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
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246

//! AVX2 scalar operation kernels
//!
//! Broadcasts scalar to vector and applies operation.

#[cfg(target_arch = "x86_64")]
use std::arch::x86_64::*;

use super::{scalar_scalar_f32, scalar_scalar_f64};
use crate::ops::BinaryOp;

const F32_LANES: usize = 8;
const F64_LANES: usize = 4;

/// AVX2 scalar operation for f32
#[target_feature(enable = "avx2")]
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;

    match op {
        BinaryOp::Add => scalar_add_f32(a, scalar, out, chunks),
        BinaryOp::Sub => scalar_sub_f32(a, scalar, out, chunks),
        BinaryOp::Mul => scalar_mul_f32(a, scalar, out, chunks),
        BinaryOp::Div => scalar_div_f32(a, scalar, out, chunks),
        BinaryOp::Max => scalar_max_f32(a, scalar, out, chunks),
        BinaryOp::Min => scalar_min_f32(a, scalar, out, chunks),
        _ => {
            scalar_scalar_f32(op, a, scalar, out, len);
            return;
        }
    }

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

/// AVX2 scalar operation for f64
#[target_feature(enable = "avx2")]
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;

    match op {
        BinaryOp::Add => scalar_add_f64(a, scalar, out, chunks),
        BinaryOp::Sub => scalar_sub_f64(a, scalar, out, chunks),
        BinaryOp::Mul => scalar_mul_f64(a, scalar, out, chunks),
        BinaryOp::Div => scalar_div_f64(a, scalar, out, chunks),
        BinaryOp::Max => scalar_max_f64(a, scalar, out, chunks),
        BinaryOp::Min => scalar_min_f64(a, scalar, out, chunks),
        _ => {
            scalar_scalar_f64(op, a, scalar, out, len);
            return;
        }
    }

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

// ============================================================================
// f32 kernels
// ============================================================================

#[target_feature(enable = "avx2")]
unsafe fn scalar_add_f32(a: *const f32, scalar: f32, out: *mut f32, chunks: usize) {
    let vs = _mm256_set1_ps(scalar);
    for i in 0..chunks {
        let offset = i * F32_LANES;
        let va = _mm256_loadu_ps(a.add(offset));
        let vr = _mm256_add_ps(va, vs);
        _mm256_storeu_ps(out.add(offset), vr);
    }
}

#[target_feature(enable = "avx2")]
unsafe fn scalar_sub_f32(a: *const f32, scalar: f32, out: *mut f32, chunks: usize) {
    let vs = _mm256_set1_ps(scalar);
    for i in 0..chunks {
        let offset = i * F32_LANES;
        let va = _mm256_loadu_ps(a.add(offset));
        let vr = _mm256_sub_ps(va, vs);
        _mm256_storeu_ps(out.add(offset), vr);
    }
}

#[target_feature(enable = "avx2")]
unsafe fn scalar_mul_f32(a: *const f32, scalar: f32, out: *mut f32, chunks: usize) {
    let vs = _mm256_set1_ps(scalar);
    for i in 0..chunks {
        let offset = i * F32_LANES;
        let va = _mm256_loadu_ps(a.add(offset));
        let vr = _mm256_mul_ps(va, vs);
        _mm256_storeu_ps(out.add(offset), vr);
    }
}

#[target_feature(enable = "avx2")]
unsafe fn scalar_div_f32(a: *const f32, scalar: f32, out: *mut f32, chunks: usize) {
    let vs = _mm256_set1_ps(scalar);
    for i in 0..chunks {
        let offset = i * F32_LANES;
        let va = _mm256_loadu_ps(a.add(offset));
        let vr = _mm256_div_ps(va, vs);
        _mm256_storeu_ps(out.add(offset), vr);
    }
}

#[target_feature(enable = "avx2")]
unsafe fn scalar_max_f32(a: *const f32, scalar: f32, out: *mut f32, chunks: usize) {
    let vs = _mm256_set1_ps(scalar);
    for i in 0..chunks {
        let offset = i * F32_LANES;
        let va = _mm256_loadu_ps(a.add(offset));
        let vr = _mm256_max_ps(va, vs);
        _mm256_storeu_ps(out.add(offset), vr);
    }
}

