trueno 0.17.1

High-performance SIMD compute library with GPU support for matrix operations
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

#![allow(missing_docs)]
//! Scalar Q4_K GEMV implementations.
//!
//! Contains the baseline scalar dot product, 4-way unrolled fused GEMV,
//! and the scalar chunk processor for parallel dispatch.

use super::super::{parse_q4k_header, SUPER_BLOCK_BYTES, SUPER_BLOCK_SIZE};

/// Scalar dot product for one Q4K super-block row.
#[inline(always)]
fn process_q4k_superblock_scalar(
    sb_data: &[u8],
    input: &[f32],
    input_offset: usize,
    in_dim: usize,
) -> f32 {
    let (d, dmin, scales, mins) = parse_q4k_header(sb_data);
    let qs = sb_data.get(16..144).expect("Q4_K: need ≥144 bytes for qs");
    let mut sum = 0.0f32;

    for chunk in 0..4 {
        let chunk_start = chunk * 64;
        let q_start = chunk * 32;

        let d1 = d * f32::from(scales[chunk * 2]);
        let dm1 = dmin * f32::from(mins[chunk * 2]);
        let d2 = d * f32::from(scales[chunk * 2 + 1]);
        let dm2 = dmin * f32::from(mins[chunk * 2 + 1]);

        for i in 0..32 {
            let input_idx = input_offset + chunk_start + i;
            if input_idx < in_dim {
                let q_val = (qs[q_start + i] & 0x0F) as f32;
                sum += (d1 * q_val - dm1) * input[input_idx];
            }
        }
        for i in 0..32 {
            let input_idx = input_offset + chunk_start + 32 + i;
            if input_idx < in_dim {
                let q_val = (qs[q_start + i] >> 4) as f32;
                sum += (d2 * q_val - dm2) * input[input_idx];
            }
        }
    }
    sum
}

pub fn matmul_q4k_f32_scalar(
    q4k_data: &[u8],
    input: &[f32],
    out_dim: usize,
    in_dim: usize,
) -> Vec<f32> {
    assert_eq!(input.len(), in_dim, "Input length mismatch");
    assert!(
        in_dim % SUPER_BLOCK_SIZE == 0 || in_dim < SUPER_BLOCK_SIZE,
        "in_dim must be multiple of 256 (or smaller for padding)"
    );

    let num_blocks_per_row = (in_dim + SUPER_BLOCK_SIZE - 1) / SUPER_BLOCK_SIZE;
    let row_bytes = num_blocks_per_row * SUPER_BLOCK_BYTES;
    let expected_size = out_dim * row_bytes;

    assert!(
        q4k_data.len() >= expected_size,
        "Q4K data too small: {} < {}",
        q4k_data.len(),
        expected_size
    );

    let mut output = vec![0.0f32; out_dim];

    for out_idx in 0..out_dim {
        let row_start = out_idx * row_bytes;
        let mut sum = 0.0f32;

        for sb_idx in 0..num_blocks_per_row {
            let sb_start = row_start + sb_idx * SUPER_BLOCK_BYTES;
            let sb_data = &q4k_data[sb_start..sb_start + SUPER_BLOCK_BYTES];
            let input_offset = sb_idx * SUPER_BLOCK_SIZE;
            sum += process_q4k_superblock_scalar(sb_data, input, input_offset, in_dim);
        }

        output[out_idx] = sum;
    }

    output
}

/// Process one Q4K nibble half (low or high) with 4-way unrolled accumulation.
#[inline(always)]
fn process_q4k_nibble_half(
    qs: &[u8],
    q_start: usize,
    input: &[f32],
    input_base: usize,
    in_dim: usize,
    d_val: f32,
    dm_val: f32,
    shift: u8,
    acc: &mut [f32; 4],
) {
    let mut i = 0;
    while i + 3 < 32 {
        let idx = input_base + i;
        if idx + 3 < in_dim {
            let q0 = ((qs[q_start + i] >> shift) & 0x0F) as f32;
            let q1 = ((qs[q_start + i + 1] >> shift) & 0x0F) as f32;
            let q2 = ((qs[q_start + i + 2] >> shift) & 0x0F) as f32;
            let q3 = ((qs[q_start + i + 3] >> shift) & 0x0F) as f32;

