tenferro-gpu 0.2.0

CubeCL-backed CUDA and WebGPU provider backends for tenferro tensors.
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

use cubecl::prelude::*;

#[cube(launch_unchecked)]
pub fn pack_lhs_dot_general<E: CubePrimitive>(
    out: &mut Tensor<E>,
    input: &Tensor<E>,
    #[comptime] free_axes: Sequence<usize>,
    #[comptime] contract_axes: Sequence<usize>,
    #[comptime] batch_axes: Sequence<usize>,
    #[comptime] input_rank: usize,
    #[comptime] out_rank: usize,
) {
    if ABSOLUTE_POS < out.len() {
        let mut input_idx = Array::<usize>::new(input_rank);

        #[unroll]
        for batch_pos in 0..batch_axes.len() {
            let input_axis = comptime! { *batch_axes.index(batch_pos) };
            input_idx[input_axis] = out.coordinate(ABSOLUTE_POS, batch_pos);
        }

        let mut free_flat = out.coordinate(ABSOLUTE_POS, out_rank - 2);
        #[unroll]
        for pos in 0..free_axes.len() {
            let input_axis = comptime! { *free_axes.index(pos) };
            let dim = input.shape(input_axis);
            input_idx[input_axis] = free_flat % dim;
            free_flat /= dim;
        }

        let mut contract_flat = out.coordinate(ABSOLUTE_POS, out_rank - 1);
        #[unroll]
        for pos in 0..contract_axes.len() {
            let input_axis = comptime! { *contract_axes.index(pos) };
            let dim = input.shape(input_axis);
            input_idx[input_axis] = contract_flat % dim;
            contract_flat /= dim;
        }

        let mut input_offset = 0usize;
        #[unroll]
        for axis in 0..input_rank {
            input_offset += input_idx[axis] * input.stride(axis);
        }
        out[ABSOLUTE_POS] = input[input_offset];
    }
}

#[cube(launch_unchecked)]
pub fn pack_rhs_dot_general<E: CubePrimitive>(
    out: &mut Tensor<E>,
    input: &Tensor<E>,
    #[comptime] contract_axes: Sequence<usize>,
    #[comptime] free_axes: Sequence<usize>,
    #[comptime] batch_axes: Sequence<usize>,
    #[comptime] input_rank: usize,
    #[comptime] out_rank: usize,
) {
    if ABSOLUTE_POS < out.len() {
        let mut input_idx = Array::<usize>::new(input_rank);

        #[unroll]
        for batch_pos in 0..batch_axes.len() {
            let input_axis = comptime! { *batch_axes.index(batch_pos) };
            input_idx[input_axis] = out.coordinate(ABSOLUTE_POS, batch_pos);
        }

        let mut contract_flat = out.coordinate(ABSOLUTE_POS, out_rank - 2);
        #[unroll]
        for pos in 0..contract_axes.len() {
            let input_axis = comptime! { *contract_axes.index(pos) };
            let dim = input.shape(input_axis);
            input_idx[input_axis] = contract_flat % dim;
            contract_flat /= dim;
        }

        let mut free_flat = out.coordinate(ABSOLUTE_POS, out_rank - 1);
        #[unroll]
        for pos in 0..free_axes.len() {
            let input_axis = comptime! { *free_axes.index(pos) };
            let dim = input.shape(input_axis);
            input_idx[input_axis] = free_flat % dim;
            free_flat /= dim;
        }

        let mut input_offset = 0usize;
        #[unroll]
        for axis in 0..input_rank {
            input_offset += input_idx[axis] * input.stride(axis);
        }
        out[ABSOLUTE_POS] = input[input_offset];
    }
}