trueno 0.16.4

High-performance SIMD compute library with GPU support for matrix operations
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

//! Backend dispatch macros for vector operations
//!
//! These macros handle routing operations to the appropriate SIMD backend
//! based on CPU feature detection and the selected Backend variant.

/// Macro to dispatch binary operations to appropriate backend
///
/// Routes operations like add, sub, mul, div to the best available SIMD backend.
/// Falls back to scalar implementation when a hardware backend is unavailable.
#[macro_export]
macro_rules! dispatch_binary_op {
    ($backend:expr, $op:ident, $a:expr, $b:expr, $result:expr) => {
        // SAFETY: CPU feature verified at runtime, slices bounds-checked
        unsafe {
            match $backend {
                $crate::Backend::Scalar => {
                    $crate::backends::scalar::ScalarBackend::$op($a, $b, $result)
                }
                #[cfg(target_arch = "x86_64")]
                $crate::Backend::SSE2 | $crate::Backend::AVX => {
                    $crate::backends::sse2::Sse2Backend::$op($a, $b, $result)
                }
                #[cfg(target_arch = "x86_64")]
                $crate::Backend::AVX2 => $crate::backends::avx2::Avx2Backend::$op($a, $b, $result),
                #[cfg(target_arch = "x86_64")]
                $crate::Backend::AVX512 => {
                    $crate::backends::avx512::Avx512Backend::$op($a, $b, $result)
                }
                #[cfg(not(target_arch = "x86_64"))]
                $crate::Backend::SSE2
                | $crate::Backend::AVX
                | $crate::Backend::AVX2
                | $crate::Backend::AVX512 => {
                    $crate::backends::scalar::ScalarBackend::$op($a, $b, $result)
                }
                #[cfg(any(target_arch = "aarch64", target_arch = "arm"))]
                $crate::Backend::NEON => $crate::backends::neon::NeonBackend::$op($a, $b, $result),
                #[cfg(not(any(target_arch = "aarch64", target_arch = "arm")))]
                $crate::Backend::NEON => {
                    $crate::backends::scalar::ScalarBackend::$op($a, $b, $result)
                }
                #[cfg(target_arch = "wasm32")]
                $crate::Backend::WasmSIMD => {
                    $crate::backends::wasm::WasmBackend::$op($a, $b, $result)
                }
                #[cfg(not(target_arch = "wasm32"))]
                $crate::Backend::WasmSIMD => {
                    $crate::backends::scalar::ScalarBackend::$op($a, $b, $result)
                }
                $crate::Backend::GPU | $crate::Backend::Auto => {
                    $crate::backends::scalar::ScalarBackend::$op($a, $b, $result)
                }
            }
        }
    };
}

/// Macro to dispatch reduction operations (return f32)
///
/// Routes operations like sum, max, min to the best available SIMD backend.
/// Falls back to scalar implementation when a hardware backend is unavailable.
#[macro_export]
macro_rules! dispatch_reduction {
    ($backend:expr, $op:ident, $data:expr) => {
        // SAFETY: CPU feature verified at runtime, slices bounds-checked
        unsafe {
            match $backend {
                $crate::Backend::Scalar => $crate::backends::scalar::ScalarBackend::$op($data),
                #[cfg(target_arch = "x86_64")]
                $crate::Backend::SSE2 | $crate::Backend::AVX => {
                    $crate::backends::sse2::Sse2Backend::$op($data)
                }
                #[cfg(target_arch = "x86_64")]
                $crate::Backend::AVX2 => $crate::backends::avx2::Avx2Backend::$op($data),
                #[cfg(target_arch = "x86_64")]
                $crate::Backend::AVX512 => $crate::backends::avx512::Avx512Backend::$op($data),
                #[cfg(not(target_arch = "x86_64"))]
                $crate::Backend::SSE2
                | $crate::Backend::AVX
                | $crate::Backend::AVX2
                | $crate::Backend::AVX512 => $crate::backends::scalar::ScalarBackend::$op($data),
                #[cfg(any(target_arch = "aarch64", target_arch = "arm"))]
                $crate::Backend::NEON => $crate::backends::neon::NeonBackend::$op($data),
                #[cfg(not(any(target_arch = "aarch64", target_arch = "arm")))]
                $crate::Backend::NEON => $crate::backends::scalar::ScalarBackend::$op($data),
                #[cfg(target_arch = "wasm32")]
                $crate::Backend::WasmSIMD => $crate::backends::wasm::WasmBackend::$op($data),
                #[cfg(not(target_arch = "wasm32"))]
                $crate::Backend::WasmSIMD => $crate::backends::scalar::ScalarBackend::$op($data),
                $crate::Backend::GPU | $crate::Backend::Auto => {
                    $crate::backends::scalar::ScalarBackend::$op($data)
                }
            }
        }
    };
}

/// Macro to dispatch unary operations (a -> result)
///
/// Routes operations like relu, sigmoid to the best available SIMD backend.
/// Falls back to scalar implementation when a hardware backend is unavailable.
#[macro_export]
macro_rules! dispatch_unary_op {
    ($backend:expr, $op:ident, $a:expr, $result:expr) => {
        // SAFETY: CPU feature verified at runtime, slices bounds-checked
        unsafe {
            match $backend {
                $crate::Backend::Scalar => {
                    $crate::backends::scalar::ScalarBackend::$op($a, $result)
                }
                #[cfg(target_arch = "x86_64")]
                $crate::Backend::SSE2 | $crate::Backend::AVX => {
                    $crate::backends::sse2::Sse2Backend::$op($a, $result)
                }
                #[cfg(target_arch = "x86_64")]
                $crate::Backend::AVX2 => $crate::backends::avx2::Avx2Backend::$op($a, $result),
                #[cfg(target_arch = "x86_64")]
                $crate::Backend::AVX512 => {
                    $crate::backends::avx512::Avx512Backend::$op($a, $result)
                }
                #[cfg(not(target_arch = "x86_64"))]
                $crate::Backend::SSE2
                | $crate::Backend::AVX
                | $crate::Backend::AVX2
                | $crate::Backend::AVX512 => {
                    $crate::backends::scalar::ScalarBackend::$op($a, $result)
                }
                #[cfg(any(target_arch = "aarch64", target_arch = "arm"))]
                $crate::Backend::NEON => $crate::backends::neon::NeonBackend::$op($a, $result),
                #[cfg(not(any(target_arch = "aarch64", target_arch = "arm")))]
                $crate::Backend::NEON => $crate::backends::scalar::ScalarBackend::$op($a, $result),
                #[cfg(target_arch = "wasm32")]
                $crate::Backend::WasmSIMD => $crate::backends::wasm::WasmBackend::$op($a, $result),
                #[cfg(not(target_arch = "wasm32"))]
                $crate::Backend::WasmSIMD => {
                    $crate::backends::scalar::ScalarBackend::$op($a, $result)
                }
                $crate::Backend::GPU | $crate::Backend::Auto => {
                    $crate::backends::scalar::ScalarBackend::$op($a, $result)
                }
            }
        }
    };
}

pub use dispatch_binary_op;
pub use dispatch_reduction;
pub use dispatch_unary_op;