any-gpu 0.1.0

Tensor engine for every GPU. AMD, NVIDIA, Intel, Apple. One codebase, zero vendor lock-in. wgpu under the hood.
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
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

// Unlicense — cochranblock.org
// Contributors: GotEmCoach, KOVA, Claude Opus 4.6
//
// Tensor manipulation: concat, transpose.

use crate::device::{GpuBuffer, GpuDevice};
use anyhow::{ensure, Result};

// --- Concat ---

#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct ConcatParams {
    n: u32,
    outer: u32,
    a_inner: u32,
    b_inner: u32,
}

const SHADER_CONCAT: &str = "
struct P { n: u32, outer: u32, a_inner: u32, b_inner: u32, }
@group(0) @binding(0) var<uniform> p: P;
@group(0) @binding(1) var<storage, read> a: array<f32>;
@group(0) @binding(2) var<storage, read> b: array<f32>;
@group(0) @binding(3) var<storage, read_write> out: array<f32>;
@compute @workgroup_size(256)
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
    let idx = gid.x + gid.y * 65535u * 256u;
    if idx >= p.n { return; }

    let combined = p.a_inner + p.b_inner;
    let outer_idx = idx / combined;
    let inner_idx = idx % combined;

    if inner_idx < p.a_inner {
        out[idx] = a[outer_idx * p.a_inner + inner_idx];
    } else {
        out[idx] = b[outer_idx * p.b_inner + (inner_idx - p.a_inner)];
    }
}
";

// --- Transpose (swap two dims) ---

#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct TransposeParams {
    n: u32,
    d0: u32,
    d1: u32,
    inner: u32,
    outer_stride: u32,
    _pad: [u32; 3],
}

const SHADER_TRANSPOSE: &str = "
struct P { n: u32, d0: u32, d1: u32, inner: u32, outer_stride: u32, _p0: u32, _p1: u32, _p2: u32, }
@group(0) @binding(0) var<uniform> p: P;
@group(0) @binding(1) var<storage, read> a: array<f32>;
@group(0) @binding(2) var<storage, read_write> out: array<f32>;
@compute @workgroup_size(256)
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
    let idx = gid.x + gid.y * 65535u * 256u;
    if idx >= p.n { return; }

    // Decompose output index: outer * (d1 * d0 * inner) + i1 * (d0 * inner) + i0 * inner + inner_idx
    let block = p.d0 * p.d1 * p.inner;
    let outer = idx / block;
    let rem = idx % block;
    let i1 = rem / (p.d0 * p.inner);
    let rem2 = rem % (p.d0 * p.inner);
    let i0 = rem2 / p.inner;
    let inner_idx = rem2 % p.inner;

    // In input, dims are (d0, d1) so source index swaps i0 and i1
    let src = outer * block + i0 * (p.d1 * p.inner) + i1 * p.inner + inner_idx;
    out[idx] = a[src];
}
";

impl GpuDevice {
    /// Concat two buffers along a given axis.
    /// `outer_size` = product of dims before concat axis.
    /// `a_inner` = a's size along concat axis * product of dims after.
    /// `b_inner` = same for b.
    pub fn concat(
        &self,
        a: &GpuBuffer, b: &GpuBuffer,
        outer_size: u32, a_inner: u32, b_inner: u32,
    ) -> Result<GpuBuffer> {
        ensure!(a.len == (outer_size * a_inner) as usize);
        ensure!(b.len == (outer_size * b_inner) as usize);
        let total = outer_size * (a_inner + b_inner);
        let out = self.alloc(total as usize);
        let params = ConcatParams { n: total, outer: outer_size, a_inner, b_inner };
        self.dispatch_shader(
            SHADER_CONCAT, Some("concat"),
            &params, &[a, b], &out,
            super::dispatch_1d(total),
        );
        Ok(out)
    }

    /// Transpose two dimensions of a tensor.
    /// Shape is [..., d0, d1, ...inner_dims].
    /// `outer_size` = product of dims before d0.
    /// `inner` = product of dims after d1.
    pub fn transpose(
        &self,
        a: &GpuBuffer,
        outer_size: u32, d0: u32, d1: u32, inner: u32,
    ) -> Result<GpuBuffer> {
        let total = outer_size * d0 * d1 * inner;
        ensure!(a.len == total as usize);
        let out = self.alloc(total as usize);
        let params = TransposeParams {
            n: total, d0, d1, inner,
            outer_stride: d0 * d1 * inner, _pad: [0; 3],
        };
        self.dispatch_shader(
            SHADER_TRANSPOSE, Some("transpose"),
            &params, &[a], &out,
            super::dispatch_1d(total),
        );
        Ok(out)
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    fn dev() -> &'static GpuDevice { &crate::ops::TEST_DEV }

