g-tensor 0.1.0

CPU-first tensor and ML library for Apple Silicon
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
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303

#![cfg(feature = "gpu")]

use g::{from_slice_f32, gpu_device_count, gpu_device_name, matmul, Result};

fn close(a: f32, b: f32) -> bool {
    // CPU and GPU accumulate the same dot products in different orders, so a
    // tiny relative error is expected at f32 magnitudes around 1e8.
    (a - b).abs() <= 1e-3 + 1e-5 * a.abs().max(b.abs())
}

#[test]
fn gpu_reports_metal_device() {
    assert!(gpu_device_count() >= 1);
    if let Some(name) = gpu_device_name() {
        assert!(!name.is_empty());
        println!("gpu device: {name}");
    }
}

#[test]
fn batched_rank3_times_rank2_matches_cpu_reference() -> Result<()> {
    // [2, 512, 512] @ [512, 512]. Sits exactly on the offload crossover so the
    // GPU path is exercised, and the identity LHS makes the expected output
    // trivial to verify without materializing a second matmul.
    let m = 512usize;
    let n = 512usize;
    let k = 512usize;
    let b = 2usize;

    let mut a = vec![0.0f32; b * m * k];
    for bi in 0..b {
        for i in 0..m {
            a[(bi * m + i) * k + i] = 1.0;
        }
    }
    let mut rhs = vec![0.0f32; k * n];
    for i in 0..k {
        for j in 0..n {
            rhs[i * n + j] = ((i * 7 + j * 3) % 997) as f32;
        }
    }

    let at = from_slice_f32(&a, &[b, m, k])?;
    let bt = from_slice_f32(&rhs, &[k, n])?;
    let y = matmul(&at, &bt)?;
    assert_eq!(y.shape(), &[b, m, n]);

    let got = y.to_vec_f32()?;
    for bi in 0..b {
        for (i, j) in [(0usize, 0usize), (1, 511), (511, 1), (511, 511)] {
            let expected = rhs[i * n + j];
            let actual = got[(bi * m + i) * n + j];
            assert!(
                close(actual, expected),
                "{bi} {i} {j}: {actual} != {expected}"
            );
        }
    }
    Ok(())
}

#[test]
fn multi_device_rank3_times_rank2_matches_single() -> Result<()> {
    let m = 512usize;
    let n = 512usize;
    let k = 512usize;
    let b = 2usize;

    let mut a = vec![0.0f32; b * m * k];
    for bi in 0..b {
        for i in 0..m {
            a[(bi * m + i) * k + i] = 1.0;
        }
    }
    let mut rhs = vec![0.0f32; k * n];
    for i in 0..k {
        for j in 0..n {
            rhs[i * n + j] = ((i * 7 + j * 3) % 997) as f32;
        }
    }

    let at = from_slice_f32(&a, &[b, m, k])?;
    let bt = from_slice_f32(&rhs, &[k, n])?;
    let single = matmul(&at, &bt)?;
    let multi = g::g_apple::matmul_multi_device(&at, &bt)?;

    let sv = single.to_vec_f32()?;
    let mv = multi.to_vec_f32()?;
    assert_eq!(sv.len(), mv.len());
    for (i, (s, m)) in sv.iter().zip(&mv).enumerate() {
        assert!(close(*s, *m), "{i}: {s} != {m}");
    }
    Ok(())
}

#[test]
fn cpu_gpu_batched_matches_single() -> Result<()> {
    let m = 512usize;
    let n = 512usize;
    let k = 512usize;
    let b = 6usize;

    let mut a = vec![0.0f32; b * m * k];
    for bi in 0..b {
        for i in 0..m {
            a[(bi * m + i) * k + i] = 1.0;
        }
    }
    let mut rhs = vec![0.0f32; k * n];
    for i in 0..k {
        for j in 0..n {
            rhs[i * n + j] = ((i * 7 + j * 3) % 997) as f32;
        }
    }

    let at = from_slice_f32(&a, &[b, m, k])?;
    let bt = from_slice_f32(&rhs, &[k, n])?;
    let cpu_gpu = matmul(&at, &bt)?;
    let single = g::g_apple::matmul(&at, &bt)?;

    let a = cpu_gpu.to_vec_f32()?;
    let c = single.to_vec_f32()?;
    assert_eq!(a.len(), c.len());
    for (i, (x, y)) in a.iter().zip(&c).enumerate() {
        assert!(close(*x, *y), "{i}: {x} != {y}");
    }
    Ok(())
}

#[test]
fn cpu_gpu_rank3_times_rank3_matches_single() -> Result<()> {
    let m = 512usize;
    let n = 512usize;
    let k = 512usize;
    let b = 6usize;

    let mut a = vec![0.0f32; b * m * k];
    let mut bmat = vec![0.0f32; b * k * n];
    for bi in 0..b {
        for i in 0..m {
            for j in 0..k {
                a[(bi * m + i) * k + j] = ((i * 3 + j * 5 + bi) % 991) as f32;
            }
        }
        for i in 0..k {
            for j in 0..n {
                bmat[(bi * k + i) * n + j] = ((i * 7 + j * 11 + bi) % 983) as f32;
            }
        }
    }

    let at = from_slice_f32(&a, &[b, m, k])?;
    let bt = from_slice_f32(&bmat, &[b, k, n])?;
    let cpu_gpu = matmul(&at, &bt)?;
    let single = g::g_apple::matmul(&at, &bt)?;

    let a = cpu_gpu.to_vec_f32()?;
    let c = single.to_vec_f32()?;
    assert_eq!(a.len(), c.len());
    for (i, (x, y)) in a.iter().zip(&c).enumerate() {
        assert!(close(*x, *y), "{i}: {x} != {y}");
    }
    Ok(())
}

