knok-compile 0.3.0

MLIR lowering and IREE compilation for knok
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

use knok_core::{Conv2dOptions, TensorType};

use super::lowerer::{Lowerer, Value};

impl Lowerer<'_, '_> {
    pub(super) fn conv2d(
        &mut self,
        input: Value,
        kernel: Value,
        options: &Conv2dOptions,
    ) -> anyhow::Result<Value> {
        let input = self.pad_conv2d_input(input, options)?;
        if options.groups > 1 {
            return self.grouped_conv2d(input, kernel, options);
        }
        self.conv2d_direct(input, kernel, options)
    }

    fn conv2d_direct(
        &mut self,
        input: Value,
        kernel: Value,
        options: &Conv2dOptions,
    ) -> anyhow::Result<Value> {
        let ty = TensorType {
            elem: input.ty.elem,
            shape: vec![
                input.ty.shape[0],
                (input.ty.shape[1] - ((kernel.ty.shape[0] - 1) * options.dilation[0] + 1))
                    / options.stride[0]
                    + 1,
                (input.ty.shape[2] - ((kernel.ty.shape[1] - 1) * options.dilation[1] + 1))
                    / options.stride[1]
                    + 1,
                kernel.ty.shape[3],
            ],
        };
        let init = self.zero_initialized_tensor(&ty)?;
        let attrs = conv2d_attrs(options);
        self.append_named_linalg(
            "linalg.conv_2d_nhwc_hwcf",
            &[input, kernel],
            init,
            &ty,
            &attrs,
        )
    }

    fn grouped_conv2d(
        &mut self,
        input: Value,
        kernel: Value,
        options: &Conv2dOptions,
    ) -> anyhow::Result<Value> {
        let groups = options.groups;
        let input_channels_per_group = input.ty.shape[3] / groups;
        let output_channels_per_group = kernel.ty.shape[3] / groups;
        let mut output = None;
        for group in 0..groups {
            let input_slice_ty = TensorType {
                elem: input.ty.elem,
                shape: vec![
                    input.ty.shape[0],
                    input.ty.shape[1],
                    input.ty.shape[2],
                    input_channels_per_group,
                ],
            };
            let kernel_slice_ty = TensorType {
                elem: kernel.ty.elem,
                shape: vec![
                    kernel.ty.shape[0],
                    kernel.ty.shape[1],
                    kernel.ty.shape[2],
                    output_channels_per_group,
                ],
            };
            let input_slice = self.slice(
                input.clone(),
                &input_slice_ty,
                &[0, 0, 0, group * input_channels_per_group],
            )?;
            let kernel_slice = self.slice(
                kernel.clone(),
                &kernel_slice_ty,
                &[0, 0, 0, group * output_channels_per_group],
            )?;
            let group_output = self.conv2d_direct(input_slice, kernel_slice, options)?;
            output = Some(match output {
                Some(previous) => self.concat(previous, group_output, 3)?,
                None => group_output,
            });
        }
        output.ok_or_else(|| anyhow::anyhow!("conv2d groups must be non-zero"))
    }

    fn pad_conv2d_input(&mut self, input: Value, options: &Conv2dOptions) -> anyhow::Result<Value> {
        let padding = options.padding;
        if padding.top == 0 && padding.bottom == 0 && padding.left == 0 && padding.right == 0 {
            return Ok(input);
        }
        let ty = TensorType {
            elem: input.ty.elem,
            shape: vec![
                input.ty.shape[0],
                input.ty.shape[1] + padding.top + padding.bottom,
                input.ty.shape[2] + padding.left + padding.right,
                input.ty.shape[3],
            ],
        };
        let zero = self.constant(ty.elem.zero_literal(), ty.elem)?;
        self.append_tensor_pad(
            input,
            &ty,
            &[0, padding.top, padding.left, 0],
            &[0, padding.bottom, padding.right, 0],
            zero,
        )
    }
}

fn conv2d_attrs(options: &Conv2dOptions) -> Vec<(String, String)> {
    if options.dilation == [1, 1] && options.stride == [1, 1] {
        Vec::new()
    } else {
        vec![
            (
                "dilations".to_string(),
                format!(
                    "dense<[{}, {}]> : vector<2xi64>",
                    options.dilation[0], options.dilation[1]
                ),
            ),
            (
                "strides".to_string(),
                format!(
                    "dense<[{}, {}]> : vector<2xi64>",
                    options.stride[0], options.stride[1]
                ),
            ),
        ]
    }
}