modelc 0.1.4

Compile model weight files to standalone executable binaries
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

use crate::runtime::tensor::Tensor;

#[cfg(target_os = "macos")]
use crate::metal::MetalBackend;

pub fn matmul(a: &Tensor, b: &Tensor) -> Tensor {
    // Try GPU acceleration on macOS
    #[cfg(target_os = "macos")]
    {
        if let Some(backend) = MetalBackend::new() {
            if let Some(result) = backend.matmul_gpu(a, b) {
                return result;
            }
        }
    }

    // Fallback to optimized CPU implementation
    matmul_cpu(a, b)
}

fn matmul_cpu(a: &Tensor, b: &Tensor) -> Tensor {
    let (m, k) = (a.shape[0], a.shape[1]);
    let n = b.shape[1];
    assert_eq!(a.shape[1], b.shape[0]);

    let mut out = vec![0.0f32; m * n];

    #[cfg(target_arch = "x86_64")]
    {
        if is_x86_feature_detected!("avx") {
            unsafe { matmul_avx(a, b, &mut out); }
            return Tensor::from_vec(out, vec![m, n]);
        }
    }

    #[cfg(target_arch = "aarch64")]
    {
        if std::arch::is_aarch64_feature_detected!("neon") {
            unsafe { matmul_neon(a, b, &mut out); }
            return Tensor::from_vec(out, vec![m, n]);
        }
    }

    // Parallel fallback for reasonably large matrices
    if m >= 8 {
        use rayon::prelude::*;
        out.par_chunks_exact_mut(n)
            .enumerate()
            .for_each(|(i, row_out)| {
                for j in 0..n {
                    let mut sum = 0.0f32;
                    for p in 0..k {
                        sum += a.data[i * k + p] * b.data[p * n + j];
                    }
                    row_out[j] = sum;
                }
            });
    } else {
        // Fallback scalar implementation
        for i in 0..m {
            for j in 0..n {
                let mut sum = 0.0f32;
                for p in 0..k {
                    sum += a.data[i * k + p] * b.data[p * n + j];
                }
                out[i * n + j] = sum;
            }
        }
    }
    Tensor::from_vec(out, vec![m, n])
}

#[cfg(target_arch = "x86_64")]
unsafe fn matmul_avx(a: &Tensor, b: &Tensor, out: &mut [f32]) {
    use std::arch::x86_64::*;
    let (m, k) = (a.shape[0], a.shape[1]);
    let n = b.shape[1];

    for i in 0..m {
        let mut j = 0;
        while j + 8 <= n {
            let mut sum = _mm256_setzero_ps();
            for p in 0..k {
                let a_val = _mm256_set1_ps(a.data[i * k + p]);
                let b_vec = _mm256_loadu_ps(b.data.as_ptr().add(p * n + j));
                sum = _mm256_add_ps(sum, _mm256_mul_ps(a_val, b_vec));
            }
            _mm256_storeu_ps(out.as_mut_ptr().add(i * n + j), sum);
            j += 8;
        }
        while j < n {
            let mut sum = 0.0f32;
            for p in 0..k {
                sum += a.data[i * k + p] * b.data[p * n + j];
            }
            out[i * n + j] = sum;
            j += 1;
        }
    }
}

#[cfg(target_arch = "aarch64")]
unsafe fn matmul_neon(a: &Tensor, b: &Tensor, out: &mut [f32]) {
    use std::arch::aarch64::*;
    let (m, k) = (a.shape[0], a.shape[1]);
    let n = b.shape[1];

    for i in 0..m {
        let mut j = 0;
        while j + 4 <= n {
            let mut sum = unsafe { vdupq_n_f32(0.0) };
            for p in 0..k {
                let a_val = unsafe { vdupq_n_f32(a.data[i * k + p]) };
                let b_vec = unsafe { vld1q_f32(b.data.as_ptr().add(p * n + j)) };
                sum = unsafe { vfmaq_f32(sum, a_val, b_vec) };
            }
            unsafe { vst1q_f32(out.as_mut_ptr().add(i * n + j), sum) };
            j += 4;
        }
        while j < n {
            let mut sum = 0.0f32;
            for p in 0..k {
                sum += a.data[i * k + p] * b.data[p * n + j];
            }
            out[i * n + j] = sum;
            j += 1;
        }
    }
}

pub fn add(a: &Tensor, b: &Tensor) -> Tensor {
    #[cfg(target_os = "macos")]
    {
        if let Some(backend) = MetalBackend::new() {
            if let Some(result) = backend.add_gpu(a, b) {
                return result;
            }
        }
    }

