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
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200

//! Matrix operations for linear algebra computations
//!
//! This module provides:
//! - `matmul`: Batched matrix multiplication with broadcasting support
//! - `dot`: Optimized 2D matrix multiplication
//!
//! Both operations support various numeric types including floating point and integers.

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

// Define all matrix operations using the macro
ops!(matmul, dot);

/// Execute a batched matrix multiplication operation with broadcasting
///
/// Performs batched matrix multiplication (C = A @ B) with support for broadcasting
/// across batch dimensions. Supports arbitrary batch dimensions for both operands.
///
/// # Arguments
/// * `kernel_name` - The matmul kernel to execute (e.g., matmul::F32)
/// * `lhs` - Pointer to left-hand side tensor (A)
/// * `rhs` - Pointer to right-hand side tensor (B)
/// * `output` - Pointer to output tensor buffer (C)
/// * `metadata` - Tensor metadata array (see layout below)
///
/// # Metadata layout
/// - metadata[0]: num_els (total number of output elements)
/// - metadata[1]: lhs_ndim (number of dimensions in lhs)
/// - metadata[2]: rhs_ndim (number of dimensions in rhs)
/// - metadata[3]: batch_ndim (number of batch dimensions in output)
/// - metadata[4..4+lhs_ndim]: lhs_shape (shape of lhs tensor)
/// - metadata[4+lhs_ndim..4+lhs_ndim+rhs_ndim]: rhs_shape (shape of rhs tensor)
/// - metadata[4+lhs_ndim+rhs_ndim..4+lhs_ndim+rhs_ndim+batch_ndim]: batch_shape (broadcasted batch shape)
/// - metadata[4+lhs_ndim+rhs_ndim+batch_ndim..4+2*lhs_ndim+rhs_ndim+batch_ndim]: lhs_strides
/// - metadata[4+2*lhs_ndim+rhs_ndim+batch_ndim..4+2*lhs_ndim+2*rhs_ndim+batch_ndim]: rhs_strides
/// - metadata[4+2*lhs_ndim+2*rhs_ndim+batch_ndim]: lhs_offset (starting offset in lhs)
/// - metadata[4+2*lhs_ndim+2*rhs_ndim+batch_ndim+1]: rhs_offset (starting offset in rhs)
/// - metadata[4+2*lhs_ndim+2*rhs_ndim+batch_ndim+2]: M (rows of lhs matrix)
/// - metadata[4+2*lhs_ndim+2*rhs_ndim+batch_ndim+3]: K (cols of lhs / rows of rhs)
/// - metadata[4+2*lhs_ndim+2*rhs_ndim+batch_ndim+4]: N (cols of rhs matrix)
///
/// # Broadcasting
/// Batch dimensions are broadcasted automatically. Dimensions of size 1 are
/// broadcasted to match the output batch shape.
///
/// # Safety
/// This function uses unsafe FFI calls to C kernels. Caller must ensure:
/// - All pointers are valid and properly aligned
/// - Metadata accurately describes tensor layout
/// - Output buffer has sufficient capacity
/// - Matrix dimensions are compatible: lhs[..., M, K] @ rhs[..., K, N] = output[..., M, N]
///
/// # Returns
/// Returns `Ok(())` on success.
pub fn call_ops_matmul(
    kernel_name: crate::kernels::macros::Kernel,
    lhs: *const c_void,
    rhs: *const c_void,
    output: *mut c_void,
    metadata: &[usize],
) -> Result<()> {
    unsafe {
        dispatch_matmul(kernel_name.0, lhs, rhs, output, metadata);
    }

    Ok(())
}

/// Execute a 2D matrix multiplication operation
///
/// Performs optimized 2D matrix multiplication (C = A @ B) for two 2D matrices only.
/// This is a simplified version of matmul without batch dimensions.
///
/// # Arguments
/// * `kernel_name` - The dot kernel to execute (e.g., dot::F32)
/// * `lhs` - Pointer to left-hand side matrix (A)
/// * `rhs` - Pointer to right-hand side matrix (B)
/// * `output` - Pointer to output matrix buffer (C)
/// * `metadata` - Tensor metadata array (see layout below)
///
/// # Metadata layout
/// - metadata[0]: M (number of rows in lhs)
/// - metadata[1]: K (number of cols in lhs / rows in rhs)
/// - metadata[2]: N (number of cols in rhs)
/// - metadata[3]: lhs_stride_m (stride for lhs rows)
/// - metadata[4]: lhs_stride_k (stride for lhs cols)
/// - metadata[5]: rhs_stride_k (stride for rhs rows)
/// - metadata[6]: rhs_stride_n (stride for rhs cols)
/// - metadata[7]: lhs_offset (starting offset in lhs)
/// - metadata[8]: rhs_offset (starting offset in rhs)
///
/// # Safety
/// This function uses unsafe FFI calls to C kernels. Caller must ensure:
/// - All pointers are valid and properly aligned
/// - Metadata accurately describes tensor layout
/// - Output buffer has sufficient capacity (M * N elements)
/// - Matrix dimensions are compatible: lhs[M, K] @ rhs[K, N] = output[M, N]
///
/// # Returns
/// Returns `Ok(())` on success.
pub fn call_ops_dot(
    kernel_name: crate::kernels::macros::Kernel,
    lhs: *const c_void,
    rhs: *const c_void,
    output: *mut c_void,
    metadata: &[usize],
) -> Result<()> {
    unsafe {
        dispatch_dot(kernel_name.0, lhs, rhs, output, metadata.as_ptr());
    }

