primal-barracuda 0.4.0

barraCuda — BARrier-free Rust Abstracted Cross-platform Unified Dimensional Algebra. Vendor-agnostic GPU compute via WGSL.
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
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
// SPDX-License-Identifier: AGPL-3.0-or-later
// DF64 transcendental functions — f64-precision math at FP32 core speed.
//
// These functions use Df64 (f32-pair) arithmetic for all intermediate
// computations, running on the massively parallel FP32 cores instead of
// the 1:64 throttled FP64 units. Precision is ~48-bit mantissa (~14
// decimal digits), suitable for Krylov solvers, molecular dynamics,
// and lattice QCD where the extra throughput matters.
//
// Requires: df64_core.wgsl (Df64, df64_add, df64_mul, df64_div, etc.)
//
// Techniques:
//   sqrt_df64:  Newton–Raphson refinement
//   exp_df64:   Cody–Waite range reduction + degree-6 Horner
//   log_df64:   atanh-based + degree-5 Horner
//   sin_df64/cos_df64: Cody–Waite π/2 reduction + minimax kernels

// ── Constants ──

const DF64_LN2_HI: f32 = 0.6931471824645996;
const DF64_LN2_LO: f32 = -1.9046542e-9;
const DF64_LOG2E: f32 = 1.4426950408889634;
const DF64_PI_HI: f32 = 3.1415927;
const DF64_PI_LO: f32 = -8.742278e-8;
const DF64_HALF_PI_HI: f32 = 1.5707964;
const DF64_HALF_PI_LO: f32 = -4.371139e-8;
const DF64_QUARTER_PI_HI: f32 = 0.7853982;
const DF64_QUARTER_PI_LO: f32 = -3.660254e-8;

fn df64_abs(a: Df64) -> Df64 {
    if a.hi < 0.0 {
        return df64_neg(a);
    }
    return a;
}

fn df64_gt(a: Df64, b: Df64) -> bool {
    return a.hi > b.hi || (a.hi == b.hi && a.lo > b.lo);
}

fn df64_lt(a: Df64, b: Df64) -> bool {
    return a.hi < b.hi || (a.hi == b.hi && a.lo < b.lo);
}

// ── sqrt_df64: Newton–Raphson ──
// x_{n+1} = x_n/2 + a/(2*x_n)
fn sqrt_df64(a: Df64) -> Df64 {
    if a.hi <= 0.0 {
        return df64_zero();
    }
    let x0 = df64_from_f32(1.0 / sqrt(a.hi));
    // Two Newton–Raphson iterations: r = a * x0, r = r + (a - r*r) * x0 / 2
    var r = df64_mul(a, x0);
    let half = df64_from_f32(0.5);
    // Refinement: r = 0.5 * (r + a/r)
    r = df64_mul(half, df64_add(r, df64_div(a, r)));
    r = df64_mul(half, df64_add(r, df64_div(a, r)));
    return r;
}

// ── exp_df64: Cody–Waite range reduction ──
// exp(x) = 2^k * exp(r) where r = x - k*ln(2), |r| < ln(2)/2
// exp(r) ≈ 1 + r + r²/2 + r³/6 + r⁴/24 + r⁵/120 + r⁶/720
fn exp_df64(a: Df64) -> Df64 {
    if a.hi > 88.0 { return df64_from_f32(3.4028235e+38); } // overflow
    if a.hi < -87.0 { return df64_zero(); } // underflow

    // k = round(x / ln2)
    let k = i32(round(a.hi * DF64_LOG2E));
    let kf = df64_from_f32(f32(k));

    // r = x - k * ln2 (Cody-Waite two-term)
    let ln2 = Df64(DF64_LN2_HI, DF64_LN2_LO);
    let r = df64_sub(a, df64_mul(kf, ln2));

    // Horner evaluation of exp(r) - 1
    let r2 = df64_mul(r, r);
    let r3 = df64_mul(r2, r);

    let c2 = df64_from_f32(0.5);
    let c3 = df64_from_f32(0.16666666666666666);
    let c4 = df64_from_f32(0.041666666666666664);
    let c5 = df64_from_f32(0.008333333333333333);
    let c6 = df64_from_f32(0.001388888888888889);