#[target_feature(enable = "avx2")]
unsafe fn scalar_min_f32(a: *const f32, scalar: f32, out: *mut f32, chunks: usize) {
    let vs = _mm256_set1_ps(scalar);
    for i in 0..chunks {
        let offset = i * F32_LANES;
        let va = _mm256_loadu_ps(a.add(offset));
        let vr = _mm256_min_ps(va, vs);
        _mm256_storeu_ps(out.add(offset), vr);
    }
}

// ============================================================================
// f64 kernels
// ============================================================================

#[target_feature(enable = "avx2")]
unsafe fn scalar_add_f64(a: *const f64, scalar: f64, out: *mut f64, chunks: usize) {
    let vs = _mm256_set1_pd(scalar);
    for i in 0..chunks {
        let offset = i * F64_LANES;
        let va = _mm256_loadu_pd(a.add(offset));
        let vr = _mm256_add_pd(va, vs);
        _mm256_storeu_pd(out.add(offset), vr);
    }
}

#[target_feature(enable = "avx2")]
unsafe fn scalar_sub_f64(a: *const f64, scalar: f64, out: *mut f64, chunks: usize) {
    let vs = _mm256_set1_pd(scalar);
    for i in 0..chunks {
        let offset = i * F64_LANES;
        let va = _mm256_loadu_pd(a.add(offset));
        let vr = _mm256_sub_pd(va, vs);
        _mm256_storeu_pd(out.add(offset), vr);
    }
}

#[target_feature(enable = "avx2")]
unsafe fn scalar_mul_f64(a: *const f64, scalar: f64, out: *mut f64, chunks: usize) {
    let vs = _mm256_set1_pd(scalar);
    for i in 0..chunks {
        let offset = i * F64_LANES;
        let va = _mm256_loadu_pd(a.add(offset));
        let vr = _mm256_mul_pd(va, vs);
        _mm256_storeu_pd(out.add(offset), vr);
    }
}

#[target_feature(enable = "avx2")]
unsafe fn scalar_div_f64(a: *const f64, scalar: f64, out: *mut f64, chunks: usize) {
    let vs = _mm256_set1_pd(scalar);
    for i in 0..chunks {
        let offset = i * F64_LANES;
        let va = _mm256_loadu_pd(a.add(offset));
        let vr = _mm256_div_pd(va, vs);
        _mm256_storeu_pd(out.add(offset), vr);
    }
}

#[target_feature(enable = "avx2")]
unsafe fn scalar_max_f64(a: *const f64, scalar: f64, out: *mut f64, chunks: usize) {
    let vs = _mm256_set1_pd(scalar);
    for i in 0..chunks {
        let offset = i * F64_LANES;
        let va = _mm256_loadu_pd(a.add(offset));
        let vr = _mm256_max_pd(va, vs);
        _mm256_storeu_pd(out.add(offset), vr);
    }
}

#[target_feature(enable = "avx2")]
unsafe fn scalar_min_f64(a: *const f64, scalar: f64, out: *mut f64, chunks: usize) {
    let vs = _mm256_set1_pd(scalar);
    for i in 0..chunks {
        let offset = i * F64_LANES;
        let va = _mm256_loadu_pd(a.add(offset));
        let vr = _mm256_min_pd(va, vs);
        _mm256_storeu_pd(out.add(offset), vr);
    }
}

// ============================================================================
// Reverse subtract (scalar - a) kernels
// ============================================================================

/// AVX2 reverse scalar subtract for f32: out[i] = scalar - a[i]
#[target_feature(enable = "avx2")]
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 = _mm256_set1_ps(scalar);

    for i in 0..chunks {
        let offset = i * F32_LANES;
        let va = _mm256_loadu_ps(a.add(offset));
        let vr = _mm256_sub_ps(vs, va);
        _mm256_storeu_ps(out.add(offset), vr);
    }

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

/// AVX2 reverse scalar subtract for f64: out[i] = scalar - a[i]
#[target_feature(enable = "avx2")]
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 = _mm256_set1_pd(scalar);

    for i in 0..chunks {
        let offset = i * F64_LANES;
        let va = _mm256_loadu_pd(a.add(offset));
        let vr = _mm256_sub_pd(vs, va);
        _mm256_storeu_pd(out.add(offset), vr);
    }

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