    Ok(())
}

/// Macro to generate extern C declarations and dispatch logic for matrix operations
///
/// This macro generates FFI bindings for all supported numeric types.
macro_rules! declare_and_dispatch_matrix {
    ($($op:ident),* $(,)?) => {
        paste::paste! {
            // Extern C declarations for all operations and types
            extern "C" {
                $(
                    fn [<$op _f8e4m3>](lhs: *const c_void, rhs: *const c_void, output: *mut c_void, metadata: *const usize);
                    fn [<$op _f8e5m2>](lhs: *const c_void, rhs: *const c_void, output: *mut c_void, metadata: *const usize);
                    fn [<$op _bf16>](lhs: *const c_void, rhs: *const c_void, output: *mut c_void, metadata: *const usize);
                    fn [<$op _f16>](lhs: *const c_void, rhs: *const c_void, output: *mut c_void, metadata: *const usize);
                    fn [<$op _f32>](lhs: *const c_void, rhs: *const c_void, output: *mut c_void, metadata: *const usize);
                    fn [<$op _f64>](lhs: *const c_void, rhs: *const c_void, output: *mut c_void, metadata: *const usize);
                    fn [<$op _u8>](lhs: *const c_void, rhs: *const c_void, output: *mut c_void, metadata: *const usize);
                    fn [<$op _u16>](lhs: *const c_void, rhs: *const c_void, output: *mut c_void, metadata: *const usize);
                    fn [<$op _u32>](lhs: *const c_void, rhs: *const c_void, output: *mut c_void, metadata: *const usize);
                    fn [<$op _u64>](lhs: *const c_void, rhs: *const c_void, output: *mut c_void, metadata: *const usize);
                    fn [<$op _i8>](lhs: *const c_void, rhs: *const c_void, output: *mut c_void, metadata: *const usize);
                    fn [<$op _i16>](lhs: *const c_void, rhs: *const c_void, output: *mut c_void, metadata: *const usize);
                    fn [<$op _i32>](lhs: *const c_void, rhs: *const c_void, output: *mut c_void, metadata: *const usize);
                    fn [<$op _i64>](lhs: *const c_void, rhs: *const c_void, output: *mut c_void, metadata: *const usize);
                )*
            }

            // Dispatch function for matmul
            unsafe fn dispatch_matmul(
                name: &str,
                lhs: *const c_void,
                rhs: *const c_void,
                output: *mut c_void,
                metadata: &[usize],
            ) {
                match name {
                    "matmul_f8e4m3" => matmul_f8e4m3(lhs, rhs, output, metadata.as_ptr()),
                    "matmul_f8e5m2" => matmul_f8e5m2(lhs, rhs, output, metadata.as_ptr()),
                    "matmul_bf16" => matmul_bf16(lhs, rhs, output, metadata.as_ptr()),
                    "matmul_f16" => matmul_f16(lhs, rhs, output, metadata.as_ptr()),
                    "matmul_f32" => matmul_f32(lhs, rhs, output, metadata.as_ptr()),
                    "matmul_f64" => matmul_f64(lhs, rhs, output, metadata.as_ptr()),
                    "matmul_u8" => matmul_u8(lhs, rhs, output, metadata.as_ptr()),
                    "matmul_u16" => matmul_u16(lhs, rhs, output, metadata.as_ptr()),
                    "matmul_u32" => matmul_u32(lhs, rhs, output, metadata.as_ptr()),
                    "matmul_u64" => matmul_u64(lhs, rhs, output, metadata.as_ptr()),
                    "matmul_i8" => matmul_i8(lhs, rhs, output, metadata.as_ptr()),
                    "matmul_i16" => matmul_i16(lhs, rhs, output, metadata.as_ptr()),
                    "matmul_i32" => matmul_i32(lhs, rhs, output, metadata.as_ptr()),
                    "matmul_i64" => matmul_i64(lhs, rhs, output, metadata.as_ptr()),
                    _ => panic!("Unsupported matmul kernel: {}", name),
                }
            }

            // Dispatch function for dot
            unsafe fn dispatch_dot(
                name: &str,
                lhs: *const c_void,
                rhs: *const c_void,
                output: *mut c_void,
                metadata: *const usize,
            ) {
                match name {
                    "dot_f8e4m3" => dot_f8e4m3(lhs, rhs, output, metadata),
                    "dot_f8e5m2" => dot_f8e5m2(lhs, rhs, output, metadata),
                    "dot_bf16" => dot_bf16(lhs, rhs, output, metadata),
                    "dot_f16" => dot_f16(lhs, rhs, output, metadata),
                    "dot_f32" => dot_f32(lhs, rhs, output, metadata),
                    "dot_f64" => dot_f64(lhs, rhs, output, metadata),
                    "dot_u8" => dot_u8(lhs, rhs, output, metadata),
                    "dot_u16" => dot_u16(lhs, rhs, output, metadata),
                    "dot_u32" => dot_u32(lhs, rhs, output, metadata),
                    "dot_u64" => dot_u64(lhs, rhs, output, metadata),
                    "dot_i8" => dot_i8(lhs, rhs, output, metadata),
                    "dot_i16" => dot_i16(lhs, rhs, output, metadata),
                    "dot_i32" => dot_i32(lhs, rhs, output, metadata),
                    "dot_i64" => dot_i64(lhs, rhs, output, metadata),
                    _ => panic!("Unsupported dot kernel: {}", name),
                }
            }
        }
    };
}

// Declare matrix operations
declare_and_dispatch_matrix!(matmul, dot);