    // p = r + r²/2 + r³(1/6 + r/24 + r²/120 + r³/720)
    var p = df64_mul(c6, r);
    p = df64_add(c5, p);
    p = df64_mul(p, r);
    p = df64_add(c4, p);
    p = df64_mul(p, r3);
    p = df64_add(df64_mul(c3, r3), p);
    p = df64_add(df64_mul(c2, r2), p);
    p = df64_add(r, p);
    p = df64_add(df64_from_f32(1.0), p);

    // Multiply by 2^k via WGSL builtin ldexp
    let scale = ldexp(1.0, k);
    return df64_scale_f32(p, scale);
}

// ── log_df64: reduction to [1, 2) then atanh series ──
// log(x) = log(m * 2^e) = e*ln(2) + log(m)
// For m in [1,2): log(m) = 2*atanh((m-1)/(m+1))
fn log_df64(a: Df64) -> Df64 {
    if a.hi <= 0.0 { return df64_from_f32(-1e38); } // -inf proxy

    // Extract exponent and mantissa via f32 frexp approximation
    var m = a.hi;
    var e = 0;
    while m >= 2.0 { m *= 0.5; e += 1; }
    while m < 1.0 { m *= 2.0; e -= 1; }

    // Now m ∈ [1, 2), recompute as DF64
    let scale_inv = ldexp(1.0, -e);  // 2^(-e)
    let mdf = df64_scale_f32(a, scale_inv);

    // s = (m - 1) / (m + 1)
    let one = df64_from_f32(1.0);
    let s = df64_div(df64_sub(mdf, one), df64_add(mdf, one));
    let s2 = df64_mul(s, s);

    // atanh(s) = s + s³/3 + s⁵/5 + s⁷/7 + s⁹/9
    let c3 = df64_from_f32(0.33333333333333333);
    let c5 = df64_from_f32(0.2);
    let c7 = df64_from_f32(0.14285714285714285);
    let c9 = df64_from_f32(0.11111111111111111);

    var p = df64_mul(c9, s2);
    p = df64_add(c7, p);
    p = df64_mul(p, s2);
    p = df64_add(c5, p);
    p = df64_mul(p, s2);
    p = df64_add(c3, p);
    p = df64_mul(p, df64_mul(s2, s));
    p = df64_add(s, p);

    // log(m) = 2 * atanh(s)
    let log_m = df64_scale_f32(p, 2.0);

    // log(x) = e * ln(2) + log(m)
    let ln2 = Df64(DF64_LN2_HI, DF64_LN2_LO);
    let e_ln2 = df64_scale_f32(ln2, f32(e));
    return df64_add(e_ln2, log_m);
}

// ── sin_df64 / cos_df64: Cody–Waite π/2 reduction + minimax ──

fn sin_kernel_df64(x: Df64) -> Df64 {
    // sin(x) ≈ x - x³/6 + x⁵/120 - x⁷/5040 for |x| < π/4
    let x2 = df64_mul(x, x);
    let c3 = df64_from_f32(-0.16666666666666666);
    let c5 = df64_from_f32(0.008333333333333333);
    let c7 = df64_from_f32(-0.0001984126984126984);

    var p = df64_mul(c7, x2);
    p = df64_add(c5, p);
    p = df64_mul(p, x2);
    p = df64_add(c3, p);
    p = df64_mul(p, df64_mul(x2, x));
    return df64_add(x, p);
}

fn cos_kernel_df64(x: Df64) -> Df64 {
    // cos(x) ≈ 1 - x²/2 + x⁴/24 - x⁶/720 for |x| < π/4
    let x2 = df64_mul(x, x);
    let c2 = df64_from_f32(-0.5);
    let c4 = df64_from_f32(0.041666666666666664);
    let c6 = df64_from_f32(-0.001388888888888889);

    var p = df64_mul(c6, x2);
    p = df64_add(c4, p);
    p = df64_mul(p, x2);
    p = df64_add(c2, p);
    p = df64_mul(p, x2);
    return df64_add(df64_from_f32(1.0), p);
}

fn sin_df64(a: Df64) -> Df64 {
    // Cody-Waite: k = round(x / (π/2)), r = x - k*(π/2)
    let k = i32(round(a.hi * 0.6366197723675814)); // 2/π
    let kf = df64_from_f32(f32(k));
    let half_pi = Df64(DF64_HALF_PI_HI, DF64_HALF_PI_LO);
    let r = df64_sub(a, df64_mul(kf, half_pi));

