simdeez 3.0.1

SIMD library to abstract over different instruction sets and widths
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
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341

use crate::math::scalar;
use crate::{Simd, SimdBaseIo, SimdBaseOps, SimdConsts, SimdFloat32, SimdInt, SimdInt32};

pub(super) type SimdI32<V> = <<V as SimdConsts>::Engine as Simd>::Vi32;

pub(super) const F32_EXPONENT_MASK: i32 = 0x7F80_0000u32 as i32;
pub(super) const F32_MANTISSA_MASK: i32 = 0x007F_FFFF;
pub(super) const F32_LOG_NORM_MANTISSA: i32 = 0x3F00_0000;
pub(super) const F32_EXPONENT_BIAS_ADJUST: i32 = 126;

// DECISION(2026-03-23): KEEP_SIMD_PORTABLE
// Function(s): f32 log2_u35 portable fallback / exp2_u35
// Why kept:
// - local benches show both kernels materially ahead of native scalar on this host
// - scalar patching already contains non-finite, zero, and subnormal edge lanes
// Revisit when:
// - a new approximation family lands or non-x86 evidence disagrees sharply

// DECISION(2026-03-23): KEEP_SIMD_PORTABLE
// Function(s): f32 sin_u35 / cos_u35 / tan_u35
// Why kept:
// - runtime-selected throughput is far above native scalar on the local machine
// - targeted boundary and mixed-lane tests cover current reduction and tan-pole handling
// Revisit when:
// - large-argument reduction strategy changes materially

// DECISION(2026-03-23): KEEP_SIMD_PORTABLE
// Function(s): f32 asinh_u35 / acosh_u35 / atanh_u35
// Why kept:
// - the restored inverse-hyperbolic paths beat native scalar in local benchmarks
// - exceptional-domain lanes already fall back to scalar references
// Revisit when:
// - the shared log/exp kernels change enough to affect the current balance

#[inline(always)]
fn any_lane_nonzero<V>(mask: SimdI32<V>) -> bool
where
    V: SimdFloat32,
    V::Engine: Simd<Vf32 = V>,
{
    unsafe {
        let lanes = mask.as_array();
        for lane in 0..V::WIDTH {
            if lanes[lane] != 0 {
                return true;
            }
        }
    }

    false
}

#[inline(always)]
pub(super) fn log2_exceptional_mask<V>(input: V) -> SimdI32<V>
where
    V: SimdFloat32,
    V::Engine: Simd<Vf32 = V>,
{
    let bits = input.bitcast_i32();
    let exponent_bits = bits & F32_EXPONENT_MASK;

    let non_positive = input
        .cmp_gt(V::zeroes())
        .bitcast_i32()
        .cmp_eq(SimdI32::<V>::zeroes());
    let subnormal_or_zero = exponent_bits.cmp_eq(SimdI32::<V>::zeroes());
    let inf_or_nan = exponent_bits.cmp_eq(SimdI32::<V>::set1(F32_EXPONENT_MASK));

    non_positive | subnormal_or_zero | inf_or_nan
}

#[inline(always)]
pub(super) fn patch_exceptional_lanes<V>(
    input: V,
    output: V,
    exceptional_mask: SimdI32<V>,
    scalar_fallback: fn(f32) -> f32,
) -> V
where
    V: SimdFloat32,
    V::Engine: Simd<Vf32 = V>,
{
    if !any_lane_nonzero::<V>(exceptional_mask) {
        return output;
    }

    unsafe {
        let input_lanes = input.as_array();
        let mask_lanes = exceptional_mask.as_array();
        let mut output_lanes = output.as_array();

        for lane in 0..V::WIDTH {
            if mask_lanes[lane] != 0 {
                output_lanes[lane] = scalar_fallback(input_lanes[lane]);
            }
        }

        V::load_from_ptr_unaligned(&output_lanes as *const V::ArrayRepresentation as *const f32)
    }
}

#[inline(always)]
pub(super) fn log2_u35<V>(input: V) -> V
where
    V: SimdFloat32,
    V::Engine: Simd<Vf32 = V>,
{
    let bits = input.bitcast_i32();
    let exponent_bits = bits & F32_EXPONENT_MASK;
    let mantissa_bits = bits & F32_MANTISSA_MASK;

    let exceptional_mask = log2_exceptional_mask(input);

    let exponent = (exponent_bits.shr(23) - F32_EXPONENT_BIAS_ADJUST).cast_f32();
    let normalized_mantissa = (mantissa_bits | F32_LOG_NORM_MANTISSA).bitcast_f32();

    let one = V::set1(1.0);
    let half = V::set1(0.5);
    let sqrt_half = V::set1(core::f32::consts::FRAC_1_SQRT_2);

    let adjust_mask = normalized_mantissa.cmp_lt(sqrt_half);
    let exponent = exponent - adjust_mask.blendv(V::zeroes(), one);
    let reduced = adjust_mask.blendv(
        normalized_mantissa - one,
        (normalized_mantissa + normalized_mantissa) - one,
    );

    let reduced_sq = reduced * reduced;

