mopenzl 0.2.3

OpenZL compression and decompression in safe Rust; encoding is byte-identical to Meta's reference library
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204

//! The bit-split nodes (`encode_bitSplit_binding.c`,
//! `encode_bitSplit_kernel.c`, `encode_bitsplit_{top8,fp,bf16}_binding.c`),
//! codec 63.
//!
//! A numeric stream is cut into bit fields, least significant first; each
//! field becomes its own numeric output of the smallest width that holds it
//! (1 to 8 bits: one byte, up to 16: two, up to 32: four, else eight). The
//! header is the input element width followed by the field widths, without
//! the last one when the fields cover the whole element. Fields may cover
//! less than the element only when the remaining top bits are all zero.
//!
//! * `bitSplit` (private node 75) takes the field widths from the copy
//!   parameter 701 (one byte each, 1 to 64 of them);
//! * `bitsplit_top8` splits the largest element's significant bits into the
//!   low remainder and the top 8 bits;
//! * `bitsplit_fp` and `bitsplit_bf16` use the IEEE binary16/32/64 and
//!   bfloat16 mantissa, exponent and sign widths.
//!
//! The reference's special kernels for the floating-point layouts write what
//! the generic kernel writes, so this port has the generic one only.

use alloc::vec::Vec;

use super::super::FnEncoder;
use crate::compressor::core::data::Stream;
use crate::compressor::core::error::NodeError;
use crate::compressor::core::exec::NodeCx;
use crate::shared::numeric::{elements, push_le};

/// `ZL_BITSPLIT_WIDTHS_PID`.
const WIDTHS_PID: i32 = 701;
/// The most fields one split takes.
const MAX_FIELDS: usize = 64;

/// `bit_split` (private node 75).
pub(crate) static BITSPLIT: FnEncoder = FnEncoder(bit_split);
/// `ZL_NODE_BITSPLIT_TOP8` (node 45).
pub(crate) static BITSPLIT_TOP8: FnEncoder = FnEncoder(top8);
/// `ZL_NODE_BITSPLIT_FP` (node 46).
pub(crate) static BITSPLIT_FP: FnEncoder = FnEncoder(fp);
/// `ZL_NODE_BITSPLIT_BF16` (node 47).
pub(crate) static BITSPLIT_BF16: FnEncoder = FnEncoder(bf16);

/// `ZL_bitSplit_outputEltWidth`.
const fn output_width(bits: u8) -> usize {
    match bits {
        0..=8 => 1,
        9..=16 => 2,
        17..=32 => 4,
        _ => 8,
    }
}

/// `ZL_bitSplit_paramsAreValid`: the sum of the widths, or `None` when one
/// is zero or the sum exceeds the element's bits.
fn checked_sum(widths: &[u8], element_bits: usize) -> Option<usize> {
    if widths.contains(&0) {
        return None;
    }
    let sum: usize = widths.iter().map(|&width| usize::from(width)).sum();
    (sum <= element_bits).then_some(sum)
}

/// `ZL_bitSplitEncode`: one stream of fields per width.
fn split(bytes: &[u8], width: usize, widths: &[u8]) -> Vec<Vec<u8>> {
    let mut outputs: Vec<Vec<u8>> = widths.iter().map(|_| Vec::new()).collect();
    for value in elements(bytes, width) {
        let mut position = 0u32;
        for (output, &bits) in outputs.iter_mut().zip(widths) {
            let mask = if bits >= 64 { u64::MAX } else { (1u64 << bits) - 1 };
            push_le(output, (value >> position) & mask, output_width(bits));
            position += u32::from(bits);
        }
    }
    outputs
}

/// `EI_bitSplit_withWidths`.
fn with_widths<'a>(
    cx: &mut NodeCx<'_, 'a>,
    input: &Stream<'a>,
    widths: &[u8],
) -> Result<(), NodeError> {
    if widths.is_empty() {
        return Err(NodeError::InvalidParam(
            "bitSplit requires at least one bit width",
        ));
    }
    if widths.len() > MAX_FIELDS {
        return Err(NodeError::InvalidParam(
            "bitSplit supports at most 64 bit widths",
        ));
    }
    let width = input.width();
    let element_bits = 8 * width;
    let sum = checked_sum(widths, element_bits).ok_or(NodeError::InvalidParam(
        "bitSplit parameter validation failed",
    ))?;
    if sum < element_bits {
        let mask = !((1u64 << sum) - 1);
        if elements(input.bytes(), width).any(|value| value & mask != 0) {
            return Err(NodeError::InvalidInput(
                "bitSplit: top bits must be zero for partial coverage",
            ));
        }
    }
    let mut header = Vec::with_capacity(1 + widths.len());
    header.push(width as u8);
    let kept = if sum == element_bits {
        widths.len() - 1
    } else {
        widths.len()
    };
    header.extend_from_slice(&widths[..kept]);
    cx.send_header(&header)?;
    for (output, &bits) in split(input.bytes(), width, widths).into_iter().zip(widths) {
        cx.emit(0, output_width(bits), output)?;
    }
    Ok(())
}

