speakrs 0.6.0

Fast Rust speaker diarization with pyannote-level accuracy and native CPU/CoreML/CUDA inference
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

// both compilation targets must use the same physical ownership maps
#[path = "../../../../crates/speakrs-cuda-kernels/src/lstmproj/exchange.rs"]
mod device_exchange;
#[path = "../../../../crates/speakrs-cuda-kernels/src/lstmproj/tiled_exchange.rs"]
mod device_tiled_exchange;

use super::super::lstmproj::RecurrencePlan;
use super::super::lstmproj::layout::{
    AlignedSpan, ExchangeLayout, GATE_COLUMNS, GROUPS, HIDDEN, ProjectionLayout, STATE_TILE,
    Schedule, THREADS, TILE_ROWS, TILED_GROUPS, TILED_ROWS, exchange, pack_directions,
    pack_input_directions, padded_input, round_tf32, tiled_exchange,
};

#[test]
fn geometry_matches_the_kernel_area() {
    assert_eq!(THREADS, 256);
    assert_eq!(GROUPS, 16);
    assert_eq!(padded_input(60), 64);
    assert_eq!(padded_input(256), 256);
}

#[test]
fn eight_unit_blocks_halve_the_blocks_per_tile_in_the_schedule() {
    // 16 blocks per tile let b64 run 8-row tiles concurrently where four-unit
    // blocks need 16-row tiles
    let wide = Schedule::for_groups(GROUPS, 64, 350, Some(350)).expect("fits");
    assert_eq!((wide.tile_rows, wide.tiles, wide.concurrent), (8, 8, true));
    let legacy = Schedule::new(64, 350, Some(350)).expect("fits");
    assert_eq!(legacy.tile_rows, 16);

    let serial = Schedule::for_groups(GROUPS, 65, GROUPS, None).expect("fits");
    assert_eq!(serial.tile_rows, TILE_ROWS);
    assert!(!serial.concurrent);
    assert_eq!(serial.launches().count(), 3);
    assert!(Schedule::for_groups(GROUPS, 1, GROUPS - 1, Some(350)).is_err());
}

#[test]
fn simt_and_tensor_packing_transpose_the_packed_rows() {
    let columns = 60;
    let source: Vec<f32> = (0..2 * GATE_COLUMNS * columns)
        .map(|i| (i % 977) as f32 * 0.25 - 100.0)
        .collect();
    let rows = pack_directions(&source, columns);
    let simt = pack_input_directions(&source, columns, ProjectionLayout::Simt);
    let tensor = pack_input_directions(&source, columns, ProjectionLayout::Tensor);
    let padded = padded_input(columns);
    for direction in 0..2 {
        let base = direction * GATE_COLUMNS * padded;
        for column in [0, 1, 5, 130, GATE_COLUMNS - 1] {
            for k in 0..padded {
                let expected = if k < columns {
                    rows[(direction * GATE_COLUMNS + column) * columns + k]
                } else {
                    0.0
                };
                assert_eq!(simt[base + k * GATE_COLUMNS + column], expected);
                assert_eq!(
                    tensor[base + column * padded + k],
                    round_tf32(expected),
                    "direction {direction} column {column} k {k}"
                );
            }
        }
    }
}

#[test]
fn tf32_rounding_keeps_ten_mantissa_bits_with_ties_to_even() {
    // 1 + 2^-11 is halfway between two TF32 values; the even neighbour is 1
    assert_eq!(round_tf32(1.0 + 2f32.powi(-11)), 1.0);
    // 1 + 3 * 2^-11 is halfway again; the even neighbour is 1 + 2^-9
    assert_eq!(round_tf32(1.0 + 3.0 * 2f32.powi(-11)), 1.0 + 2f32.powi(-9));
    assert_eq!(round_tf32(1.0 + 2f32.powi(-12)), 1.0);
    assert_eq!(round_tf32(-3.5), -3.5);
    assert!(round_tf32(f32::NAN).is_nan());
    assert_eq!(round_tf32(f32::INFINITY), f32::INFINITY);
    assert_eq!(round_tf32(1.0 + 2f32.powi(-12)).to_bits() & 0x1fff, 0);
}

#[test]
fn aligned_span_starts_on_a_two_mebibyte_boundary_inside_its_allocation() {
    // allocations are naturally aligned for their element type
    let align = 2 * 1024 * 1024;
    for address in [0u64, 8, 256, align as u64 - 8, align as u64 + 512] {
        let len = 1000;
        let span = AlignedSpan::<u64>::new(address, len);
        let range = span.range();
        assert_eq!((address as usize + range.start * 8) % align, 0);
        assert_eq!(range.len(), len);
        assert!(range.end <= AlignedSpan::<u64>::allocation_len(len));
    }
}

