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
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266

//! AVX2 fused activation-mul kernels
//!
//! Vectorized implementations of fused activation * multiplication using 256-bit registers.
//! Functions take two inputs (a, b) and compute activation(a) * b in a single pass.

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

use super::super::math::avx2::{exp_f32, exp_f64, tanh_f32, tanh_f64};
use super::{
    gelu_mul_scalar_f32, gelu_mul_scalar_f64, relu_mul_scalar_f32, relu_mul_scalar_f64,
    sigmoid_mul_scalar_f32, sigmoid_mul_scalar_f64, silu_mul_scalar_f32, silu_mul_scalar_f64,
};

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

/// AVX2 silu_mul for f32
///
/// Computes: (a / (1 + exp(-a))) * b
#[target_feature(enable = "avx2", enable = "fma")]
pub unsafe fn silu_mul_f32(a: *const f32, b: *const f32, out: *mut f32, len: usize) {
    let chunks = len / F32_LANES;
    let one = _mm256_set1_ps(1.0);

    for c in 0..chunks {
        let offset = c * F32_LANES;
        let x = _mm256_loadu_ps(a.add(offset));
        let y = _mm256_loadu_ps(b.add(offset));
        let neg_x = _mm256_sub_ps(_mm256_setzero_ps(), x);
        let exp_neg_x = exp_f32(neg_x);
        let activation = _mm256_div_ps(x, _mm256_add_ps(one, exp_neg_x));
        let result = _mm256_mul_ps(activation, y);
        _mm256_storeu_ps(out.add(offset), result);
    }

    let processed = chunks * F32_LANES;
    if processed < len {
        silu_mul_scalar_f32(
            a.add(processed),
            b.add(processed),
            out.add(processed),
            len - processed,
        );
    }
}

/// AVX2 silu_mul for f64
///
/// Computes: (a / (1 + exp(-a))) * b
#[target_feature(enable = "avx2", enable = "fma")]
pub unsafe fn silu_mul_f64(a: *const f64, b: *const f64, out: *mut f64, len: usize) {
    let chunks = len / F64_LANES;
    let one = _mm256_set1_pd(1.0);

    for c in 0..chunks {
        let offset = c * F64_LANES;
        let x = _mm256_loadu_pd(a.add(offset));
        let y = _mm256_loadu_pd(b.add(offset));
        let neg_x = _mm256_sub_pd(_mm256_setzero_pd(), x);
        let exp_neg_x = exp_f64(neg_x);
        let activation = _mm256_div_pd(x, _mm256_add_pd(one, exp_neg_x));
        let result = _mm256_mul_pd(activation, y);
        _mm256_storeu_pd(out.add(offset), result);
    }

    let processed = chunks * F64_LANES;
    if processed < len {
        silu_mul_scalar_f64(
            a.add(processed),
            b.add(processed),
            out.add(processed),
            len - processed,
        );
    }
}

/// AVX2 gelu_mul for f32
///
/// Computes: 0.5 * a * (1 + tanh(sqrt(2/pi) * (a + 0.044715 * a^3))) * b
#[target_feature(enable = "avx2", enable = "fma")]
pub unsafe fn gelu_mul_f32(a: *const f32, b: *const f32, out: *mut f32, len: usize) {
    let chunks = len / F32_LANES;
    let half = _mm256_set1_ps(0.5);
    let one = _mm256_set1_ps(1.0);
    let sqrt_2_over_pi = _mm256_set1_ps(0.7978845608);
    let tanh_coef = _mm256_set1_ps(0.044715);

    for c in 0..chunks {
        let offset = c * F32_LANES;
        let x = _mm256_loadu_ps(a.add(offset));
        let y = _mm256_loadu_ps(b.add(offset));

        let x_cubed = _mm256_mul_ps(_mm256_mul_ps(x, x), x);
        let inner = _mm256_mul_ps(sqrt_2_over_pi, _mm256_fmadd_ps(tanh_coef, x_cubed, x));

        let tanh_inner = tanh_f32(inner);
        let activation = _mm256_mul_ps(half, _mm256_mul_ps(x, _mm256_add_ps(one, tanh_inner)));

        let result = _mm256_mul_ps(activation, y);
        _mm256_storeu_ps(out.add(offset), result);
    }

    let processed = chunks * F32_LANES;
    if processed < len {
        gelu_mul_scalar_f32(
            a.add(processed),
            b.add(processed),
            out.add(processed),
            len - processed,
        );
    }
}

