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

//! AVX-512 fused elementwise kernels (512-bit)

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

use super::super::{
    fused_add_mul_scalar_f32 as fused_add_mul_fallback_f32,
    fused_add_mul_scalar_f64 as fused_add_mul_fallback_f64,
    fused_mul_add_scalar_f32 as fused_mul_add_fallback_f32,
    fused_mul_add_scalar_f64 as fused_mul_add_fallback_f64, fused_mul_add_scalar_loop_f32,
    fused_mul_add_scalar_loop_f64,
};

const F32_LANES: usize = 16;
const F64_LANES: usize = 8;

/// AVX-512 fused_mul_add for f32: out = a * b + c
#[target_feature(enable = "avx512f")]
pub unsafe fn fused_mul_add_f32(
    a: *const f32,
    b: *const f32,
    c: *const f32,
    out: *mut f32,
    len: usize,
) {
    let chunks = len / F32_LANES;

    for i in 0..chunks {
        let offset = i * F32_LANES;
        let va = _mm512_loadu_ps(a.add(offset));
        let vb = _mm512_loadu_ps(b.add(offset));
        let vc = _mm512_loadu_ps(c.add(offset));
        let result = _mm512_fmadd_ps(va, vb, vc);
        _mm512_storeu_ps(out.add(offset), result);
    }

    let processed = chunks * F32_LANES;
    if processed < len {
        fused_mul_add_fallback_f32(
            a.add(processed),
            b.add(processed),
            c.add(processed),
            out.add(processed),
            len - processed,
        );
    }
}

/// AVX-512 fused_mul_add for f64: out = a * b + c
#[target_feature(enable = "avx512f")]
pub unsafe fn fused_mul_add_f64(
    a: *const f64,
    b: *const f64,
    c: *const f64,
    out: *mut f64,
    len: usize,
) {
    let chunks = len / F64_LANES;

    for i in 0..chunks {
        let offset = i * F64_LANES;
        let va = _mm512_loadu_pd(a.add(offset));
        let vb = _mm512_loadu_pd(b.add(offset));
        let vc = _mm512_loadu_pd(c.add(offset));
        let result = _mm512_fmadd_pd(va, vb, vc);
        _mm512_storeu_pd(out.add(offset), result);
    }

    let processed = chunks * F64_LANES;
    if processed < len {
        fused_mul_add_fallback_f64(
            a.add(processed),
            b.add(processed),
            c.add(processed),
            out.add(processed),
            len - processed,
        );
    }
}

/// AVX-512 fused_add_mul for f32: out = (a + b) * c
#[target_feature(enable = "avx512f")]
pub unsafe fn fused_add_mul_f32(
    a: *const f32,
    b: *const f32,
    c: *const f32,
    out: *mut f32,
    len: usize,
) {
    let chunks = len / F32_LANES;

    for i in 0..chunks {
        let offset = i * F32_LANES;
        let va = _mm512_loadu_ps(a.add(offset));
        let vb = _mm512_loadu_ps(b.add(offset));
        let vc = _mm512_loadu_ps(c.add(offset));
        let sum = _mm512_add_ps(va, vb);
        let result = _mm512_mul_ps(sum, vc);
        _mm512_storeu_ps(out.add(offset), result);
    }

    let processed = chunks * F32_LANES;
    if processed < len {
        fused_add_mul_fallback_f32(
            a.add(processed),
            b.add(processed),
            c.add(processed),
            out.add(processed),
            len - processed,
        );
    }
}

/// AVX-512 fused_add_mul for f64: out = (a + b) * c
#[target_feature(enable = "avx512f")]
pub unsafe fn fused_add_mul_f64(
    a: *const f64,
    b: *const f64,
    c: *const f64,
    out: *mut f64,
    len: usize,
) {
    let chunks = len / F64_LANES;

    for i in 0..chunks {
        let offset = i * F64_LANES;
        let va = _mm512_loadu_pd(a.add(offset));
        let vb = _mm512_loadu_pd(b.add(offset));
        let vc = _mm512_loadu_pd(c.add(offset));
        let sum = _mm512_add_pd(va, vb);
        let result = _mm512_mul_pd(sum, vc);
        _mm512_storeu_pd(out.add(offset), result);
    }

    let processed = chunks * F64_LANES;
    if processed < len {
        fused_add_mul_fallback_f64(
            a.add(processed),
            b.add(processed),
            c.add(processed),
            out.add(processed),
            len - processed,
        );
    }
}

/// AVX-512 fused_mul_add_scalar for f32: out = a * scale + bias
#[target_feature(enable = "avx512f")]
pub unsafe fn fused_mul_add_scalar_f32(
    a: *const f32,
    scale: f32,
    bias: f32,
    out: *mut f32,
    len: usize,
) {
    let chunks = len / F32_LANES;
    let vscale = _mm512_set1_ps(scale);
    let vbias = _mm512_set1_ps(bias);

    for i in 0..chunks {
        let offset = i * F32_LANES;
        let va = _mm512_loadu_ps(a.add(offset));
        let result = _mm512_fmadd_ps(va, vscale, vbias);
        _mm512_storeu_ps(out.add(offset), result);
    }

    let processed = chunks * F32_LANES;
    if processed < len {
        fused_mul_add_scalar_loop_f32(
            a.add(processed),
            scale,
            bias,
            out.add(processed),
            len - processed,
        );
    }
}

/// AVX-512 fused_mul_add_scalar for f64: out = a * scale + bias
#[target_feature(enable = "avx512f")]
pub unsafe fn fused_mul_add_scalar_f64(
    a: *const f64,
    scale: f64,
    bias: f64,
    out: *mut f64,
    len: usize,
) {
    let chunks = len / F64_LANES;
    let vscale = _mm512_set1_pd(scale);
    let vbias = _mm512_set1_pd(bias);

    for i in 0..chunks {
        let offset = i * F64_LANES;
        let va = _mm512_loadu_pd(a.add(offset));
        let result = _mm512_fmadd_pd(va, vscale, vbias);
        _mm512_storeu_pd(out.add(offset), result);
    }

    let processed = chunks * F64_LANES;
    if processed < len {
        fused_mul_add_scalar_loop_f64(
            a.add(processed),
            scale,
            bias,
            out.add(processed),
            len - processed,
        );
    }
}