    let quadrant = ((k % 4) + 4) % 4;
    switch quadrant {
        case 0: { return sin_kernel_df64(r); }
        case 1: { return cos_kernel_df64(r); }
        case 2: { return df64_neg(sin_kernel_df64(r)); }
        case 3: { return df64_neg(cos_kernel_df64(r)); }
        default: { return df64_zero(); }
    }
}

fn cos_df64(a: Df64) -> Df64 {
    let k = i32(round(a.hi * 0.6366197723675814));
    let kf = df64_from_f32(f32(k));
    let half_pi = Df64(DF64_HALF_PI_HI, DF64_HALF_PI_LO);
    let r = df64_sub(a, df64_mul(kf, half_pi));

    let quadrant = ((k % 4) + 4) % 4;
    switch quadrant {
        case 0: { return cos_kernel_df64(r); }
        case 1: { return df64_neg(sin_kernel_df64(r)); }
        case 2: { return df64_neg(cos_kernel_df64(r)); }
        case 3: { return sin_kernel_df64(r); }
        default: { return df64_zero(); }
    }
}

// ── pow_df64: exp(b * log(a)) ──

fn pow_df64(base: Df64, exponent: Df64) -> Df64 {
    if base.hi <= 0.0 {
        if base.hi == 0.0 { return df64_zero(); }
        return df64_from_f32(-1e38); // NaN proxy
    }
    return exp_df64(df64_mul(exponent, log_df64(base)));
}

// ── tanh_df64: (exp(2x) - 1) / (exp(2x) + 1) ──

fn tanh_df64(a: Df64) -> Df64 {
    if a.hi > 10.0 { return df64_from_f32(1.0); }
    if a.hi < -10.0 { return df64_from_f32(-1.0); }
    let two_x = df64_scale_f32(a, 2.0);
    let e2x = exp_df64(two_x);
    let one = df64_from_f32(1.0);
    return df64_div(df64_sub(e2x, one), df64_add(e2x, one));
}

// ── atan_df64: Taylor for |x| < 0.5, argument reduction for larger ──
// atan(x) = x - x³/3 + x⁵/5 - ... for |x| < 0.5
// atan(x) = π/2 - atan(1/x) for |x| > 1
// atan(x) = π/4 + atan((x-1)/(x+1)) for 0.5 ≤ x ≤ 1
fn atan_kernel_df64(x: Df64) -> Df64 {
    // Taylor: atan(x) = x - x³/3 + x⁵/5 - x⁷/7 + ... (12 terms for ~14 digits at |x|=0.5)
    let x2 = df64_mul(x, x);
    let c3 = df64_from_f32(-0.33333333333333333);
    let c5 = df64_from_f32(0.2);
    let c7 = df64_from_f32(-0.14285714285714285);
    let c9 = df64_from_f32(0.11111111111111111);
    let c11 = df64_from_f32(-0.09090909090909091);
    let c13 = df64_from_f32(0.07692307692307693);
    let c15 = df64_from_f32(-0.06666666666666667);
    let c17 = df64_from_f32(0.05882352941176471);
    let c19 = df64_from_f32(-0.05263157894736842);
    let c21 = df64_from_f32(0.04761904761904762);
    let c23 = df64_from_f32(-0.04347826086956522);
    let c25 = df64_from_f32(0.04);

    var p = df64_mul(c25, x2);
    p = df64_add(c23, p);
    p = df64_mul(p, x2);
    p = df64_add(c21, p);
    p = df64_mul(p, x2);
    p = df64_add(c19, p);
    p = df64_mul(p, x2);
    p = df64_add(c17, p);
    p = df64_mul(p, x2);
    p = df64_add(c15, p);
    p = df64_mul(p, x2);
    p = df64_add(c13, p);
    p = df64_mul(p, x2);
    p = df64_add(c11, p);
    p = df64_mul(p, x2);
    p = df64_add(c9, p);
    p = df64_mul(p, x2);
    p = df64_add(c7, p);
    p = df64_mul(p, x2);
    p = df64_add(c5, p);
    p = df64_mul(p, x2);
    p = df64_add(c3, p);
    p = df64_mul(p, df64_mul(x2, x));
    return df64_add(x, p);
}

