191;
        ref_cb[27] = vec![8279, 16494, 24709, 32924, 41139, 49354, 57569, 65536];
        data_sample_at_11111[28] = 69;
        ref_cb[28] = vec![8332, 16600, 24868, 33136, 41404, 49672, 57940, 65536];
        data_sample_at_11111[29] = 163;
        ref_cb[29] = vec![8228, 16549, 24870, 33191, 41512, 49833, 58154, 65536];
        data_sample_at_11111[30] = 252;
        ref_cb[30] = vec![8280, 16496, 24712, 32928, 41144, 49360, 57576, 65536];
        data_sample_at_11111[31] = 0;
        ref_cb[31] = vec![8332, 16600, 24868, 33136, 41404, 49672, 57940, 65536];
        data_sample_at_11111[32] = 173;
        ref_cb[32] = vec![8224, 16544, 24864, 33184, 41504, 49824, 58144, 65536];
        data_sample_at_11111[33] = 42;
        ref_cb[33] = vec![8275, 16486, 24697, 32908, 41119, 49330, 57541, 65536];
        data_sample_at_11111[34] = 70;
        ref_cb[34] = vec![8326, 16588, 24850, 33112, 41374, 49636, 57898, 65536];
        data_sample_at_11111[35] = 174;
        ref_cb[35] = vec![8214, 16527, 24840, 33153, 41466, 49779, 58092, 65536];
        data_sample_at_11111[36] = 235;
        ref_cb[36] = vec![8264, 16464, 24664, 32864, 41064, 49264, 57464, 65536];
        data_sample_at_11111[37] = 186;
        ref_cb[37] = vec![8314, 16564, 24814, 33064, 41314, 49564, 57814, 65536];
        data_sample_at_11111[38] = 0;
        ref_cb[38] = vec![8198, 16498, 24798, 33098, 41398, 49698, 57998, 65536];
        data_sample_at_11111[39] = 157;
        ref_cb[39] = vec![8247, 16597, 24947, 33297, 41647, 49997, 58347, 65536];
        data_sample_at_11111[40] = 126;
        ref_cb[40] = vec![8296, 16528, 24760, 32992, 41224, 49456, 57688, 65536];
        data_sample_at_11111[41] = 49;
        ref_cb[41] = vec![8345, 16626, 24907, 33188, 41469, 49750, 58031, 65536];
        data_sample_at_11111[42] = 36;
        ref_cb[42] = vec![8224, 16554, 24884, 33214, 41544, 49874, 58204, 65536];
        data_sample_at_11111[43] = 0;
        ref_cb[43] = vec![8272, 16480, 24688, 32896, 41104, 49312, 57520, 65536];
        data_sample_at_11111[44] = 236;
        ref_cb[44] = vec![8320, 16576, 24832, 33088, 41344, 49600, 57856, 65536];
        data_sample_at_11111[45] = 105;
        ref_cb[45] = vec![8195, 16499, 24803, 33107, 41411, 49715, 58019, 65536];
        data_sample_at_11111[46] = 0;
        ref_cb[46] = vec![8242, 16594, 24946, 33298, 41650, 50002, 58354, 65536];
        data_sample_at_11111[47] = 24;
        ref_cb[47] = vec![8289, 16514, 24739, 32964, 41189, 49414, 57639, 65536];
        data_sample_at_11111[48] = 126;
        ref_cb[48] = vec![8336, 16608, 24880, 33152, 41424, 49696, 57968, 65536];
        data_sample_at_11111[49] = 0;
        ref_cb[49] = vec![8206, 16525, 24844, 33163, 41482, 49801, 58120, 65536];
        data_sample_at_11111[50] = 70;
        ref_cb[50] = vec![8252, 16618, 24984, 33350, 41716, 50082, 58448, 65536];
        data_sample_at_11111[51] = 236;
        ref_cb[51] = vec![8298, 16532, 24766, 33000, 41234, 49468, 57702, 65536];
        data_sample_at_11111[52] = 0;
        ref_cb[52] = vec![8344, 16624, 24904, 33184, 41464, 49744, 58024, 65536];
        data_sample_at_11111[53] = 12;
        ref_cb[53] = vec![8209, 16535, 24861, 33187, 41513, 49839, 58165, 65536];
        data_sample_at_11111[54] = 236;
        ref_cb[54] = vec![8254, 16626, 24998, 33370, 41742, 50114, 58486, 65536];
