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
144
145
146
147
148
149
150
151
152
153

use itertools::Itertools;

use kn_cuda_sys::wrapper::handle::{CudaDevice, CudaStream};
use kn_cuda_sys::wrapper::rtc::args::KernelArgs;
use kn_cuda_sys::wrapper::rtc::core::{CuFunction, Dim3};
use kn_cuda_sys::wrapper::status::Status;

use crate::autokernel::common::{
    c_array_string, c_nested_array_string, ceil_div, compile_cached_kernel, fill_replacements, KernelKey,
};
use crate::device_tensor::DeviceTensor;
use crate::shape::StridedShape;

#[derive(Debug)]
pub struct ReduceKernel {
    _code: ReduceCode,
    _reduced_axes: Vec<usize>,

    function: CuFunction,

    input_shape: StridedShape,
    output_shape: StridedShape,
}

const REDUCE_SOURCE: &str = include_str!("reduce.cu");

#[derive(Debug, Clone)]
pub struct ReduceCode {
    pub ty: String,
    pub identity: String,
    pub operation: String,
    pub post_process: String,
}

impl ReduceKernel {
    pub fn new(
        device: CudaDevice,
        code: ReduceCode,
        input_shape: &StridedShape,
        output_shape: &StridedShape,
        reduced_axes: &[usize],
    ) -> Self {
        // check that axes are unique and in-bounds
        assert!(
            reduced_axes.iter().all_unique(),
            "Reduced axes must be unique, got {:?}",
            reduced_axes
        );
        for &axis in reduced_axes {
            assert!(
                axis < input_shape.rank(),
                "Reduced axis out of bounds for shape {:?}",
                input_shape
            );
        }

        // split strides and shapes
        let mut input_kept_shape = vec![];
        let mut input_reduced_shape = vec![];
        let mut input_kept_strides = vec![];
        let mut input_reduced_strides = vec![];

        for axis in 0..input_shape.rank() {
            let size = input_shape.shape()[axis];
            let stride = input_shape.strides()[axis];

            if reduced_axes.contains(&axis) {
                input_reduced_shape.push(size);
                input_reduced_strides.push(stride);
            } else {
                input_kept_shape.push(size);
                input_kept_strides.push(stride);
            }
        }

        // check that things make sense
        let kept_size: usize = input_kept_shape.iter().copied().product();
        let reduction_size: usize = input_reduced_shape.iter().copied().product();

        assert_eq!(input_kept_shape, output_shape.shape(), "Output shape mismatch");
        assert_eq!(kept_size, output_shape.size());
        assert_eq!(input_shape.size(), kept_size * reduction_size);

        // build replacements
        let kept_shape_dense = StridedShape::new_simple(input_kept_shape.clone());
        let reduced_shape_dense = StridedShape::new_simple(input_reduced_shape.clone());

        // pad arrays to ensure they never become zero-sized
        let mut kept_stides_dense = kept_shape_dense.strides().to_vec();
        kept_stides_dense.push(0);
        input_kept_strides.push(0);
        let mut output_kept_strides = output_shape.strides().to_vec();
        output_kept_strides.push(0);

        let replacements = vec![
            ("$KEPT_RANK$", format!("{}", input_kept_shape.len())),
            ("$REDUCED_RANK$", format!("{}", input_reduced_shape.len())),
            ("$KEPT_SIZE$", format!("{}", kept_size)),
            ("$REDUCTION_SIZE$", format!("{}", reduction_size)),
            ("$KEPT_STRIDES_DENSE$", c_array_string(&kept_stides_dense)),
            ("$REDUCED_STRIDES_DENSE$", c_array_string(reduced_shape_dense.strides())),
            (
                "$KEPT_STRIDES$",
                c_nested_array_string(&[input_kept_strides, output_kept_strides]),
            ),
            ("$REDUCED_STRIDES$", c_array_string(&input_reduced_strides)),
            ("$TYPE$", code.ty.clone()),
            ("$IDENTITY$", code.identity.clone()),
            ("$OPERATION$", code.operation.clone()),
            ("$POST_PROCESS$", code.post_process.clone()),
        ];

        // compile the kernel
        let source = fill_replacements(REDUCE_SOURCE, &replacements);
        let key = KernelKey {
            device,
            source,
            func_name: "reduce_kernel".to_owned(),
        };
        let function = compile_cached_kernel(key);

        // wrap everything up
        ReduceKernel {
            function,
            _code: code,
            _reduced_axes: reduced_axes.to_owned(),
            input_shape: input_shape.clone(),
            output_shape: output_shape.clone(),
        }
    }

    pub unsafe fn run(&self, stream: &CudaStream, input: &DeviceTensor, output: &DeviceTensor) {
        assert_eq!(input.strided_shape(), &self.input_shape);
        assert_eq!(output.strided_shape(), &self.output_shape);

        let mut args = KernelArgs::new();
        args.push(input.ptr().ptr());
        args.push(output.ptr().ptr());
        let args = args.finish();

        let warps = self.output_shape.size();
        // TODO see what the effect of increasing this is
        let warps_per_block = 16;
        let threads_per_warp = 32;

        let threads_per_block = (threads_per_warp * warps_per_block) as u32;
        let blocks = ceil_div((warps * threads_per_warp) as u32, threads_per_block as u32);

        self.function
            .launch_kernel(Dim3::single(blocks), Dim3::single(threads_per_block), 0, &stream, &args)
            .unwrap();
    }
}