vkml 0.0.2

High-level Vulkan-based machine learning library
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

#[derive(Clone, Copy, Debug)]
struct AxisState {
    axis: usize,
    dim: u64,
    limit: u32,
    dispatches: u64,
    local_size: u32,
}

pub fn optimal_workgroup_size(
    max_workgroup_size: [u32; 3],
    max_workgroup_invocations: u32,
    dims: [Option<u64>; 3],
) -> [u32; 3] {
    let mut axes: Vec<AxisState> = dims
        .iter()
        .enumerate()
        .filter_map(|(axis, &maybe_dim)| {
            maybe_dim.filter(|&dim| dim > 0).map(|dim| AxisState {
                axis,
                dim,
                limit: max_workgroup_size[axis],
                dispatches: 1,
                local_size: 1,
            })
        })
        .collect();

    if axes.is_empty() {
        return [1, 1, 1];
    }

    let max_inv = max_workgroup_invocations.max(1);

    for axis in &mut axes {
        let limit = axis.limit.min(max_inv).max(1);
        axis.limit = limit;

        if axis.dim <= limit as u64 {
            axis.local_size = axis.dim as u32;
            axis.dispatches = 1;
        } else {
            axis.dispatches = div_ceil_u64(axis.dim, limit as u64);
            axis.local_size = div_ceil_u32(axis.dim, axis.dispatches).min(limit).max(1);
        }
    }

    let mut product = total_product(&axes);

    while product > max_inv as u64 {
        let mut best: Option<(usize, u64, u64, u32, u64)> = None;

        for (idx, axis) in axes.iter().enumerate() {
            if axis.local_size <= 1 {
                continue;
            }

            let mut next_dispatches = axis.dispatches + 1;
            let mut next_size = div_ceil_u32(axis.dim, next_dispatches)
                .min(axis.limit)
                .max(1);

            while next_size == axis.local_size && next_dispatches < axis.dim {
                next_dispatches += 1;
                next_size = div_ceil_u32(axis.dim, next_dispatches)
                    .min(axis.limit)
                    .max(1);
            }

            if next_size >= axis.local_size {
                continue;
            }

            let new_product = product / axis.local_size.max(1) as u64 * next_size.max(1) as u64;

            let new_waste = projected_waste(&axes, idx, next_size, next_dispatches);

            let candidate = (idx, new_product, next_dispatches, next_size, new_waste);
            match &best {
                None => best = Some(candidate),
                Some((_, best_product, _, best_size, best_waste)) => {
                    if new_product < *best_product
                        || (new_product == *best_product && new_waste < *best_waste)
                        || (new_product == *best_product
                            && new_waste == *best_waste
                            && next_size < *best_size)
                    {
                        best = Some(candidate);
                    }
                }
            }
        }

        let Some((idx, _, next_dispatches, next_size, _)) = best else {
            break;
        };

        axes[idx].dispatches = next_dispatches;
        axes[idx].local_size = next_size.min(axes[idx].limit).max(1);
        product = total_product(&axes);
    }

    let mut result = [1u32; 3];
    for axis in axes {
        result[axis.axis] = axis.local_size;
    }

    result
}

fn div_ceil_u64(value: u64, divisor: u64) -> u64 {
    if divisor == 0 {
        return 1;
    }
    value.div_ceil(divisor)
}

fn div_ceil_u32(value: u64, divisor: u64) -> u32 {
    div_ceil_u64(value, divisor).min(u32::MAX as u64) as u32
}

fn axis_waste(dim: u64, local_size: u64, dispatches: u64) -> u64 {
    dispatches.saturating_mul(local_size).saturating_sub(dim)
}

fn projected_waste(axes: &[AxisState], idx: usize, next_size: u32, next_dispatches: u64) -> u64 {
    axes.iter()
        .enumerate()
        .map(|(i, axis)| {
            if i == idx {
                axis_waste(axis.dim, next_size as u64, next_dispatches)
            } else {
                axis_waste(axis.dim, axis.local_size as u64, axis.dispatches)
            }
        })
        .sum()
}

fn total_product(axes: &[AxisState]) -> u64 {
    axes.iter().fold(1u64, |acc, axis| {
        acc.saturating_mul(axis.local_size.max(1) as u64)
    })
}