    if a.shape == b.shape {
        let data: Vec<f32> = a.data.iter().zip(&b.data).map(|(x, y)| x + y).collect();
        Tensor::from_vec(data, a.shape.clone())
    } else if a.shape.len() == 2 && b.shape.len() == 1 && a.shape[1] == b.shape[0] {
        let _rows = a.shape[0];
        let cols = a.shape[1];
        let data: Vec<f32> = a
            .data
            .chunks_exact(cols)
            .flat_map(|row| row.iter().zip(&b.data).map(|(x, y)| x + y))
            .collect();
        Tensor::from_vec(data, a.shape.clone())
    } else if a.shape.len() == 1 && b.shape.len() == 2 && a.shape[0] == b.shape[1] {
        add(b, a)
    } else {
        panic!("add: incompatible shapes {:?} and {:?}", a.shape, b.shape);
    }
}

pub fn mul_scalar(a: &Tensor, s: f32) -> Tensor {
    #[cfg(target_os = "macos")]
    {
        if let Some(backend) = MetalBackend::new() {
            if let Some(result) = backend.mul_scalar_gpu(a, s) {
                return result;
            }
        }
    }

    let data: Vec<f32> = a.data.iter().map(|x| x * s).collect();
    Tensor::from_vec(data, a.shape.clone())
}

pub fn softmax(a: &Tensor, axis: usize) -> Tensor {
    #[cfg(target_os = "macos")]
    {
        if let Some(backend) = MetalBackend::new() {
            if let Some(result) = backend.softmax_gpu(a, axis) {
                return result;
            }
        }
    }

    let rank = a.shape.len();
    assert!(axis < rank);

    let dim = a.shape[axis];
    let outer: usize = a.shape[..axis].iter().product();
    let inner: usize = a.shape[axis + 1..].iter().product();
    let stride = dim * inner;

    let mut data = Vec::with_capacity(a.data.len());

    for o in 0..outer {
        for i in 0..inner {
            let base = o * stride + i;

            // Single pass: find max, compute exp, accumulate sum
            let mut max = a.data[base + i];
            for d in 0..dim {
                let val = a.data[base + d * inner];
                if val > max {
                    max = val;
                }
            }

            let mut sum = 0.0f32;
            let temp_offset = data.len();
            for d in 0..dim {
                let val = (a.data[base + d * inner] - max).exp();
                data.push(val);
                sum += val;
            }

            // Normalize the values we just added
            for d in 0..dim {
                data[temp_offset + d] /= sum;
            }
        }
    }
    Tensor::from_vec(data, a.shape.clone())
}

pub fn relu(a: &Tensor) -> Tensor {
    #[cfg(target_os = "macos")]
    {
        if let Some(backend) = MetalBackend::new() {
            if let Some(result) = backend.relu_gpu(a) {
                return result;
            }
        }
    }

    let data: Vec<f32> = a.data.iter().map(|x| x.max(0.0)).collect();
    Tensor::from_vec(data, a.shape.clone())
}

pub fn layer_norm(a: &Tensor, weight: &Tensor, bias: &Tensor, eps: f32) -> Tensor {
    #[cfg(target_os = "macos")]
    {
        if let Some(backend) = MetalBackend::new() {
            if let Some(result) = backend.layer_norm_gpu(a, weight, bias, eps) {
                return result;
            }
        }
    }

    assert!(!a.shape.is_empty());
    let last_dim = *a.shape.last().unwrap();
    let n_elements = a.data.len();
    let n_vectors = n_elements / last_dim;

    let mut data = Vec::with_capacity(n_elements);

    for i in 0..n_vectors {
        let base = i * last_dim;
        let slice = &a.data[base..base + last_dim];

        // Single-pass: compute mean
        let mean: f32 = slice.iter().sum::<f32>() / last_dim as f32;

        // Second pass: compute variance and apply normalization
        let var: f32 = slice.iter().map(|x| (x - mean).powi(2)).sum::<f32>() / last_dim as f32;
        let inv_std = 1.0 / (var + eps).sqrt();

        // Apply normalization in single loop
        for (j, &val) in slice.iter().enumerate() {
            data.push((val - mean) * inv_std * weight.data[j] + bias.data[j]);
        }
    }
    Tensor::from_vec(data, a.shape.clone())
}

pub fn linear(x: &Tensor, weight: &Tensor, bias: Option<&Tensor>) -> Tensor {
    let mut out = matmul(x, weight);
    if let Some(b) = bias {
        out = add(&out, b);
    }
    out
}