            acc[0] = (d_val * q0 - dm_val).mul_add(input[idx], acc[0]);
            acc[1] = (d_val * q1 - dm_val).mul_add(input[idx + 1], acc[1]);
            acc[2] = (d_val * q2 - dm_val).mul_add(input[idx + 2], acc[2]);
            acc[3] = (d_val * q3 - dm_val).mul_add(input[idx + 3], acc[3]);
        }
        i += 4;
    }
    while i < 32 {
        let idx = input_base + i;
        if idx < in_dim {
            let q_val = ((qs[q_start + i] >> shift) & 0x0F) as f32;
            acc[0] = (d_val * q_val - dm_val).mul_add(input[idx], acc[0]);
        }
        i += 1;
    }
}

/// Fused Q4_K matrix-vector multiply (optimized with 4-way unrolling)
///
/// This version uses 4 independent accumulators to improve instruction-level
/// parallelism while maintaining scalar correctness.
///
/// # Arguments
/// Same as `matmul_q4k_f32_scalar`
pub fn matmul_q4k_f32(q4k_data: &[u8], input: &[f32], out_dim: usize, in_dim: usize) -> Vec<f32> {
    assert_eq!(input.len(), in_dim, "Input length mismatch");

    let num_blocks_per_row = (in_dim + SUPER_BLOCK_SIZE - 1) / SUPER_BLOCK_SIZE;
    let row_bytes = num_blocks_per_row * SUPER_BLOCK_BYTES;

    let mut output = vec![0.0f32; out_dim];

    for out_idx in 0..out_dim {
        let row_start = out_idx * row_bytes;
        let mut acc = [0.0f32; 4];

        for sb_idx in 0..num_blocks_per_row {
            let sb_start = row_start + sb_idx * SUPER_BLOCK_BYTES;
            let sb_data = &q4k_data[sb_start..sb_start + SUPER_BLOCK_BYTES];

            let (d, dmin, scales, mins) = parse_q4k_header(sb_data);
            let qs = sb_data.get(16..144).expect("Q4_K: need ≥144 bytes for qs");
            let input_offset = sb_idx * SUPER_BLOCK_SIZE;

            for chunk in 0..4 {
                let chunk_start = chunk * 64;
                let q_start = chunk * 32;

                let d1 = d * f32::from(scales[chunk * 2]);
                let dm1 = dmin * f32::from(mins[chunk * 2]);
                let d2 = d * f32::from(scales[chunk * 2 + 1]);
                let dm2 = dmin * f32::from(mins[chunk * 2 + 1]);

                let base_low = input_offset + chunk_start;
                process_q4k_nibble_half(qs, q_start, input, base_low, in_dim, d1, dm1, 0, &mut acc);

                let base_high = input_offset + chunk_start + 32;
                process_q4k_nibble_half(
                    qs, q_start, input, base_high, in_dim, d2, dm2, 4, &mut acc,
                );
            }
        }

        output[out_idx] = (acc[0] + acc[1]) + (acc[2] + acc[3]);
    }

    output
}

/// Scalar chunk processor for parallel dispatch.
pub(crate) fn compute_chunk_q4k_scalar(
    q4k_data: &[u8],
    input: &[f32],
    chunk: &mut [f32],
    start_row: usize,
    out_dim: usize,
    in_dim: usize,
    num_blocks_per_row: usize,
    row_bytes: usize,
) {
    for (local_idx, out_val) in chunk.iter_mut().enumerate() {
        let out_idx = start_row + local_idx;
        if out_idx >= out_dim {
            break;
        }

        let row_start = out_idx * row_bytes;
        let mut sum = 0.0f32;

        for sb_idx in 0..num_blocks_per_row {
            let sb_start = row_start + sb_idx * SUPER_BLOCK_BYTES;
            if sb_start + SUPER_BLOCK_BYTES > q4k_data.len() {
                break;
            }
            let sb_data = &q4k_data[sb_start..sb_start + SUPER_BLOCK_BYTES];
            let input_offset = sb_idx * SUPER_BLOCK_SIZE;
            sum += process_q4k_superblock_scalar(sb_data, input, input_offset, in_dim);
        }

        *out_val = sum;
    }
}