fn atan_df64(x: Df64) -> Df64 {
    let half = df64_from_f32(0.5);
    let one = df64_from_f32(1.0);
    let pi_half = Df64(DF64_HALF_PI_HI, DF64_HALF_PI_LO);
    let pi_quarter = Df64(DF64_QUARTER_PI_HI, DF64_QUARTER_PI_LO);

    let ax = df64_abs(x);
    if df64_gt(ax, one) {
        // |x| > 1: atan(x) = π/2 - atan(1/x), sign preserved
        let inv_x = df64_div(one, x);
        let result = atan_kernel_df64(inv_x);
        if x.hi >= 0.0 {
            return df64_sub(pi_half, result);
        } else {
            return df64_sub(df64_neg(pi_half), result);
        }
    }
    if df64_gt(ax, half) {
        // 0.5 < |x| ≤ 1: atan(x) = π/4 + atan((x-1)/(x+1))
        let t = df64_div(df64_sub(x, one), df64_add(x, one));
        let result = atan_kernel_df64(t);
        return df64_add(pi_quarter, result);
    }
    return atan_kernel_df64(x);
}

// ── asin_df64: iterative (WGSL forbids recursion) ──
// For |x| ≤ 0.5: asin(x) = atan(x / sqrt(1 - x²))
// For |x| > 0.5: asin(x) = π/2 - 2·asin(sqrt((1-|x|)/2))
//   The half-angle argument sqrt((1-|x|)/2) is always < 0.5, so the inner
//   asin hits the base case directly — no further recursion needed.
fn asin_df64(x: Df64) -> Df64 {
    let half = df64_from_f32(0.5);
    let one = df64_from_f32(1.0);
    let pi_half = Df64(DF64_HALF_PI_HI, DF64_HALF_PI_LO);

    var ax = x;
    var negate = false;
    if x.hi < 0.0 {
        ax = df64_neg(x);
        negate = true;
    }

    var result: Df64;
    if df64_gt(ax, half) {
        let s = sqrt_df64(df64_scale_f32(df64_sub(one, ax), 0.5));
        let s2 = df64_mul(s, s);
        let base = atan_df64(df64_div(s, sqrt_df64(df64_sub(one, s2))));
        result = df64_sub(pi_half, df64_scale_f32(base, 2.0));
    } else {
        let ax2 = df64_mul(ax, ax);
        result = atan_df64(df64_div(ax, sqrt_df64(df64_sub(one, ax2))));
    }

    if negate {
        return df64_neg(result);
    }
    return result;
}

// ── acos_df64: acos(x) = π/2 - asin(x) ──
fn acos_df64(x: Df64) -> Df64 {
    let pi_half = Df64(DF64_HALF_PI_HI, DF64_HALF_PI_LO);
    return df64_sub(pi_half, asin_df64(x));
}

// ── atan2_df64: four-quadrant arctangent ──
fn atan2_df64(y: Df64, x: Df64) -> Df64 {
    let zero = df64_zero();
    let pi = Df64(DF64_PI_HI, DF64_PI_LO);
    let pi_half = Df64(DF64_HALF_PI_HI, DF64_HALF_PI_LO);

    if x.hi == 0.0 && x.lo == 0.0 {
        if y.hi == 0.0 && y.lo == 0.0 { return zero; }
        if y.hi >= 0.0 { return pi_half; }
        return df64_neg(pi_half);
    }
    let q = df64_div(y, x);
    let at = atan_df64(q);
    if df64_gt(x, zero) {
        return at;
    }
    if df64_lt(y, zero) {
        return df64_sub(at, pi);
    }
    return df64_add(at, pi);
}

// ── sinh_df64: sinh(x) = (exp(x) - exp(-x)) / 2 ──
fn sinh_df64(x: Df64) -> Df64 {
    let ex = exp_df64(x);
    let emx = exp_df64(df64_neg(x));
    return df64_scale_f32(df64_sub(ex, emx), 0.5);
}

