tl-gpu 0.3.8

GPU tensor compute for ThinkingLanguage via wgpu
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

// WGSL compute shaders for GPU tensor operations

/// Elementwise binary operation: add(0), sub(1), mul(2), div(3)
pub const ELEMENTWISE_BINARY: &str = r#"
@group(0) @binding(0) var<storage, read> a: array<f32>;
@group(0) @binding(1) var<storage, read> b: array<f32>;
@group(0) @binding(2) var<storage, read_write> result: array<f32>;

struct Params {
    len: u32,
    op: u32,
}
@group(0) @binding(3) var<uniform> params: Params;

@compute @workgroup_size(256)
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
    let idx = gid.x;
    if (idx >= params.len) {
        return;
    }
    let va = a[idx];
    let vb = b[idx];
    switch (params.op) {
        case 0u: { result[idx] = va + vb; }
        case 1u: { result[idx] = va - vb; }
        case 2u: { result[idx] = va * vb; }
        case 3u: { result[idx] = va / vb; }
        default: { result[idx] = 0.0; }
    }
}
"#;

/// Scalar multiplication: scale all elements by a scalar
pub const SCALAR_MUL: &str = r#"
@group(0) @binding(0) var<storage, read> input: array<f32>;
@group(0) @binding(1) var<storage, read_write> result: array<f32>;

struct Params {
    len: u32,
    scalar: f32,
}
@group(0) @binding(2) var<uniform> params: Params;

@compute @workgroup_size(256)
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
    let idx = gid.x;
    if (idx >= params.len) {
        return;
    }
    result[idx] = input[idx] * params.scalar;
}
"#;

/// Parallel sum reduction using shared memory
pub const REDUCE_SUM: &str = r#"
@group(0) @binding(0) var<storage, read> input: array<f32>;
@group(0) @binding(1) var<storage, read_write> result: array<f32>;

struct Params {
    len: u32,
}
@group(0) @binding(2) var<uniform> params: Params;

var<workgroup> shared: array<f32, 256>;

@compute @workgroup_size(256)
fn main(
    @builtin(local_invocation_id) lid: vec3<u32>,
    @builtin(global_invocation_id) gid: vec3<u32>,
    @builtin(workgroup_id) wid: vec3<u32>,
) {
    let idx = gid.x;
    if (idx < params.len) {
        shared[lid.x] = input[idx];
    } else {
        shared[lid.x] = 0.0;
    }
    workgroupBarrier();

    // Tree reduction
    var stride: u32 = 128u;
    loop {
        if (stride == 0u) {
            break;
        }
        if (lid.x < stride) {
            shared[lid.x] = shared[lid.x] + shared[lid.x + stride];
        }
        workgroupBarrier();
        stride = stride >> 1u;
    }

    if (lid.x == 0u) {
        result[wid.x] = shared[0];
    }
}
"#;

/// Naive 2D matrix multiplication: C = A * B
/// A is [M, K], B is [K, N], C is [M, N]
pub const MATMUL: &str = r#"
@group(0) @binding(0) var<storage, read> a: array<f32>;
@group(0) @binding(1) var<storage, read> b: array<f32>;
@group(0) @binding(2) var<storage, read_write> result: array<f32>;

struct Params {
    m: u32,
    k: u32,
    n: u32,
}
@group(0) @binding(3) var<uniform> params: Params;

@compute @workgroup_size(16, 16)
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
    let row = gid.y;
    let col = gid.x;
    if (row >= params.m || col >= params.n) {
        return;
    }
    var sum: f32 = 0.0;
    for (var i: u32 = 0u; i < params.k; i = i + 1u) {
        sum = sum + a[row * params.k + i] * b[i * params.n + col];
    }
    result[row * params.n + col] = sum;
}
"#;

/// 2D matrix transpose
pub const TRANSPOSE: &str = r#"
@group(0) @binding(0) var<storage, read> input: array<f32>;
@group(0) @binding(1) var<storage, read_write> result: array<f32>;

struct Params {
    rows: u32,
    cols: u32,
}
@group(0) @binding(2) var<uniform> params: Params;

@compute @workgroup_size(16, 16)
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
    let row = gid.y;
    let col = gid.x;
    if (row >= params.rows || col >= params.cols) {
        return;
    }
    result[col * params.rows + row] = input[row * params.cols + col];
}
"#;