/// AVX2 gelu_mul for f64
///
/// Computes: 0.5 * a * (1 + tanh(sqrt(2/pi) * (a + 0.044715 * a^3))) * b
#[target_feature(enable = "avx2", enable = "fma")]
pub unsafe fn gelu_mul_f64(a: *const f64, b: *const f64, out: *mut f64, len: usize) {
    let chunks = len / F64_LANES;
    let half = _mm256_set1_pd(0.5);
    let one = _mm256_set1_pd(1.0);
    let sqrt_2_over_pi = _mm256_set1_pd(0.7978845608028654);
    let tanh_coef = _mm256_set1_pd(0.044715);

    for c in 0..chunks {
        let offset = c * F64_LANES;
        let x = _mm256_loadu_pd(a.add(offset));
        let y = _mm256_loadu_pd(b.add(offset));

        let x_cubed = _mm256_mul_pd(_mm256_mul_pd(x, x), x);
        let inner = _mm256_mul_pd(sqrt_2_over_pi, _mm256_fmadd_pd(tanh_coef, x_cubed, x));

        let tanh_inner = tanh_f64(inner);
        let activation = _mm256_mul_pd(half, _mm256_mul_pd(x, _mm256_add_pd(one, tanh_inner)));

        let result = _mm256_mul_pd(activation, y);
        _mm256_storeu_pd(out.add(offset), result);
    }

    let processed = chunks * F64_LANES;
    if processed < len {
        gelu_mul_scalar_f64(
            a.add(processed),
            b.add(processed),
            out.add(processed),
            len - processed,
        );
    }
}

/// AVX2 relu_mul for f32
///
/// Computes: max(0, a) * b
#[target_feature(enable = "avx2", enable = "fma")]
pub unsafe fn relu_mul_f32(a: *const f32, b: *const f32, out: *mut f32, len: usize) {
    let chunks = len / F32_LANES;
    let zero = _mm256_setzero_ps();

    for c in 0..chunks {
        let offset = c * F32_LANES;
        let x = _mm256_loadu_ps(a.add(offset));
        let y = _mm256_loadu_ps(b.add(offset));
        let activation = _mm256_max_ps(zero, x);
        let result = _mm256_mul_ps(activation, y);
        _mm256_storeu_ps(out.add(offset), result);
    }

    let processed = chunks * F32_LANES;
    if processed < len {
        relu_mul_scalar_f32(
            a.add(processed),
            b.add(processed),
            out.add(processed),
            len - processed,
        );
    }
}

/// AVX2 relu_mul for f64
///
/// Computes: max(0, a) * b
#[target_feature(enable = "avx2", enable = "fma")]
pub unsafe fn relu_mul_f64(a: *const f64, b: *const f64, out: *mut f64, len: usize) {
    let chunks = len / F64_LANES;
    let zero = _mm256_setzero_pd();

    for c in 0..chunks {
        let offset = c * F64_LANES;
        let x = _mm256_loadu_pd(a.add(offset));
        let y = _mm256_loadu_pd(b.add(offset));
        let activation = _mm256_max_pd(zero, x);
        let result = _mm256_mul_pd(activation, y);
        _mm256_storeu_pd(out.add(offset), result);
    }

    let processed = chunks * F64_LANES;
    if processed < len {
        relu_mul_scalar_f64(
            a.add(processed),
            b.add(processed),
            out.add(processed),
            len - processed,
        );
    }
}

/// AVX2 sigmoid_mul for f32
///
/// Computes: (1 / (1 + exp(-a))) * b
#[target_feature(enable = "avx2", enable = "fma")]
pub unsafe fn sigmoid_mul_f32(a: *const f32, b: *const f32, out: *mut f32, len: usize) {
    let chunks = len / F32_LANES;
    let one = _mm256_set1_ps(1.0);

    for c in 0..chunks {
        let offset = c * F32_LANES;
        let x = _mm256_loadu_ps(a.add(offset));
        let y = _mm256_loadu_ps(b.add(offset));
        let neg_x = _mm256_sub_ps(_mm256_setzero_ps(), x);
        let exp_neg_x = exp_f32(neg_x);
        let activation = _mm256_div_ps(one, _mm256_add_ps(one, exp_neg_x));
        let result = _mm256_mul_ps(activation, y);
        _mm256_storeu_ps(out.add(offset), result);
    }

    let processed = chunks * F32_LANES;
    if processed < len {
        sigmoid_mul_scalar_f32(
            a.add(processed),
            b.add(processed),
            out.add(processed),
            len - processed,
        );
    }
}

/// AVX2 sigmoid_mul for f64
///
/// Computes: (1 / (1 + exp(-a))) * b
#[target_feature(enable = "avx2", enable = "fma")]
pub unsafe fn sigmoid_mul_f64(a: *const f64, b: *const f64, out: *mut f64, len: usize) {
    let chunks = len / F64_LANES;
    let one = _mm256_set1_pd(1.0);

    for c in 0..chunks {
        let offset = c * F64_LANES;
        let x = _mm256_loadu_pd(a.add(offset));
        let y = _mm256_loadu_pd(b.add(offset));
        let neg_x = _mm256_sub_pd(_mm256_setzero_pd(), x);
        let exp_neg_x = exp_f64(neg_x);
        let activation = _mm256_div_pd(one, _mm256_add_pd(one, exp_neg_x));
        let result = _mm256_mul_pd(activation, y);
        _mm256_storeu_pd(out.add(offset), result);
    }

    let processed = chunks * F64_LANES;
    if processed < len {
        sigmoid_mul_scalar_f64(
            a.add(processed),
            b.add(processed),
            out.add(processed),
            len - processed,
        );
    }
}