metaltile-std 0.1.0

MetalTile kernel standard library — benchmark metadata and type definitions
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

//! Copyright 2026 0xClandestine, Ekryski, TheTom, Ambisphaeric
//! SPDX-License-Identifier: Apache-2.0
//! Elementwise binary ops — #[kernel] DSL vs MLX metal/binary.metal

use metaltile::{bench_kernel, kernel};

#[bench_kernel(
    op="binary",
    subop="add",
    class=Binary,
    input_a=Unit,
    input_b=Half,
    tol=1e-6,
    mlx="vvn_Add{tn}",
    metal_file="binary.metal",
)]
#[kernel]
pub fn vector_add<T>(a: Tensor<T>, b: Tensor<T>, c: Tensor<T>) {
    let idx = program_id(0);
    store(c[idx], load(a[idx]) + load(b[idx]));
}

#[bench_kernel(
    op="binary",
    subop="mul",
    class=Binary,
    input_a=Unit,
    input_b=Half,
    tol=1e-6,
    mlx="vvn_Multiply{tn}",
    metal_file="binary.metal",
)]
#[kernel]
pub fn mt_mul<T>(a: Tensor<T>, b: Tensor<T>, out: Tensor<T>) {
    let idx = program_id(0);
    store(out[idx], load(a[idx]) * load(b[idx]));
}

#[bench_kernel(
    op="binary",
    subop="sub",
    class=Binary,
    input_a=Unit,
    input_b=Half,
    tol=1e-6,
    mlx="vvn_Subtract{tn}",
    metal_file="binary.metal",
)]
#[kernel]
pub fn mt_sub<T>(a: Tensor<T>, b: Tensor<T>, out: Tensor<T>) {
    let idx = program_id(0);
    store(out[idx], load(a[idx]) - load(b[idx]));
}

#[bench_kernel(
    op="binary",
    subop="div",
    class=Binary,
    input_a=Unit,
    input_b=Half,
    tol=1e-6,
    mlx="vvn_Divide{tn}",
    metal_file="binary.metal",
)]
#[kernel]
pub fn mt_div<T>(a: Tensor<T>, b: Tensor<T>, out: Tensor<T>) {
    let idx = program_id(0);
    store(out[idx], load(a[idx]) / load(b[idx]));
}

#[bench_kernel(
    op="binary",
    subop="maximum",
    class=Binary,
    input_a=Unit,
    input_b=Half,
    tol=1e-6,
    mlx="vvn_Maximum{tn}",
    metal_file="binary.metal",
)]
#[kernel]
pub fn mt_max_elem<T>(a: Tensor<T>, b: Tensor<T>, out: Tensor<T>) {
    let idx = program_id(0);
    store(out[idx], max(load(a[idx]), load(b[idx])));
}

#[bench_kernel(
    op="binary",
    subop="minimum",
    class=Binary,
    input_a=Unit,
    input_b=Half,
    tol=1e-6,
    mlx="vvn_Minimum{tn}",
    metal_file="binary.metal",
)]
#[kernel]
pub fn mt_min_elem<T>(a: Tensor<T>, b: Tensor<T>, out: Tensor<T>) {
    let idx = program_id(0);
    store(out[idx], min(load(a[idx]), load(b[idx])));
}

#[bench_kernel(
    op="binary",
    subop="pow",
    class=Binary,
    input_a=Unit,
    input_b=Half,
    tol=1e-4,
    mlx="vvn_Power{tn}",
    metal_file="binary.metal",
)]
#[kernel]
pub fn mt_pow<T>(a: Tensor<T>, b: Tensor<T>, out: Tensor<T>) {
    let idx = program_id(0);
    store(out[idx], pow(load(a[idx]), load(b[idx])));
}

#[bench_kernel(
    op="binary",
    subop="atan2",
    class=Binary,
    input_a=Signed,
    input_b=Half,
    // tol=1e-3 — f16 atan2 compounds two-input ULP drift (~3e-4).
    tol=1e-3,
    mlx="vvn_ArcTan2{tn}",
    metal_file="binary.metal",
)]
#[kernel]
pub fn mt_atan2<T>(y: Tensor<T>, x: Tensor<T>, out: Tensor<T>) {
    let idx = program_id(0);
    store(out[idx], atan2(load(y[idx]), load(x[idx])));
}

#[bench_kernel(
    op="binary",
    subop="remainder",
    class=Binary,
    input_a=Signed,
    input_b=Half,
    // tol=1e-4 — f16 fmod(dividend, divisor) stays within ULP of MLX.
    tol=1e-4,
    mlx="vvn_Remainder{tn}",
    metal_file="binary.metal",
)]
#[kernel]
pub fn mt_remainder<T>(a: Tensor<T>, b: Tensor<T>, out: Tensor<T>) {
    let idx = program_id(0);
    store(out[idx], remainder(load(a[idx]), load(b[idx])));
}

#[bench_kernel(
    op="binary",
    subop="logaddexp",
    class=Binary,
    input_a=Signed,
    input_b=Signed,
    // tol=1e-2 — f16 drifts ~1.2e-4 and bf16 drifts ~7.8e-3 on signed
    // input. `logaddexp(a, b) = log(exp(a) + exp(b))` compounds three
    // transcendentals so half/bfloat ULP drift accumulates.
    tol=1e-2,
    mlx="vvn_LogAddExp{tn}",
    metal_file="binary.metal",
)]
#[kernel]
pub fn mt_logaddexp<T>(a: Tensor<T>, b: Tensor<T>, out: Tensor<T>) {
    let idx = program_id(0);
    store(out[idx], log(exp(load(a[idx])) + exp(load(b[idx]))));
}