        data_sample_at_11111[55] = 0;
        ref_cb[55] = vec![8299, 16534, 24769, 33004, 41239, 49474, 57709, 65536];
        data_sample_at_11111[56] = 173;
        ref_cb[56] = vec![8344, 16624, 24904, 33184, 41464, 49744, 58024, 65536];
        data_sample_at_11111[57] = 196;
        ref_cb[57] = vec![8204, 16529, 24854, 33179, 41504, 49829, 58154, 65536];
        data_sample_at_11111[58] = 0;
        ref_cb[58] = vec![8248, 16618, 24988, 33358, 41728, 50098, 58468, 65536];
        data_sample_at_11111[59] = 159;
        ref_cb[59] = vec![8292, 16520, 24748, 32976, 41204, 49432, 57660, 65536];
        data_sample_at_11111[60] = 178;
        ref_cb[60] = vec![8336, 16608, 24880, 33152, 41424, 49696, 57968, 65536];
        data_sample_at_11111[61] = 0;
        ref_cb[61] = vec![8380, 16696, 25012, 33328, 41644, 49960, 58276, 65536];
        data_sample_at_11111[62] = 10;
        ref_cb[62] = vec![8234, 16594, 24954, 33314, 41674, 50034, 58394, 65536];
        data_sample_at_11111[63] = 101;
        ref_cb[63] = vec![8277, 16490, 24703, 32916, 41129, 49342, 57555, 65536];
        data_sample_at_11111[64] = 0;
        ref_cb[64] = vec![8320, 16576, 24832, 33088, 41344, 49600, 57856, 65536];
        data_sample_at_11111[65] = 15;
        ref_cb[65] = vec![8363, 16662, 24961, 33260, 41559, 49858, 58157, 65536];
        data_sample_at_11111[66] = 147;
        ref_cb[66] = vec![8212, 16554, 24896, 33238, 41580, 49922, 58264, 65536];
        data_sample_at_11111[67] = 0;
        ref_cb[67] = vec![8254, 16639, 25024, 33409, 41794, 50179, 58564, 65536];
        data_sample_at_11111[68] = 0;
        ref_cb[68] = vec![8296, 16528, 24760, 32992, 41224, 49456, 57688, 65536];
        data_sample_at_11111[69] = 227;
        ref_cb[69] = vec![8338, 16612, 24886, 33160, 41434, 49708, 57982, 65536];
        data_sample_at_11111[70] = 126;
        ref_cb[70] = vec![8380, 16696, 25012, 33328, 41644, 49960, 58276, 65536];
        data_sample_at_11111[71] = 0;
        ref_cb[71] = vec![8223, 16581, 24939, 33297, 41655, 50013, 58371, 65536];
        data_sample_at_11111[72] = 101;
        ref_cb[72] = vec![8264, 16464, 24664, 32864, 41064, 49264, 57464, 65536];
        data_sample_at_11111[73] = 186;
        ref_cb[73] = vec![8305, 16546, 24787, 33028, 41269, 49510, 57751, 65536];
        data_sample_at_11111[74] = 52;
        ref_cb[74] = vec![8346, 16628, 24910, 33192, 41474, 49756, 58038, 65536];
        data_sample_at_11111[75] = 0;
        ref_cb[75] = vec![8387, 16710, 25033, 33356, 41679, 50002, 58325, 65536];
        data_sample_at_11111[76] = 70;
        ref_cb[76] = vec![8224, 16588, 24952, 33316, 41680, 50044, 58408, 65536];
        data_sample_at_11111[77] = 228;
        ref_cb[77] = vec![8264, 16464, 24664, 32864, 41064, 49264, 57464, 65536];
        data_sample_at_11111[78] = 0;
        ref_cb[78] = vec![8304, 16544, 24784, 33024, 41264, 49504, 57744, 65536];
        data_sample_at_11111[79] = 0;
        ref_cb[79] = vec![8344, 16624, 24904, 33184, 41464, 49744, 58024, 65536];
        data_sample_at_11111[80] = 50;
        ref_cb[80] = vec![8384, 16704, 25024, 33344, 41664, 49984, 58304, 65536];
        data_sample_at_11111[81] = 214;
        ref_cb[81] = vec![8215, 16575, 24935, 33295, 41655, 50015, 58375, 65536];
        data_sample_at_11111[82] = 0;
