hodu_cpu_kernels 0.2.4

hodu cpu kernels
Documentation 
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
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169

//! Windowing operations for tensors
//!
//! This module provides sliding window reduction operations:
//! - reduce_window_max: Maximum value in each window
//! - reduce_window_min: Minimum value in each window
//! - reduce_window_sum: Sum of values in each window
//! - reduce_window_mean: Mean (average) of values in each window
//!
//! These operations apply a reduction function over sliding windows of the input tensor,
//! with support for configurable window size, stride, and padding.

use crate::{
    error::Result,
    kernels::{macros::ops, Kernel},
};
use core::ffi::c_void;

ops!(
    reduce_window_max,
    reduce_window_mean,
    reduce_window_sum,
    reduce_window_min
);

extern "C" {
    fn reduce_window_max_f8e4m3(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_max_f8e5m2(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_max_bf16(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_max_f16(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_max_f32(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_max_f64(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_max_i8(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_max_i16(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_max_i32(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_max_i64(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_max_u8(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_max_u16(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_max_u32(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_max_u64(input: *const c_void, output: *mut c_void, metadata: *const usize);

    fn reduce_window_mean_f8e4m3(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_mean_f8e5m2(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_mean_bf16(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_mean_f16(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_mean_f32(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_mean_f64(input: *const c_void, output: *mut c_void, metadata: *const usize);

    fn reduce_window_sum_f8e4m3(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_sum_f8e5m2(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_sum_bf16(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_sum_f16(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_sum_f32(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_sum_f64(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_sum_i8(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_sum_i16(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_sum_i32(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_sum_i64(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_sum_u8(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_sum_u16(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_sum_u32(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_sum_u64(input: *const c_void, output: *mut c_void, metadata: *const usize);

    fn reduce_window_min_f8e4m3(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_min_f8e5m2(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_min_bf16(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_min_f16(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_min_f32(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_min_f64(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_min_i8(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_min_i16(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_min_i32(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_min_i64(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_min_u8(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_min_u16(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_min_u32(input: *const c_void, output: *mut c_void, metadata: *const usize);
    fn reduce_window_min_u64(input: *const c_void, output: *mut c_void, metadata: *const usize);
}

/// Call windowing reduction operation by kernel name
///
/// Applies a reduction operation over sliding windows of the input tensor.
///
/// # Metadata layout
/// - metadata[0]: output_size (total number of elements in output)
/// - metadata[1]: num_dims (number of dimensions)
/// - metadata[2..2+num_dims]: input_shape
/// - metadata[2+num_dims..2+2*num_dims]: input_strides
/// - metadata[2+2*num_dims]: offset (starting offset in input)
/// - metadata[3+2*num_dims..3+3*num_dims]: window_shape (size of window in each dimension)
/// - metadata[3+3*num_dims..3+4*num_dims]: strides (step size in each dimension)
/// - metadata[3+4*num_dims..3+4*num_dims+2*num_dims]: padding (before and after for each dimension)
/// - metadata[3+6*num_dims..]: output_shape
///
/// # Padding layout
/// For each dimension i, padding is specified as [pad_before_i, pad_after_i].
/// Padded areas are treated as negative infinity for max, positive infinity for min,
/// and zero for sum/mean operations.
///
/// # Safety
/// - `input` must point to valid tensor data of the appropriate type
/// - `output` must point to a valid output buffer with sufficient capacity
/// - Metadata must accurately describe the tensor layout and window parameters
pub fn call_ops_reduce_window(
    kernel_name: Kernel,
    input: *const c_void,
    output: *mut c_void,
    metadata: &[usize],
) -> Result<()> {
    let kernel_str = kernel_name.0;
    unsafe {
        match kernel_str {
            "reduce_window_max_f8e4m3" => reduce_window_max_f8e4m3(input, output, metadata.as_ptr()),
            "reduce_window_max_f8e5m2" => reduce_window_max_f8e5m2(input, output, metadata.as_ptr()),
            "reduce_window_max_bf16" => reduce_window_max_bf16(input, output, metadata.as_ptr()),
            "reduce_window_max_f16" => reduce_window_max_f16(input, output, metadata.as_ptr()),
            "reduce_window_max_f32" => reduce_window_max_f32(input, output, metadata.as_ptr()),
            "reduce_window_max_f64" => reduce_window_max_f64(input, output, metadata.as_ptr()),
            "reduce_window_max_i8" => reduce_window_max_i8(input, output, metadata.as_ptr()),
            "reduce_window_max_i16" => reduce_window_max_i16(input, output, metadata.as_ptr()),
            "reduce_window_max_i32" => reduce_window_max_i32(input, output, metadata.as_ptr()),
            "reduce_window_max_i64" => reduce_window_max_i64(input, output, metadata.as_ptr()),
            "reduce_window_max_u8" => reduce_window_max_u8(input, output, metadata.as_ptr()),
            "reduce_window_max_u16" => reduce_window_max_u16(input, output, metadata.as_ptr()),
            "reduce_window_max_u32" => reduce_window_max_u32(input, output, metadata.as_ptr()),
            "reduce_window_max_u64" => reduce_window_max_u64(input, output, metadata.as_ptr()),