    let mut poly = V::set1(7.037_683_6e-2);
    poly = (poly * reduced) + V::set1(-1.151_461e-1);
    poly = (poly * reduced) + V::set1(1.167_699_9e-1);
    poly = (poly * reduced) + V::set1(-1.242_014_1e-1);
    poly = (poly * reduced) + V::set1(1.424_932_3e-1);
    poly = (poly * reduced) + V::set1(-1.666_805_8e-1);
    poly = (poly * reduced) + V::set1(2.000_071_5e-1);
    poly = (poly * reduced) + V::set1(-2.499_999_4e-1);
    poly = (poly * reduced) + V::set1(3.333_333e-1);

    poly *= reduced;
    poly *= reduced_sq;
    poly += exponent * V::set1(-2.121_944_4e-4);
    poly -= half * reduced_sq;

    let ln_x = reduced + poly + (exponent * V::set1(0.693_359_4));
    let fast = ln_x * V::set1(core::f32::consts::LOG2_E);

    patch_exceptional_lanes(input, fast, exceptional_mask, scalar::log2_u35_f32)
}

#[inline(always)]
pub(super) fn exp2_u35<V>(input: V) -> V
where
    V: SimdFloat32,
    V::Engine: Simd<Vf32 = V>,
{
    let finite_mask = input.cmp_eq(input).bitcast_i32();
    let in_lower_bound = input.cmp_gte(V::set1(-126.0)).bitcast_i32();
    let in_upper_bound = input.cmp_lte(V::set1(126.0)).bitcast_i32();
    let fast_mask = finite_mask & in_lower_bound & in_upper_bound;
    let exceptional_mask = fast_mask.cmp_eq(SimdI32::<V>::zeroes());

    let integral = input.floor().cast_i32();
    let fractional = input - integral.cast_f32();
    let reduced = fractional * V::set1(core::f32::consts::LN_2);

    let mut poly = V::set1(1.987_569_1e-4);
    poly = (poly * reduced) + V::set1(1.398_2e-3);
    poly = (poly * reduced) + V::set1(8.333_452e-3);
    poly = (poly * reduced) + V::set1(4.166_579_6e-2);
    poly = (poly * reduced) + V::set1(1.666_666_5e-1);
    poly = (poly * reduced) + V::set1(5e-1);

    let reduced_sq = reduced * reduced;
    let exp_reduced = (poly * reduced_sq) + reduced + V::set1(1.0);

    let exp_bits = (integral + 127).shl(23);
    let scale = exp_bits.bitcast_f32();
    let fast = exp_reduced * scale;

    patch_exceptional_lanes(input, fast, exceptional_mask, scalar::exp2_u35_f32)
}

#[inline(always)]
fn trig_exceptional_mask<V>(input: V) -> SimdI32<V>
where
    V: SimdFloat32,
    V::Engine: Simd<Vf32 = V>,
{
    let finite_mask = input.cmp_eq(input).bitcast_i32();
    let within_fast_range = input.abs().cmp_lte(V::set1(8192.0)).bitcast_i32();
    let non_zero_mask = input.cmp_neq(V::zeroes()).bitcast_i32();
    let fast_mask = finite_mask & within_fast_range & non_zero_mask;
    fast_mask.cmp_eq(SimdI32::<V>::zeroes())
}

#[inline(always)]
fn sin_cos_fast<V>(input: V) -> (V, V)
where
    V: SimdFloat32,
    V::Engine: Simd<Vf32 = V>,
{
    let two_over_pi = V::set1(core::f32::consts::FRAC_2_PI);
    let n = (input * two_over_pi).round().cast_i32();

    let n_f = n.cast_f32();
    let r = ((input - n_f * V::set1(1.570_312_5)) - n_f * V::set1(4.837_513e-4))
        - n_f * V::set1(7.549_789_4e-8);
    let r2 = r * r;

    let sin_poly = (((V::set1(-2.388_985_9e-8) * r2 + V::set1(2.752_556_2e-6)) * r2
        + V::set1(-1.984_127e-4))
        * r2
        + V::set1(8.333_331e-3))
        * r2
        + V::set1(-1.666_666_7e-1);
    let sin_r = ((sin_poly * r2) * r) + r;

    let cos_poly = (((V::set1(-2.605_161_5e-7) * r2 + V::set1(2.476_049_5e-5)) * r2
        + V::set1(-1.388_837_8e-3))
        * r2
        + V::set1(4.166_664_6e-2))
        * r2
        + V::set1(-5e-1);
    let cos_r = (cos_poly * r2) + V::set1(1.0);

    let q = n & SimdI32::<V>::set1(3);
    let q0 = q.cmp_eq(SimdI32::<V>::zeroes()).bitcast_f32();
    let q1 = q.cmp_eq(SimdI32::<V>::set1(1)).bitcast_f32();
    let q2 = q.cmp_eq(SimdI32::<V>::set1(2)).bitcast_f32();

    let mut sin_out = q0.blendv(V::zeroes(), sin_r);
    sin_out = q1.blendv(sin_out, cos_r);
    sin_out = q2.blendv(sin_out, -sin_r);
    sin_out = (q0 | q1 | q2).cmp_eq(V::zeroes()).blendv(sin_out, -cos_r);

    let mut cos_out = q0.blendv(V::zeroes(), cos_r);
    cos_out = q1.blendv(cos_out, -sin_r);
    cos_out = q2.blendv(cos_out, -cos_r);
    cos_out = (q0 | q1 | q2).cmp_eq(V::zeroes()).blendv(cos_out, sin_r);