        ref_cb[82] = vec![8254, 16654, 25054, 33454, 41854, 50254, 58654, 65536];
        data_sample_at_11111[83] = 0;
        ref_cb[83] = vec![8293, 16522, 24751, 32980, 41209, 49438, 57667, 65536];
        data_sample_at_11111[84] = 50;
        ref_cb[84] = vec![8332, 16600, 24868, 33136, 41404, 49672, 57940, 65536];
        data_sample_at_11111[85] = 69;
        ref_cb[85] = vec![8371, 16678, 24985, 33292, 41599, 49906, 58213, 65536];
        data_sample_at_11111[86] = 0;
        ref_cb[86] = vec![8196, 16542, 24888, 33234, 41580, 49926, 58272, 65536];
        data_sample_at_11111[87] = 0;
        ref_cb[87] = vec![8234, 16619, 25004, 33389, 41774, 50159, 58544, 65536];
        data_sample_at_11111[88] = 70;
        ref_cb[88] = vec![8272, 16480, 24688, 32896, 41104, 49312, 57520, 65536];
        data_sample_at_11111[89] = 136;
        ref_cb[89] = vec![8310, 16556, 24802, 33048, 41294, 49540, 57786, 65536];
        data_sample_at_11111[90] = 0;
        ref_cb[90] = vec![8348, 16632, 24916, 33200, 41484, 49768, 58052, 65536];
        data_sample_at_11111[91] = 0;
        ref_cb[91] = vec![8386, 16708, 25030, 33352, 41674, 49996, 58318, 65536];
        data_sample_at_11111[92] = 101;
        ref_cb[92] = vec![8204, 16564, 24924, 33284, 41644, 50004, 58364, 65536];
        data_sample_at_11111[93] = 36;
        ref_cb[93] = vec![8241, 16639, 25037, 33435, 41833, 50231, 58629, 65536];
        data_sample_at_11111[94] = 196;
        ref_cb[94] = vec![8278, 16492, 24706, 32920, 41134, 49348, 57562, 65536];
        data_sample_at_11111[95] = 0;
        ref_cb[95] = vec![8315, 16566, 24817, 33068, 41319, 49570, 57821, 65536];
        data_sample_at_11111[96] = 0;
        ref_cb[96] = vec![8352, 16640, 24928, 33216, 41504, 49792, 58080, 65536];
        data_sample_at_11111[97] = 24;
        ref_cb[97] = vec![8389, 16714, 25039, 33364, 41689, 50014, 58339, 65536];
        data_sample_at_11111[98] = 8;
        ref_cb[98] = vec![8200, 16562, 24924, 33286, 41648, 50010, 58372, 65536];
        data_sample_at_11111[99] = 0;
        ref_cb[99] = vec![8236, 16635, 25034, 33433, 41832, 50231, 58630, 65536];
        data_sample_at_11111[100] = 0;
        ref_cb[100] = vec![8272, 16480, 24688, 32896, 41104, 49312, 57520, 65536];
        data_sample_at_11111[101] = 125;
        ref_cb[101] = vec![8308, 16552, 24796, 33040, 41284, 49528, 57772, 65536];
        data_sample_at_11111[102] = 173;
        ref_cb[102] = vec![8344, 16624, 24904, 33184, 41464, 49744, 58024, 65536];
        data_sample_at_11111[103] = 126;
        ref_cb[103] = vec![8380, 16696, 25012, 33328, 41644, 49960, 58276, 65536];
        data_sample_at_11111[104] = 0;
        ref_cb[104] = vec![8416, 16768, 25120, 33472, 41824, 50176, 58528, 65536];
        data_sample_at_11111[105] = 0;
        ref_cb[105] = vec![8219, 16607, 24995, 33383, 41771, 50159, 58547, 65536];
        data_sample_at_11111[106] = 159;
        ref_cb[106] = vec![8254, 16678, 25102, 33526, 41950, 50374, 58798, 65536];
        data_sample_at_11111[107] = 210;
        ref_cb[107] = vec![8289, 16514, 24739, 32964, 41189, 49414, 57639, 65536];
        data_sample_at_11111[108] = 178;
        ref_cb[108] = vec![8324, 16584, 24844, 33104, 41364, 49624, 57884, 65536];
        data_sample_at_11111[109] = 0;
        ref_cb[109] = vec![8359, 16654, 24949, 33244, 41539, 49834, 58129, 65536];
        data_sample_at_11111[110] = 0;