#[test]
fn exchange_words_stay_inside_their_tile_and_producer_regions() {
    assert_eq!(exchange::TILE, device_exchange::TILE);
    assert_eq!(exchange::SLOT, device_exchange::SLOT);
    assert_eq!(exchange::COPIES, device_exchange::COPIES);
    let mut bulk = std::collections::BTreeSet::new();
    for logical in 0..32 * HIDDEN {
        let word = exchange::bulk_word(logical);
        assert_eq!(word, device_exchange::bulk_word(logical));
        assert!(word < exchange::SLOT);
        assert!(bulk.insert(word), "two hidden values share word {word}");
        // eight consecutive units of one row share one producer's line
        assert_eq!(
            word / exchange::LINE,
            exchange::bulk_word(logical / 8 * 8) / exchange::LINE
        );
    }

    for flag in [1, 2, 3, 590, u32::MAX] {
        let mut single = std::collections::BTreeSet::new();
        for unit in 0..HIDDEN {
            for copy in 0..exchange::COPIES {
                let word = exchange::single_copy(exchange::single_base(unit), copy, flag);
                assert_eq!(
                    word,
                    device_exchange::single_copy(device_exchange::single_base(unit), copy, flag)
                );
                assert!(word < exchange::SLOT);
                // a producer group of eight units keeps its own 4 KiB region
                assert_eq!(word / 512, unit / 8);
                assert!(single.insert(word), "flag {flag}: copies collide at {word}");
            }
        }
    }

    let layout = ExchangeLayout::new(3, STATE_TILE);
    assert_eq!(layout.direction(1), 3 * STATE_TILE + exchange::PAD);
    assert_eq!(layout.words(), 2 * layout.direction(1));
}

#[test]
fn tiled_exchange_gives_each_producer_row_one_line() {
    assert_eq!(tiled_exchange::TILE, device_tiled_exchange::TILE);
    assert_eq!(tiled_exchange::SLOT, device_tiled_exchange::SLOT);
    assert_eq!(tiled_exchange::ROWS, device_tiled_exchange::ROWS);
    assert_eq!(tiled_exchange::LINE, device_tiled_exchange::LINE);
    assert_eq!(tiled_exchange::SLOT % tiled_exchange::LINE, 0);
    assert_eq!(tiled_exchange::TILE % tiled_exchange::LINE, 0);

    let units = HIDDEN / TILED_GROUPS;
    let mut words = std::collections::BTreeSet::new();
    for row in 0..TILED_ROWS {
        for unit in 0..HIDDEN {
            let word = tiled_exchange::word(row, unit);
            assert_eq!(word, device_tiled_exchange::word(row, unit));
            assert!(word + tiled_exchange::PAD < tiled_exchange::SLOT + 1);
            assert!(words.insert(word), "two hidden values share word {word}");
            // a block's units of one row fill exactly one 128-byte line
            let first = tiled_exchange::word(row, unit / units * units);
            assert_eq!(word / tiled_exchange::LINE, first / tiled_exchange::LINE);
            assert_eq!(first % tiled_exchange::LINE, 0);
        }
    }
}

#[test]
fn tiled_recurrence_needs_full_tiles_and_both_directions_resident() {
    // two directions of eight blocks per tile
    let tiled = |tiles| Some(RecurrencePlan::Tiled { tiles });
    assert_eq!(RecurrencePlan::tiled(32, Some(72)), tiled(4));
    assert_eq!(RecurrencePlan::tiled(32, Some(64)), tiled(4));
    assert_eq!(RecurrencePlan::tiled(32, Some(63)), None);
    assert_eq!(RecurrencePlan::tiled(33, Some(72)), None);
    assert_eq!(RecurrencePlan::tiled(33, Some(80)), tiled(5));
    assert_eq!(RecurrencePlan::tiled(9, Some(32)), tiled(2));
    assert_eq!(RecurrencePlan::tiled(9, Some(31)), None);
    // one tile leaves most SMs idle, and the wide kernel was as fast or faster
    assert_eq!(RecurrencePlan::tiled(8, Some(1000)), None);
    // an unknown budget must not run both grids together
    assert_eq!(RecurrencePlan::tiled(32, None), None);
    // a lone window keeps the wide kernel, though one tile would fit
    assert_eq!(RecurrencePlan::tiled(1, Some(1000)), None);
    assert_eq!(RecurrencePlan::tiled_fit(1, Some(16)), tiled(1));
}

#[test]
fn resident_recurrence_runs_only_on_measured_parts_that_hold_the_matrix() {
    use crate::inference::cuda::ComputeCapability;

    // A100 and A30 opt-in shared memory per block
    assert!(RecurrencePlan::resident_device(
        ComputeCapability::new(8, 0),
        166_912
    ));
    // one direction's matrix is 256 KiB; half must fit beside the static hidden vectors
    assert!(!RecurrencePlan::resident_device(
        ComputeCapability::new(8, 0),
        131_071
    ));
    // Ampere consumer, Ada and Blackwell consumer parts allow about 100 KiB
    for (major, minor) in [(8, 6), (8, 9), (12, 0)] {
        assert!(!RecurrencePlan::resident_device(
            ComputeCapability::new(major, minor),
            101_376
        ));
    }
    // Hopper holds the matrix but was not measured
    assert!(!RecurrencePlan::resident_device(
        ComputeCapability::new(9, 0),
        232_448
    ));
}