cubek-pool 0.3.0

CubeK: Pool
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

use super::{
    super::{
        accumulator_dtype, adaptive_end_index as end_index, adaptive_start_index as start_index,
        decompose_linear, shape_divmod,
    },
    pool2d::{Position, view4d},
};
use crate::definition::{AdaptiveAvgPoolOptions, PoolError};
use cubecl::{
    CubeDim, calculate_cube_count_elemwise,
    num_traits::Zero,
    prelude::{TensorBinding, *},
    std::{FastDivmod, tensor::ViewMut},
    tensor_vector_size_parallel,
};

/// One over the sum and its tap: the read is memory bound, and wider vectors measured slower.
const LIVE_VECTORS: usize = 3;
#[cube(launch, address_type = "dynamic")]
fn adaptive_avg_pool2d_direct<E: Numeric, EA: Numeric, N: Size>(
    input: &Tensor<Vector<E, N>>,
    mut output: ViewMut<'_, Vector<E, N>, Position>,
    out_shape: Sequence<FastDivmod<usize>>,
    working_units: usize,
    #[define(E)] _dtype: ElemType,
    #[define(EA)] _acc_dtype: ElemType,
) {
    if ABSOLUTE_POS >= working_units {
        terminate!();
    }

    let (b, oh, ow, c) = decompose_linear(ABSOLUTE_POS * output.vector_size(), &out_shape);

    let (_, out_h, out_w, _) = output.shape();
    let (in_stride_h, in_stride_w) = (input.stride(1), input.stride(2));
    let (in_h, in_w) = (input.shape(1), input.shape(2));

    let ih_start = start_index(oh, out_h, in_h);
    let ih_end = end_index(oh, out_h, in_h);

    let iw_start = start_index(ow, out_w, in_w);
    let iw_end = end_index(ow, out_w, in_w);

    let mut sum = Vector::<EA, N>::zero();

    let index_input_base = b * input.stride(0) + c * input.stride(3);

    for ih in ih_start..ih_end {
        let index_input_2 = ih * in_stride_h;

        for iw in iw_start..iw_end {
            let index_input_3 = iw * in_stride_w;

            let index_input = index_input_base + index_input_2 + index_input_3;
            sum += Vector::cast_from(input[index_input / input.vector_size()]);
        }
    }

    let num_ih = ih_end - ih_start;
    let num_iw = iw_end - iw_start;

    let average = sum / Vector::cast_from(num_ih * num_iw);
    output.write((b, oh, ow, c), Vector::cast_from(average));
}

pub(crate) fn adaptive_avg_pool2d_launch(
    client: &Client,
    input: TensorBinding,
    output: TensorBinding,
    _options: AdaptiveAvgPoolOptions<2>,
    dtype: ElemType,
) -> Result<(), PoolError> {
    let acc_dtype = accumulator_dtype(dtype);
    let vector_size = tensor_vector_size_parallel(
        client
            .properties()
            .vector_sizes_in_registers(acc_dtype.size(), LIVE_VECTORS),
        &input.shape,
        &input.strides,
        input.shape.len() - 1,
    );

    let working_units = output.shape.iter().product::<usize>() / vector_size as usize;
    let cube_dim = CubeDim::new(client, working_units);
    let cube_count = calculate_cube_count_elemwise(client, working_units, cube_dim);

    let address_type = input
        .required_address_type(dtype.size())
        .max(output.required_address_type(dtype.size()));

    adaptive_avg_pool2d_direct::launch(
        client,
        cube_count,
        cube_dim,
        address_type,
        vector_size,
        input.into_tensor_arg(),
        view4d(output.clone(), vector_size),
        shape_divmod(&output),
        working_units,
        dtype,
        acc_dtype,
    );

    Ok(())
}