        ref_cb[110] = vec![8394, 16724, 25054, 33384, 41714, 50044, 58374, 65536];
        data_sample_at_11111[111] = 170;
        ref_cb[111] = vec![8429, 16794, 25159, 33524, 41889, 50254, 58619, 65536];
        data_sample_at_11111[112] = 173;
        ref_cb[112] = vec![8224, 16624, 25024, 33424, 41824, 50224, 58624, 65536];
        data_sample_at_11111[113] = 235;
        ref_cb[113] = vec![8258, 16452, 24646, 32840, 41034, 49228, 57422, 65536];
        data_sample_at_11111[114] = 0;
        ref_cb[114] = vec![8292, 16520, 24748, 32976, 41204, 49432, 57660, 65536];
        data_sample_at_11111[115] = 0;
        ref_cb[115] = vec![8326, 16588, 24850, 33112, 41374, 49636, 57898, 65536];
        data_sample_at_11111[116] = 0;
        ref_cb[116] = vec![8360, 16656, 24952, 33248, 41544, 49840, 58136, 65536];
        data_sample_at_11111[117] = 24;
        ref_cb[117] = vec![8394, 16724, 25054, 33384, 41714, 50044, 58374, 65536];
        data_sample_at_11111[118] = 228;
        ref_cb[118] = vec![8428, 16792, 25156, 33520, 41884, 50248, 58612, 65536];
        data_sample_at_11111[119] = 0;
        ref_cb[119] = vec![8215, 16613, 25011, 33409, 41807, 50205, 58603, 65536];
        data_sample_at_11111[120] = 0;
        ref_cb[120] = vec![8248, 16680, 25112, 33544, 41976, 50408, 58840, 65536];
        data_sample_at_11111[121] = 0;
        ref_cb[121] = vec![8281, 16498, 24715, 32932, 41149, 49366, 57583, 65536];
        data_sample_at_11111[122] = 101;
        ref_cb[122] = vec![8314, 16564, 24814, 33064, 41314, 49564, 57814, 65536];
        data_sample_at_11111[123] = 174;
        ref_cb[123] = vec![8347, 16630, 24913, 33196, 41479, 49762, 58045, 65536];
        data_sample_at_11111[124] = 126;
        ref_cb[124] = vec![8380, 16696, 25012, 33328, 41644, 49960, 58276, 65536];
        data_sample_at_11111[125] = 0;
        ref_cb[125] = vec![8413, 16762, 25111, 33460, 41809, 50158, 58507, 65536];
        data_sample_at_11111[126] = 0;
        ref_cb[126] = vec![8192, 16574, 24956, 33338, 41720, 50102, 58484, 65536];
        data_sample_at_11111[127] = 0;
        ref_cb[127] = vec![8224, 16639, 25054, 33469, 41884, 50299, 58714, 65536];

        // Now run the loop with this reference data.
        for i in 0..128 {
            let data = get_triggering_base_data(65536, i);

            // This check is here so that the tests written against this chunker
            // can verify that the test data input is correct.
            assert_eq!(data[11111], data_sample_at_11111[i]);

            // Uncomment to create the line above.
            // eprintln!("data_sample_at_11111[{i}]={};", data[11111]);

            // Now, run the chunks through the default chunker.
            let chunks = ChunkerTestWrapper::default().next_block(&data, true);

            // Get the boundaries indices as determined by the size of the chunks above.
            let chunk_boundaries: Vec<usize> = get_chunk_boundaries(&chunks);

            // Uncomment this to generate the table above.
            // eprintln!("ref_cb[{i}]=vec!{chunk_boundaries:?};");

            assert_eq!(chunk_boundaries, ref_cb[i]);
        }

        // eprintln!("assert_eq!(chunk_boundaries, vec!{chunk_boundaries:?})");
        // assert_eq!(chunk_boundaries, vec![131072, 262144, 393216, 524288, 655360, 786432, 917504, 1000000])
    }
}