runmat-accelerate 0.4.4

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
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153

pub const KRON_SHADER_F64: &str = r#"
const MAX_RANK: u32 = 128u;

struct PackedValue {
    value: u32,
    _pad0: u32,
    _pad1: u32,
    _pad2: u32,
};

alias PackedArray = array<PackedValue, MAX_RANK>;

struct Tensor {
    data: array<f64>,
};

struct Params {
    len: u32,
    offset: u32,
    rank: u32,
    _pad: u32,
    shape_a: PackedArray,
    shape_b: PackedArray,
    shape_out: PackedArray,
    stride_a: PackedArray,
    stride_b: PackedArray,
};

@group(0) @binding(0) var<storage, read> A: Tensor;
@group(0) @binding(1) var<storage, read> B: Tensor;
@group(0) @binding(2) var<storage, read_write> Out: Tensor;
@group(0) @binding(3) var<uniform> params: Params;

@compute @workgroup_size(@WG@)
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
    let idx = gid.x;
    if idx >= params.len {
        return;
    }

    let global_index = params.offset + idx;
    var remaining = global_index;
    var index_a: u32 = 0u;
    var index_b: u32 = 0u;
    var dim: u32 = 0u;

    loop {
        if dim >= params.rank {
            break;
        }
        let out_dim = params.shape_out[dim].value;
        var coord_out: u32 = 0u;
        if out_dim != 0u {
            coord_out = remaining % out_dim;
            remaining = remaining / out_dim;
        }

        let dim_b = params.shape_b[dim].value;
        var coord_b: u32 = 0u;
        var coord_a: u32 = 0u;
        if dim_b != 0u {
            coord_b = coord_out % dim_b;
            coord_a = coord_out / dim_b;
        }

        let stride_a = params.stride_a[dim].value;
        let stride_b = params.stride_b[dim].value;
        index_a = index_a + coord_a * stride_a;
        index_b = index_b + coord_b * stride_b;

        dim = dim + 1u;
    }

    Out.data[global_index] = A.data[index_a] * B.data[index_b];
}
"#;

pub const KRON_SHADER_F32: &str = r#"
const MAX_RANK: u32 = 128u;

struct PackedValue {
    value: u32,
    _pad0: u32,
    _pad1: u32,
    _pad2: u32,
};

alias PackedArray = array<PackedValue, MAX_RANK>;

struct Tensor {
    data: array<f32>,
};

struct Params {
    len: u32,
    offset: u32,
    rank: u32,
    _pad: u32,
    shape_a: PackedArray,
    shape_b: PackedArray,
    shape_out: PackedArray,
    stride_a: PackedArray,
    stride_b: PackedArray,
};

@group(0) @binding(0) var<storage, read> A: Tensor;
@group(0) @binding(1) var<storage, read> B: Tensor;
@group(0) @binding(2) var<storage, read_write> Out: Tensor;
@group(0) @binding(3) var<uniform> params: Params;

@compute @workgroup_size(@WG@)
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
    let idx = gid.x;
    if idx >= params.len {
        return;
    }

    let global_index = params.offset + idx;
    var remaining = global_index;
    var index_a: u32 = 0u;
    var index_b: u32 = 0u;
    var dim: u32 = 0u;

    loop {
        if dim >= params.rank {
            break;
        }
        let out_dim = params.shape_out[dim].value;
        var coord_out: u32 = 0u;
        if out_dim != 0u {
            coord_out = remaining % out_dim;
            remaining = remaining / out_dim;
        }

        let dim_b = params.shape_b[dim].value;
        var coord_b: u32 = 0u;
        var coord_a: u32 = 0u;
        if dim_b != 0u {
            coord_b = coord_out % dim_b;
            coord_a = coord_out / dim_b;
        }

        let stride_a = params.stride_a[dim].value;
        let stride_b = params.stride_b[dim].value;
        index_a = index_a + coord_a * stride_a;
        index_b = index_b + coord_b * stride_b;

        dim = dim + 1u;
    }

    Out.data[global_index] = A.data[index_a] * B.data[index_b];
}
"#;