    #[test]
    fn test_concat_flat() {
        let result = dev().read(&dev().concat(&dev().upload(&[1.0, 2.0, 3.0]), &dev().upload(&[4.0, 5.0, 6.0]), 1, 3, 3).unwrap()).unwrap();
        assert_eq!(result, vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0]);
    }

    #[test]
    fn test_concat_asymmetric() {
        // Different sized inner dims: a has 2 elements, b has 3 per outer block
        let a = dev().upload(&[1.0, 2.0]);
        let b = dev().upload(&[3.0, 4.0, 5.0]);
        let result = dev().read(&dev().concat(&a, &b, 1, 2, 3).unwrap()).unwrap();
        assert_eq!(result, vec![1.0, 2.0, 3.0, 4.0, 5.0]);
    }

    #[test]
    fn test_concat_batched_channel_axis() {
        // batch=2, concat 1-channel and 2-channel tensors along C, spatial=2
        // a: [batch=2, c=1, spatial=2] = [10, 20, 30, 40]
        // b: [batch=2, c=2, spatial=2] = [1, 2, 3, 4, 5, 6, 7, 8]
        let a = dev().upload(&[10.0, 20.0, 30.0, 40.0]);
        let b = dev().upload(&[1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0]);
        // outer=batch=2, a_inner=1*2=2, b_inner=2*2=4
        let result = dev().read(&dev().concat(&a, &b, 2, 2, 4).unwrap()).unwrap();
        // batch 0: [10,20, 1,2,3,4], batch 1: [30,40, 5,6,7,8]
        assert_eq!(result, vec![10.0, 20.0, 1.0, 2.0, 3.0, 4.0, 30.0, 40.0, 5.0, 6.0, 7.0, 8.0]);
    }

    #[test]
    fn test_transpose_2d() {
        let a = dev().upload(&[1.0, 2.0, 3.0, 4.0, 5.0, 6.0]);
        let result = dev().read(&dev().transpose(&a, 1, 2, 3, 1).unwrap()).unwrap();
        assert_eq!(result, vec![1.0, 4.0, 2.0, 5.0, 3.0, 6.0]);
    }

    #[test]
    fn test_transpose_square() {
        // 3x3 -> 3x3 transpose
        let a = dev().upload(&[1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0]);
        let result = dev().read(&dev().transpose(&a, 1, 3, 3, 1).unwrap()).unwrap();
        assert_eq!(result, vec![1.0, 4.0, 7.0, 2.0, 5.0, 8.0, 3.0, 6.0, 9.0]);
    }

    #[test]
    fn test_transpose_batched() {
        let a = dev().upload(&[
            1.0, 2.0, 3.0, 4.0, 5.0, 6.0,
            7.0, 8.0, 9.0, 10.0, 11.0, 12.0,
        ]);
        let result = dev().read(&dev().transpose(&a, 2, 2, 3, 1).unwrap()).unwrap();
        assert_eq!(result, vec![
            1.0, 4.0, 2.0, 5.0, 3.0, 6.0,
            7.0, 10.0, 8.0, 11.0, 9.0, 12.0,
        ]);
    }

    #[test]
    fn test_transpose_1x_n() {
        // 1xN transpose = Nx1 (column vector)
        let a = dev().upload(&[1.0, 2.0, 3.0, 4.0, 5.0]);
        let result = dev().read(&dev().transpose(&a, 1, 1, 5, 1).unwrap()).unwrap();
        // 5x1 is same flat data (no-op for 1-row)
        assert_eq!(result, vec![1.0, 2.0, 3.0, 4.0, 5.0]);
    }

    #[test]
    fn test_transpose_roundtrip() {
        // Transpose twice = identity
        let data: Vec<f32> = (0..20).map(|i| i as f32).collect(); // 4x5
        let t1 = dev().transpose(&dev().upload(&data), 1, 4, 5, 1).unwrap();
        let t2 = dev().transpose(&t1, 1, 5, 4, 1).unwrap();
        let result = dev().read(&t2).unwrap();
        assert_eq!(result, data);
    }
}