#[test]
fn cpu_gpu_rank2_matches_single() -> Result<()> {
    let m = 1536usize;
    let n = 512usize;
    let k = 512usize;

    let mut a = vec![0.0f32; m * k];
    let mut b = vec![0.0f32; k * n];
    for i in 0..m {
        for j in 0..k {
            a[i * k + j] = ((i * 3 + j * 5) % 991) as f32;
        }
    }
    for i in 0..k {
        for j in 0..n {
            b[i * n + j] = ((i * 7 + j * 11) % 983) as f32;
        }
    }

    let at = from_slice_f32(&a, &[m, k])?;
    let bt = from_slice_f32(&b, &[k, n])?;
    let cpu_gpu = matmul(&at, &bt)?;
    let single = g::g_apple::matmul(&at, &bt)?;

    let a = cpu_gpu.to_vec_f32()?;
    let c = single.to_vec_f32()?;
    assert_eq!(a.len(), c.len());
    for (i, (x, y)) in a.iter().zip(&c).enumerate() {
        assert!(close(*x, *y), "{i}: {x} != {y}");
    }
    Ok(())
}

#[test]
fn multi_device_single_batch_rank3_matches_single() -> Result<()> {
    let m = 768usize;
    let n = 768usize;
    let k = 768usize;

    let mut a = vec![0.0f32; m * k];
    for i in 0..m {
        a[i * k + i] = 1.0;
    }
    let mut rhs = vec![0.0f32; k * n];
    for i in 0..k {
        for j in 0..n {
            rhs[i * n + j] = ((i * 7 + j * 3) % 997) as f32;
        }
    }
    let at = from_slice_f32(&a, &[1, m, k])?;
    let bt = from_slice_f32(&rhs, &[k, n])?;

    let single = g::g_apple::matmul(&at, &bt)?;
    let multi = g::g_apple::matmul_multi_device(&at, &bt)?;
    let sv = single.to_vec_f32()?;
    let mv = multi.to_vec_f32()?;
    assert_eq!(sv.len(), mv.len());
    for (i, (s, m)) in sv.iter().zip(&mv).enumerate() {
        assert!(close(*s, *m), "{i}: {s} != {m}");
    }
    Ok(())
}

#[test]
fn multi_device_rank2_matches_single() -> Result<()> {
    let m = 512usize;
    let n = 512usize;
    let k = 1024usize;

    let mut a = vec![0.0f32; m * k];
    let mut b = vec![0.0f32; k * n];
    for i in 0..m {
        for j in 0..k {
            a[i * k + j] = ((i * 3 + j * 5) % 991) as f32;
        }
    }
    for i in 0..k {
        for j in 0..n {
            b[i * n + j] = ((i * 7 + j * 11) % 983) as f32;
        }
    }

    let at = from_slice_f32(&a, &[m, k])?;
    let bt = from_slice_f32(&b, &[k, n])?;
    let single = matmul(&at, &bt)?;
    let multi = g::g_apple::matmul_multi_device(&at, &bt)?;

    let sv = single.to_vec_f32()?;
    let mv = multi.to_vec_f32()?;
    assert_eq!(sv.len(), mv.len());
    for (i, (s, m)) in sv.iter().zip(&mv).enumerate() {
        assert!(close(*s, *m), "{i}: {s} != {m}");
    }
    Ok(())
}

#[test]
fn batched_rank3_times_rank3_keeps_batch_pairing() -> Result<()> {
    // [2, 512, 512] @ [2, 512, 512]. A diagonal LHS scaled per batch means
    // batch bi of the output must only depend on batch bi of the RHS.
    let m = 512usize;
    let n = 512usize;
    let k = 512usize;
    let b = 2usize;

    let mut a = vec![0.0f32; b * m * k];
    let mut rhs = vec![0.0f32; b * k * n];
    for bi in 0..b {
        let scale = (bi + 1) as f32;
        for i in 0..m {
            a[(bi * m + i) * k + i] = scale;
        }
        for i in 0..k {
            for j in 0..n {
                rhs[(bi * k + i) * n + j] = ((i * 5 + j * 11 + bi * 3) % 991) as f32;
            }
        }
    }

    let at = from_slice_f32(&a, &[b, m, k])?;
    let bt = from_slice_f32(&rhs, &[b, k, n])?;
    let y = matmul(&at, &bt)?;
    assert_eq!(y.shape(), &[b, m, n]);

    let got = y.to_vec_f32()?;
    for bi in 0..b {
        let scale = (bi + 1) as f32;
        for (i, j) in [(0usize, 0usize), (7, 500), (500, 7), (511, 511)] {
            let expected = scale * rhs[(bi * k + i) * n + j];
            let actual = got[(bi * m + i) * n + j];
            assert!(
                close(actual, expected),
                "{bi} {i} {j}: {actual} != {expected}"
            );
        }
    }
    Ok(())
}