    (sin_out, cos_out)
}

#[inline(always)]
pub(super) fn sin_u35<V>(input: V) -> V
where
    V: SimdFloat32,
    V::Engine: Simd<Vf32 = V>,
{
    let exceptional_mask = trig_exceptional_mask(input);
    let (sin_fast, _) = sin_cos_fast(input);
    patch_exceptional_lanes(input, sin_fast, exceptional_mask, scalar::sin_u35_f32)
}

#[inline(always)]
pub(super) fn cos_u35<V>(input: V) -> V
where
    V: SimdFloat32,
    V::Engine: Simd<Vf32 = V>,
{
    let exceptional_mask = trig_exceptional_mask(input);
    let (_, cos_fast) = sin_cos_fast(input);
    patch_exceptional_lanes(input, cos_fast, exceptional_mask, scalar::cos_u35_f32)
}

#[inline(always)]
pub(super) fn tan_u35<V>(input: V) -> V
where
    V: SimdFloat32,
    V::Engine: Simd<Vf32 = V>,
{
    let base_exceptional = trig_exceptional_mask(input);
    let (sin_fast, cos_fast) = sin_cos_fast(input);
    let dangerous = cos_fast.abs().cmp_lt(V::set1(1.0e-4)).bitcast_i32();
    let exceptional_mask = base_exceptional | dangerous;
    let fast = sin_fast / cos_fast;
    patch_exceptional_lanes(input, fast, exceptional_mask, scalar::tan_u35_f32)
}

#[inline(always)]
pub(super) fn asinh_u35<V>(input: V) -> V
where
    V: SimdFloat32,
    V::Engine: Simd<Vf32 = V>,
{
    let finite_mask = input.cmp_eq(input).bitcast_i32();
    let abs_x = input.abs();
    let tiny_mask = abs_x.cmp_lt(V::set1(0.05)).bitcast_i32();
    let large_mask = abs_x.cmp_gt(V::set1(1.0e19)).bitcast_i32();
    let zero_mask = input.cmp_eq(V::zeroes()).bitcast_i32();
    let exceptional_mask =
        finite_mask.cmp_eq(SimdI32::<V>::zeroes()) | tiny_mask | large_mask | zero_mask;

    let radicand = (abs_x * abs_x) + V::set1(1.0);
    let magnitude = log2_u35(abs_x + radicand.sqrt()) * V::set1(core::f32::consts::LN_2);
    let negative_mask = input.cmp_lt(V::zeroes());
    let fast = negative_mask.blendv(magnitude, -magnitude);

    patch_exceptional_lanes(input, fast, exceptional_mask, scalar::asinh_u35_f32)
}

#[inline(always)]
pub(super) fn acosh_u35<V>(input: V) -> V
where
    V: SimdFloat32,
    V::Engine: Simd<Vf32 = V>,
{
    let finite_mask = input.cmp_eq(input).bitcast_i32();
    let in_domain_mask = input.cmp_gte(V::set1(1.0)).bitcast_i32();
    let fast_mask = finite_mask & in_domain_mask;
    let exceptional_mask = fast_mask.cmp_eq(SimdI32::<V>::zeroes());

    let root_term = ((input - V::set1(1.0)).sqrt()) * ((input + V::set1(1.0)).sqrt());
    let fast = log2_u35(input + root_term) * V::set1(core::f32::consts::LN_2);

    patch_exceptional_lanes(input, fast, exceptional_mask, scalar::acosh_u35_f32)
}

#[inline(always)]
pub(super) fn atanh_u35<V>(input: V) -> V
where
    V: SimdFloat32,
    V::Engine: Simd<Vf32 = V>,
{
    let finite_mask = input.cmp_eq(input).bitcast_i32();
    let abs_x = input.abs();
    let strict_domain_mask = abs_x.cmp_lt(V::set1(1.0)).bitcast_i32();
    let non_zero_mask = input.cmp_neq(V::zeroes()).bitcast_i32();
    let stable_range_mask = abs_x.cmp_lte(V::set1(0.75)).bitcast_i32();
    let away_from_zero_mask = abs_x.cmp_gte(V::set1(0.05)).bitcast_i32();
    let fast_mask =
        finite_mask & strict_domain_mask & non_zero_mask & stable_range_mask & away_from_zero_mask;
    let exceptional_mask = fast_mask.cmp_eq(SimdI32::<V>::zeroes());

    let one = V::set1(1.0);
    let ratio = (one + input) / (one - input);
    let fast = log2_u35(ratio) * V::set1(0.5 * core::f32::consts::LN_2);

    patch_exceptional_lanes(input, fast, exceptional_mask, scalar::atanh_u35_f32)
}