runmat-accelerate 0.4.1

Pluggable GPU acceleration layer for RunMat (CUDA, ROCm, Metal, Vulkan/Spir-V)
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

use std::fmt::Write;

pub fn build_ind2sub_shader(
    scalar_ty: &str,
    dims: &[u32],
    strides: &[u32],
    total: u32,
    workgroup_size: u32,
    epsilon: &str,
) -> String {
    assert_eq!(dims.len(), strides.len());

    let mut shader = String::new();
    writeln!(shader, "struct Tensor {{ data: array<{scalar_ty}>, }};").unwrap();
    writeln!(
        shader,
        "struct ErrorState {{ code: u32, dim: u32, extra: u32, pad: u32 }};"
    )
    .unwrap();
    writeln!(
        shader,
        "struct Params {{ len: u32, _pad0: u32, _pad1: u32, _pad2: u32, }};"
    )
    .unwrap();
    writeln!(
        shader,
        "@group(0) @binding(0) var<storage, read> input: Tensor;"
    )
    .unwrap();
    for (idx, _) in dims.iter().enumerate() {
        writeln!(
            shader,
            "@group(0) @binding({}) var<storage, read_write> output{}: Tensor;",
            idx + 1,
            idx
        )
        .unwrap();
    }
    let error_binding = dims.len() + 1;
    writeln!(
        shader,
        "@group(0) @binding({}) var<storage, read_write> error: ErrorState;",
        error_binding
    )
    .unwrap();
    let params_binding = dims.len() + 2;
    writeln!(
        shader,
        "@group(0) @binding({}) var<uniform> params: Params;",
        params_binding
    )
    .unwrap();
    writeln!(shader, "const EPSILON: {scalar_ty} = {epsilon};").unwrap();
    writeln!(shader, "const ONE: {scalar_ty} = {scalar_ty}(1.0);").unwrap();
    writeln!(shader, "const TOTAL_F: {scalar_ty} = {scalar_ty}({total});").unwrap();
    writeln!(shader, "const TOTAL: u32 = {total}u;").unwrap();
    // Finite check helper using x==x (not NaN) and a large bound.
    let max_val = if scalar_ty == "f32" {
        "3.4028234663852886e38"
    } else {
        "1.7976931348623157e308"
    };
    writeln!(
        shader,
        "fn isfinite_scalar(x: {scalar_ty}) -> bool {{ return (x == x) && (abs(x) < {scalar_ty}({max_val})); }}"
    )
    .unwrap();
    for (idx, dim) in dims.iter().enumerate() {
        writeln!(shader, "const DIM_{idx}: u32 = {dim}u;").unwrap();
    }
    for (idx, stride) in strides.iter().enumerate() {
        writeln!(shader, "const STRIDE_{idx}: u32 = {stride}u;").unwrap();
    }
    writeln!(
        shader,
        "fn set_error(code: u32, extra: u32) {{
    if (error.code != 0u) {{ return; }}
    error.code = code;
    error.dim = 0u;
    error.extra = extra;
}}"
    )
    .unwrap();
    writeln!(
        shader,
        "@compute @workgroup_size({workgroup_size}, 1, 1)
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {{
    let idx = gid.x;
    if (idx >= params.len) {{
        return;
    }}
    if (error.code != 0u) {{
        return;
    }}
    let raw: {scalar_ty} = input.data[idx];
    if (!isfinite_scalar(raw)) {{
        set_error(1u, 0u);
        return;
    }}
    let rounded: {scalar_ty} = round(raw);
    if (abs(rounded - raw) > EPSILON) {{
        set_error(2u, 0u);
        return;
    }}
    if (rounded < ONE) {{
        set_error(3u, 0u);
        return;
    }}
    if (rounded > TOTAL_F) {{
        set_error(4u, 0u);
        return;
    }}
    let rounded_u32: u32 = u32(rounded);
    if (rounded_u32 == 0u) {{
        set_error(3u, 0u);
        return;
    }}
    let zero_based: u32 = rounded_u32 - 1u;
"
    )
    .unwrap();

    for (idx, _) in dims.iter().enumerate() {
        writeln!(
            shader,
            "    let coord_{idx}: u32 = (zero_based / STRIDE_{idx}) % DIM_{idx};
    output{idx}.data[idx] = {scalar_ty}(coord_{idx} + 1u);"
        )
        .unwrap();
    }

    writeln!(shader, "}}").unwrap();

    shader
}