fpzip-rs 0.1.0

Lossless and lossy floating-point compression for multi-dimensional arrays
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
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282

use crate::codec::range_encoder::RangeEncoder;
use crate::codec::rc_qs_model::RCQsModel;
use crate::core::pc_map;

/// PC_BIT_MAX threshold: bits <= 8 use small alphabet, bits > 8 use wide alphabet.
const PC_BIT_MAX: u32 = 8;

/// Returns the number of symbols for the given bit width.
pub fn symbol_count(bits: u32) -> usize {
    if bits > PC_BIT_MAX {
        // Wide alphabet: 2 * bits + 1
        (2 * bits + 1) as usize
    } else {
        // Small alphabet: 2 * (1 << bits) - 1
        (2 * (1u32 << bits) - 1) as usize
    }
}

/// Predictive coder encoder for float values.
pub struct PCEncoderFloat<'a> {
    encoder: &'a mut RangeEncoder,
    model: &'a mut RCQsModel,
    bits: u32,
}

impl<'a> PCEncoderFloat<'a> {
    pub fn new(encoder: &'a mut RangeEncoder, model: &'a mut RCQsModel, bits: u32) -> Self {
        Self {
            encoder,
            model,
            bits,
        }
    }

    /// Encodes a mapped value with prediction. Returns the actual value (masked).
    #[inline]
    pub fn encode(&mut self, actual: u32, predicted: u32) -> u32 {
        let actual = pc_map::mask_u32(actual, self.bits);
        let predicted = pc_map::mask_u32(predicted, self.bits);

        if self.bits > PC_BIT_MAX {
            self.encode_wide(actual, predicted);
        } else {
            self.encode_narrow(actual, predicted);
        }
        actual
    }

    #[inline]
    fn encode_wide(&mut self, actual: u32, predicted: u32) {
        let bias = self.bits;
        if predicted < actual {
            let d = actual - predicted;
            let k = 31 - d.leading_zeros();
            self.encoder.encode_with_model(bias + 1 + k, self.model);
            self.encoder.encode_uint(d - (1u32 << k), k as i32);
        } else if predicted > actual {
            let d = predicted - actual;
            let k = 31 - d.leading_zeros();
            self.encoder.encode_with_model(bias - 1 - k, self.model);
            self.encoder.encode_uint(d - (1u32 << k), k as i32);
        } else {
            self.encoder.encode_with_model(bias, self.model);
        }
    }

    #[inline]
    fn encode_narrow(&mut self, actual: u32, predicted: u32) {
        let bias = (1u32 << self.bits) - 1;
        let symbol = bias.wrapping_add(actual).wrapping_sub(predicted);
        self.encoder.encode_with_model(symbol, self.model);
    }
}

/// Predictive coder encoder for double values.
pub struct PCEncoderDouble<'a> {
    encoder: &'a mut RangeEncoder,
    model: &'a mut RCQsModel,
    bits: u32,
}

impl<'a> PCEncoderDouble<'a> {
    pub fn new(encoder: &'a mut RangeEncoder, model: &'a mut RCQsModel, bits: u32) -> Self {
        Self {
            encoder,
            model,
            bits,
        }
    }

    /// Encodes a mapped value with prediction. Returns the actual value (masked).
    #[inline]
    pub fn encode(&mut self, actual: u64, predicted: u64) -> u64 {
        let actual = pc_map::mask_u64(actual, self.bits);
        let predicted = pc_map::mask_u64(predicted, self.bits);

        if self.bits > PC_BIT_MAX {
            self.encode_wide(actual, predicted);
        } else {
            self.encode_narrow(actual, predicted);
        }
        actual
    }

    #[inline]
    fn encode_wide(&mut self, actual: u64, predicted: u64) {
        let bias = self.bits;
        if predicted < actual {
            let d = actual - predicted;
            let k = 63 - d.leading_zeros();
            self.encoder.encode_with_model(bias + 1 + k, self.model);
            self.encoder.encode_ulong(d - (1u64 << k), k as i32);
        } else if predicted > actual {
            let d = predicted - actual;
            let k = 63 - d.leading_zeros();
            self.encoder.encode_with_model(bias - 1 - k, self.model);
            self.encoder.encode_ulong(d - (1u64 << k), k as i32);
        } else {
            self.encoder.encode_with_model(bias, self.model);
        }
    }

