vkml 0.0.3

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

use crate::{instruction, tensor::TensorDesc, utils::error::VKMLError};

use super::{Layer, execution::LayerExecution};

#[derive(Clone)]
pub enum ElementWiseOperation {
    Add,
    Subtract,
    Multiply,
    Divide,
    Maximum,
    Minimum,
}

impl ElementWiseOperation {
    fn name(&self) -> String {
        match self {
            ElementWiseOperation::Add => "Add".to_string(),
            ElementWiseOperation::Subtract => "Subtract".to_string(),
            ElementWiseOperation::Multiply => "Multiply".to_string(),
            ElementWiseOperation::Divide => "Divide".to_string(),
            ElementWiseOperation::Maximum => "Maximum".to_string(),
            ElementWiseOperation::Minimum => "Minimum".to_string(),
        }
    }
}

#[derive(Clone)]
pub struct ElementWiseLayer {
    pub operation: ElementWiseOperation,
}

impl ElementWiseLayer {
    pub fn new(operation: ElementWiseOperation) -> Self {
        Self { operation }
    }
}

impl Layer for ElementWiseLayer {
    fn output_shapes(
        &self,
        _batch_size: i64,
        input_shapes: &[&TensorDesc],
    ) -> Result<Vec<TensorDesc>, VKMLError> {
        if input_shapes.len() < 2 {
            return Err(VKMLError::Layer(format!(
                "Element-wise operation requires at least 2 inputs, got {}",
                input_shapes.len()
            )));
        }

        // All inputs must have the same shape
        let first_shape = input_shapes[0];
        for shape in &input_shapes[1..] {
            if shape.dims() != first_shape.dims() {
                return Err(VKMLError::Layer(format!(
                    "Element-wise operations require matching dimensions: {:?} vs {:?}",
                    first_shape.dims(),
                    shape.dims()
                )));
            }
        }

        // Output has the same shape as inputs
        Ok(vec![first_shape.clone()])
    }

    fn input_requirements(&self) -> (usize, Option<usize>) {
        (2, None) // At least 2 inputs, no upper limit
    }

    fn name(&self) -> String {
        self.operation.name()
    }

    fn build_layer_exec(
        &self,
        _batch_size: i64,
        input_shapes: &[&TensorDesc],
    ) -> Result<LayerExecution, VKMLError> {
        if input_shapes.len() < 2 {
            return Err(VKMLError::Layer(format!(
                "Element-wise operation requires at least 2 inputs, got {}",
                input_shapes.len()
            )));
        }

        // Check that all inputs have the same shape
        let first_shape = input_shapes[0];
        for shape in &input_shapes[1..] {
            if shape.dims() != first_shape.dims() {
                return Err(VKMLError::Layer(format!(
                    "Element-wise operations require matching dimensions: {:?} vs {:?}",
                    first_shape.dims(),
                    shape.dims()
                )));
            }
        }

        let tensors = vec![
            first_shape.clone(),
            first_shape.clone(),
            first_shape.clone(),
        ];

        // Create element-wise instruction
        let element_wise_instruction = match self.operation {
            ElementWiseOperation::Add => instruction::add(0, 1, 2),
            ElementWiseOperation::Subtract => instruction::sub(0, 1, 2),
            ElementWiseOperation::Multiply => instruction::mul(0, 1, 2),
            ElementWiseOperation::Divide => instruction::div(0, 1, 2),
            ElementWiseOperation::Maximum => instruction::max(0, 1, 2),
            ElementWiseOperation::Minimum => instruction::min(0, 1, 2),
        };

        // Get input mappings using the trait method
        // Note: For element-wise ops, we only use the first two inputs regardless of how many are provided
        let input_mappings = self.map_input_tensors(input_shapes.len().min(2));

        Ok(LayerExecution {
            tensors,
            instructions: vec![element_wise_instruction],
            outputs: vec![2],
            input_mappings,
        })
    }
}