numr 0.5.2

High-performance numerical computing with multi-backend GPU acceleration (CPU/CUDA/WebGPU)
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
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215

//! NEON scalar operation kernels for ARM64
//!
//! Provides vectorized tensor-scalar operations using 128-bit NEON registers.
//!
//! # Supported Operations
//!
//! - Add: a + scalar
//! - Sub: a - scalar
//! - Mul: a * scalar
//! - Div: a / scalar
//! - Max: max(a, scalar)
//! - Min: min(a, scalar)
//!
//! # Unsupported (scalar fallback)
//!
//! - Pow: requires libm, no direct SIMD instruction
//! - Atan2: requires libm

#[cfg(target_arch = "aarch64")]
use std::arch::aarch64::*;

use crate::ops::BinaryOp;

const F32_LANES: usize = 4;
const F64_LANES: usize = 2;

/// NEON scalar operation for f32
///
/// # Safety
/// - CPU must support NEON (always true on AArch64)
/// - `a` and `out` must point to `len` valid f32 elements
#[cfg(target_arch = "aarch64")]
#[target_feature(enable = "neon")]
pub unsafe fn scalar_f32(op: BinaryOp, a: *const f32, scalar: f32, out: *mut f32, len: usize) {
    let chunks = len / F32_LANES;
    let remainder = len % F32_LANES;

    let vs = vdupq_n_f32(scalar);

    match op {
        BinaryOp::Add => {
            for i in 0..chunks {
                let offset = i * F32_LANES;
                let va = vld1q_f32(a.add(offset));
                let result = vaddq_f32(va, vs);
                vst1q_f32(out.add(offset), result);
            }
        }
        BinaryOp::Sub => {
            for i in 0..chunks {
                let offset = i * F32_LANES;
                let va = vld1q_f32(a.add(offset));
                let result = vsubq_f32(va, vs);
                vst1q_f32(out.add(offset), result);
            }
        }
        BinaryOp::Mul => {
            for i in 0..chunks {
                let offset = i * F32_LANES;
                let va = vld1q_f32(a.add(offset));
                let result = vmulq_f32(va, vs);
                vst1q_f32(out.add(offset), result);
            }
        }
        BinaryOp::Div => {
            for i in 0..chunks {
                let offset = i * F32_LANES;
                let va = vld1q_f32(a.add(offset));
                let result = vdivq_f32(va, vs);
                vst1q_f32(out.add(offset), result);
            }
        }
        BinaryOp::Max => {
            for i in 0..chunks {
                let offset = i * F32_LANES;
                let va = vld1q_f32(a.add(offset));
                let result = vmaxq_f32(va, vs);
                vst1q_f32(out.add(offset), result);
            }
        }
        BinaryOp::Min => {
            for i in 0..chunks {
                let offset = i * F32_LANES;
                let va = vld1q_f32(a.add(offset));
                let result = vminq_f32(va, vs);
                vst1q_f32(out.add(offset), result);
            }
        }
        BinaryOp::Pow | BinaryOp::Atan2 => {
            // Fallback to scalar for unsupported ops
            super::super::scalar_scalar_f32(op, a, scalar, out, len);
            return;
        }
    }

    // Scalar tail
    if remainder > 0 {
        let offset = chunks * F32_LANES;
        super::super::scalar_scalar_f32(op, a.add(offset), scalar, out.add(offset), remainder);
    }
}

/// NEON scalar operation for f64
///
/// # Safety
/// - CPU must support NEON (always true on AArch64)
/// - `a` and `out` must point to `len` valid f64 elements
#[cfg(target_arch = "aarch64")]
#[target_feature(enable = "neon")]
pub unsafe fn scalar_f64(op: BinaryOp, a: *const f64, scalar: f64, out: *mut f64, len: usize) {
    let chunks = len / F64_LANES;
    let remainder = len % F64_LANES;

    let vs = vdupq_n_f64(scalar);

    match op {
        BinaryOp::Add => {
            for i in 0..chunks {
                let offset = i * F64_LANES;
                let va = vld1q_f64(a.add(offset));
                let result = vaddq_f64(va, vs);
                vst1q_f64(out.add(offset), result);
            }
        }
        BinaryOp::Sub => {
            for i in 0..chunks {
                let offset = i * F64_LANES;
                let va = vld1q_f64(a.add(offset));
                let result = vsubq_f64(va, vs);
                vst1q_f64(out.add(offset), result);
            }
        }
        BinaryOp::Mul => {
            for i in 0..chunks {
                let offset = i * F64_LANES;
                let va = vld1q_f64(a.add(offset));
                let result = vmulq_f64(va, vs);
                vst1q_f64(out.add(offset), result);
            }
        }
        BinaryOp::Div => {
            for i in 0..chunks {
                let offset = i * F64_LANES;
                let va = vld1q_f64(a.add(offset));
                let result = vdivq_f64(va, vs);
                vst1q_f64(out.add(offset), result);
            }
        }
        BinaryOp::Max => {
            for i in 0..chunks {
                let offset = i * F64_LANES;
                let va = vld1q_f64(a.add(offset));
                let result = vmaxq_f64(va, vs);
                vst1q_f64(out.add(offset), result);
            }
        }
        BinaryOp::Min => {
            for i in 0..chunks {
                let offset = i * F64_LANES;
                let va = vld1q_f64(a.add(offset));
                let result = vminq_f64(va, vs);
                vst1q_f64(out.add(offset), result);
            }
        }
        BinaryOp::Pow | BinaryOp::Atan2 => {
            // Fallback to scalar for unsupported ops
            super::super::scalar_scalar_f64(op, a, scalar, out, len);
            return;
        }
    }

    // Scalar tail
    if remainder > 0 {
        let offset = chunks * F64_LANES;
        super::super::scalar_scalar_f64(op, a.add(offset), scalar, out.add(offset), remainder);
    }
}

/// NEON reverse scalar subtract for f32: out[i] = scalar - a[i]
pub unsafe fn rsub_scalar_f32(a: *const f32, scalar: f32, out: *mut f32, len: usize) {
    let chunks = len / F32_LANES;
    let remainder = len % F32_LANES;
    let vs = vdupq_n_f32(scalar);

    for i in 0..chunks {
        let offset = i * F32_LANES;
        let va = vld1q_f32(a.add(offset));
        let vr = vsubq_f32(vs, va);
        vst1q_f32(out.add(offset), vr);
    }

    for i in 0..remainder {
        let offset = chunks * F32_LANES + i;
        *out.add(offset) = scalar - *a.add(offset);
    }
}

/// NEON reverse scalar subtract for f64: out[i] = scalar - a[i]
pub unsafe fn rsub_scalar_f64(a: *const f64, scalar: f64, out: *mut f64, len: usize) {
    let chunks = len / F64_LANES;
    let remainder = len % F64_LANES;
    let vs = vdupq_n_f64(scalar);

    for i in 0..chunks {
        let offset = i * F64_LANES;
        let va = vld1q_f64(a.add(offset));
        let vr = vsubq_f64(vs, va);
        vst1q_f64(out.add(offset), vr);
    }

    for i in 0..remainder {
        let offset = chunks * F64_LANES + i;
        *out.add(offset) = scalar - *a.add(offset);
    }
}