/// `EI_bitSplit`.
fn bit_split<'a>(cx: &mut NodeCx<'_, 'a>, inputs: &[Stream<'a>]) -> Result<(), NodeError> {
    let widths = cx.copy_param(WIDTHS_PID).ok_or(NodeError::InvalidParam(
        "bitSplit requires bit widths parameter",
    ))?;
    let widths = widths.to_vec();
    with_widths(cx, &inputs[0], &widths)
}

/// `EI_bitsplit_top8`.
fn top8<'a>(cx: &mut NodeCx<'_, 'a>, inputs: &[Stream<'a>]) -> Result<(), NodeError> {
    let input = &inputs[0];
    let max = elements(input.bytes(), input.width()).max().unwrap_or(0);
    let effective = 64 - max.leading_zeros() as u8;
    if effective <= 8 {
        with_widths(cx, input, &[effective.max(1)])
    } else {
        with_widths(cx, input, &[effective - 8, 8])
    }
}

/// `EI_bitsplit_fp`.
fn fp<'a>(cx: &mut NodeCx<'_, 'a>, inputs: &[Stream<'a>]) -> Result<(), NodeError> {
    let input = &inputs[0];
    let widths: [u8; 3] = match input.width() {
        2 => [10, 5, 1],
        4 => [23, 8, 1],
        8 => [52, 11, 1],
        _ => {
            return Err(NodeError::InvalidInput(
                "bitsplit_fp requires IEEE 754 floats (2, 4, or 8-byte elements)",
            ));
        }
    };
    with_widths(cx, input, &widths)
}

/// `EI_bitsplit_bf16`.
fn bf16<'a>(cx: &mut NodeCx<'_, 'a>, inputs: &[Stream<'a>]) -> Result<(), NodeError> {
    let input = &inputs[0];
    if input.width() != 2 {
        return Err(NodeError::InvalidInput(
            "bitsplit_bf16 requires 2-byte elements (bfloat16)",
        ));
    }
    with_widths(cx, input, &[7, 8, 1])
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn output_widths_follow_the_field_size() {
        assert_eq!(
            [1u8, 8, 9, 16, 17, 32, 33, 64].map(output_width),
            [1, 1, 2, 2, 4, 4, 8, 8]
        );
    }

    #[test]
    fn widths_are_validated_against_the_element() {
        assert_eq!(checked_sum(&[4, 4], 8), Some(8));
        assert_eq!(checked_sum(&[4, 3], 8), Some(7));
        assert_eq!(checked_sum(&[4, 5], 8), None);
        assert_eq!(checked_sum(&[4, 0], 8), None);
    }

    #[test]
    fn fields_are_cut_least_significant_first() {
        let bytes: Vec<u8> = [0xabcdu16, 0x0123].iter().flat_map(|v| v.to_le_bytes()).collect();
        let outputs = split(&bytes, 2, &[4, 4, 8]);
        assert_eq!(outputs, [alloc::vec![0xd, 0x3], alloc::vec![0xc, 0x2], alloc::vec![0xab, 0x01]]);
        // A 52/11/1 split of binary64.
        let value = (1u64 << 63) | (0x3ffu64 << 52) | 0x5_0000_0000_0001;
        let outputs = split(&value.to_le_bytes(), 8, &[52, 11, 1]);
        assert_eq!(outputs[0], (value & ((1 << 52) - 1)).to_le_bytes());
        assert_eq!(outputs[1], 0x3ffu16.to_le_bytes());
        assert_eq!(outputs[2], [1]);
        // A 64-bit field takes the whole element.
        assert_eq!(split(&u64::MAX.to_le_bytes(), 8, &[64])[0], u64::MAX.to_le_bytes());
    }
}