            "reduce_window_mean_f8e4m3" => reduce_window_mean_f8e4m3(input, output, metadata.as_ptr()),
            "reduce_window_mean_f8e5m2" => reduce_window_mean_f8e5m2(input, output, metadata.as_ptr()),
            "reduce_window_mean_bf16" => reduce_window_mean_bf16(input, output, metadata.as_ptr()),
            "reduce_window_mean_f16" => reduce_window_mean_f16(input, output, metadata.as_ptr()),
            "reduce_window_mean_f32" => reduce_window_mean_f32(input, output, metadata.as_ptr()),
            "reduce_window_mean_f64" => reduce_window_mean_f64(input, output, metadata.as_ptr()),

            "reduce_window_sum_f8e4m3" => reduce_window_sum_f8e4m3(input, output, metadata.as_ptr()),
            "reduce_window_sum_f8e5m2" => reduce_window_sum_f8e5m2(input, output, metadata.as_ptr()),
            "reduce_window_sum_bf16" => reduce_window_sum_bf16(input, output, metadata.as_ptr()),
            "reduce_window_sum_f16" => reduce_window_sum_f16(input, output, metadata.as_ptr()),
            "reduce_window_sum_f32" => reduce_window_sum_f32(input, output, metadata.as_ptr()),
            "reduce_window_sum_f64" => reduce_window_sum_f64(input, output, metadata.as_ptr()),
            "reduce_window_sum_i8" => reduce_window_sum_i8(input, output, metadata.as_ptr()),
            "reduce_window_sum_i16" => reduce_window_sum_i16(input, output, metadata.as_ptr()),
            "reduce_window_sum_i32" => reduce_window_sum_i32(input, output, metadata.as_ptr()),
            "reduce_window_sum_i64" => reduce_window_sum_i64(input, output, metadata.as_ptr()),
            "reduce_window_sum_u8" => reduce_window_sum_u8(input, output, metadata.as_ptr()),
            "reduce_window_sum_u16" => reduce_window_sum_u16(input, output, metadata.as_ptr()),
            "reduce_window_sum_u32" => reduce_window_sum_u32(input, output, metadata.as_ptr()),
            "reduce_window_sum_u64" => reduce_window_sum_u64(input, output, metadata.as_ptr()),

            "reduce_window_min_f8e4m3" => reduce_window_min_f8e4m3(input, output, metadata.as_ptr()),
            "reduce_window_min_f8e5m2" => reduce_window_min_f8e5m2(input, output, metadata.as_ptr()),
            "reduce_window_min_bf16" => reduce_window_min_bf16(input, output, metadata.as_ptr()),
            "reduce_window_min_f16" => reduce_window_min_f16(input, output, metadata.as_ptr()),
            "reduce_window_min_f32" => reduce_window_min_f32(input, output, metadata.as_ptr()),
            "reduce_window_min_f64" => reduce_window_min_f64(input, output, metadata.as_ptr()),
            "reduce_window_min_i8" => reduce_window_min_i8(input, output, metadata.as_ptr()),
            "reduce_window_min_i16" => reduce_window_min_i16(input, output, metadata.as_ptr()),
            "reduce_window_min_i32" => reduce_window_min_i32(input, output, metadata.as_ptr()),
            "reduce_window_min_i64" => reduce_window_min_i64(input, output, metadata.as_ptr()),
            "reduce_window_min_u8" => reduce_window_min_u8(input, output, metadata.as_ptr()),
            "reduce_window_min_u16" => reduce_window_min_u16(input, output, metadata.as_ptr()),
            "reduce_window_min_u32" => reduce_window_min_u32(input, output, metadata.as_ptr()),
            "reduce_window_min_u64" => reduce_window_min_u64(input, output, metadata.as_ptr()),

            _ => panic!("Unsupported reduce_window kernel: {:?}", kernel_name),
        }
    }

    Ok(())
}