    #[inline]
    fn encode_narrow(&mut self, actual: u64, predicted: u64) {
        let bias = (1u64 << self.bits) - 1;
        let symbol = bias.wrapping_add(actual).wrapping_sub(predicted);
        self.encoder.encode_with_model(symbol as u32, self.model);
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::codec::range_decoder::RangeDecoder;
    use crate::core::pc_decoder::{PCDecoderDouble, PCDecoderFloat};

    fn round_trip_float(actual: u32, predicted: u32, bits: u32) -> u32 {
        let symbols = symbol_count(bits);
        let mut enc = RangeEncoder::new();
        let mut model = RCQsModel::with_defaults(true, symbols);
        {
            let mut pc = PCEncoderFloat::new(&mut enc, &mut model, bits);
            pc.encode(actual, predicted);
        }
        let data = enc.finish();

        let mut dec = RangeDecoder::new(&data);
        dec.init();
        let mut dmodel = RCQsModel::with_defaults(false, symbols);
        let mut pcd = PCDecoderFloat::new(&mut dec, &mut dmodel, bits);
        pcd.decode(predicted)
    }

    fn round_trip_double(actual: u64, predicted: u64, bits: u32) -> u64 {
        let symbols = symbol_count(bits);
        let mut enc = RangeEncoder::new();
        let mut model = RCQsModel::with_defaults(true, symbols);
        {
            let mut pc = PCEncoderDouble::new(&mut enc, &mut model, bits);
            pc.encode(actual, predicted);
        }
        let data = enc.finish();

        let mut dec = RangeDecoder::new(&data);
        dec.init();
        let mut dmodel = RCQsModel::with_defaults(false, symbols);
        let mut pcd = PCDecoderDouble::new(&mut dec, &mut dmodel, bits);
        pcd.decode(predicted)
    }

    #[test]
    fn float_perfect_prediction() {
        assert_eq!(round_trip_float(100, 100, 32), 100);
    }

    #[test]
    fn float_underprediction() {
        assert_eq!(round_trip_float(200, 100, 32), 200);
    }

    #[test]
    fn float_overprediction() {
        assert_eq!(round_trip_float(50, 200, 32), 50);
    }

    #[test]
    fn float_all_delta_sizes() {
        let predicted = 0u32;
        for k in 0..31 {
            let delta = 1u32 << k;
            let actual = predicted.wrapping_add(delta);
            assert_eq!(
                round_trip_float(actual, predicted, 32),
                actual,
                "k={k} under"
            );
            let actual2 = predicted.wrapping_sub(delta);
            assert_eq!(
                round_trip_float(actual2, predicted, 32),
                actual2,
                "k={k} over"
            );
        }
    }

    #[test]
    fn double_perfect_prediction() {
        assert_eq!(round_trip_double(100, 100, 64), 100);
    }

    #[test]
    fn double_underprediction() {
        assert_eq!(round_trip_double(200, 100, 64), 200);
    }

    #[test]
    fn double_overprediction() {
        assert_eq!(round_trip_double(50, 200, 64), 50);
    }

    #[test]
    fn float_sequence() {
        let symbols = symbol_count(32);
        let mut enc = RangeEncoder::new();
        let mut model = RCQsModel::with_defaults(true, symbols);
        let pairs: Vec<(u32, u32)> =
            vec![(100, 100), (200, 100), (50, 200), (0, 0), (0xFFFFFFFF, 0)];
        {
            let mut pc = PCEncoderFloat::new(&mut enc, &mut model, 32);
            for &(a, p) in &pairs {
                pc.encode(a, p);
            }
        }
        let data = enc.finish();

        let mut dec = RangeDecoder::new(&data);
        dec.init();
        let mut dmodel = RCQsModel::with_defaults(false, symbols);
        let mut pcd = PCDecoderFloat::new(&mut dec, &mut dmodel, 32);
        for &(a, p) in &pairs {
            assert_eq!(pcd.decode(p), a);
        }
    }

    #[test]
    fn double_all_delta_sizes() {
        let predicted = 0u64;
        for k in 0..63 {
            let delta = 1u64 << k;
            let actual = predicted.wrapping_add(delta);
            assert_eq!(
                round_trip_double(actual, predicted, 64),
                actual,
                "k={k} under"
            );
        }
    }

    // Small alphabet tests (bits <= 8)
    #[test]
    fn float_narrow_round_trip() {
        for bits in [2, 4, 8] {
            let mask = (1u32 << bits) - 1;
            for a in 0..=mask.min(15) {
                for p in 0..=mask.min(15) {
                    let result = round_trip_float(a, p, bits);
                    assert_eq!(result, a, "bits={bits} a={a} p={p}");
                }
            }
        }
    }

    #[test]
    fn float_reduced_precision_wide() {
        // bits=16 uses wide alphabet
        let mask = 0xFFFFu32;
        assert_eq!(round_trip_float(0, 0, 16), 0);
        assert_eq!(round_trip_float(mask, 0, 16), mask);
        assert_eq!(round_trip_float(100, 200, 16), 100);
        assert_eq!(round_trip_float(200, 100, 16), 200);
    }
}