// ── cosh_df64: cosh(x) = (exp(x) + exp(-x)) / 2 ──
fn cosh_df64(x: Df64) -> Df64 {
    let ex = exp_df64(x);
    let emx = exp_df64(df64_neg(x));
    return df64_scale_f32(df64_add(ex, emx), 0.5);
}

// ── gamma_df64: Lanczos approximation ──
// Γ(z) = √(2π) · (z + g - 0.5)^(z-0.5) · e^(-(z+g-0.5)) · A_g(z)
// For x < 0.5: reflection Γ(z) = π / (sin(πz) · Γ(1-z))
const DF64_SQRT_2PI: f32 = 2.5066282746310002;
const DF64_LANCZOS_G: f32 = 7.0;

fn gamma_lanczos_df64(z: Df64) -> Df64 {
    // Lanczos coefficients p[0]..p[8]
    let p0 = df64_from_f32(0.99999999999980993);
    let p1 = df64_from_f32(676.5203681218851);
    let p2 = df64_from_f32(-1259.1392167224028);
    let p3 = df64_from_f32(771.32342877765313);
    let p4 = df64_from_f32(-176.61502916214059);
    let p5 = df64_from_f32(12.507343278686905);
    let p6 = df64_from_f32(-0.13857109526572012);
    let p7 = df64_from_f32(9.9843695780195716e-6);
    let p8 = df64_from_f32(1.5056327351493116e-7);

    let g = df64_from_f32(DF64_LANCZOS_G);
    let half = df64_from_f32(0.5);

    // z_plus = z + g - 0.5
    let z_plus = df64_sub(df64_add(z, g), half);

    // A_g(z) = p0 + p1/(z+1) + p2/(z+2) + ... + p8/(z+8)
    var ag = df64_div(p1, df64_add(z, df64_from_f32(1.0)));
    ag = df64_add(ag, df64_div(p2, df64_add(z, df64_from_f32(2.0))));
    ag = df64_add(ag, df64_div(p3, df64_add(z, df64_from_f32(3.0))));
    ag = df64_add(ag, df64_div(p4, df64_add(z, df64_from_f32(4.0))));
    ag = df64_add(ag, df64_div(p5, df64_add(z, df64_from_f32(5.0))));
    ag = df64_add(ag, df64_div(p6, df64_add(z, df64_from_f32(6.0))));
    ag = df64_add(ag, df64_div(p7, df64_add(z, df64_from_f32(7.0))));
    ag = df64_add(ag, df64_div(p8, df64_add(z, df64_from_f32(8.0))));
    ag = df64_add(p0, ag);

    // Γ(z) = √(2π) · z_plus^(z-0.5) · e^(-z_plus) · A_g(z)
    let exp_z_minus_half = df64_sub(z, half);
    let base_pow = pow_df64(z_plus, exp_z_minus_half);
    let exp_neg = exp_df64(df64_neg(z_plus));

    var result = df64_mul(df64_from_f32(DF64_SQRT_2PI), base_pow);
    result = df64_mul(result, exp_neg);
    result = df64_mul(result, ag);
    return result;
}

fn gamma_df64(x: Df64) -> Df64 {
    let half = df64_from_f32(0.5);
    let one = df64_from_f32(1.0);
    let pi = Df64(DF64_PI_HI, DF64_PI_LO);

    if df64_lt(x, half) {
        // Reflection: Γ(x) = π / (sin(πx) · Γ(1-x))
        let one_minus_x = df64_sub(one, x);
        let gamma_1_minus_x = gamma_lanczos_df64(one_minus_x);
        let pi_x = df64_mul(pi, x);
        let sin_pi_x = sin_df64(pi_x);
        return df64_div(pi, df64_mul(sin_pi_x, gamma_1_minus_x));
    }
    return gamma_lanczos_df64(x);
}

// ── erf_df64: Abramowitz & Stegun 7.1.26 rational approximation ──
// erf(x) = 1 - (a₁t + a₂t² + a₃t³ + a₄t⁴ + a₅t⁵) · exp(-x²), t = 1/(1 + 0.3275911·|x|)
// For negative x: erf(-x) = -erf(x) — handle via |x| then negate result
fn erf_df64(x: Df64) -> Df64 {
    let one = df64_from_f32(1.0);
    let negate = x.hi < 0.0;
    let ax = df64_abs(x);

    // t = 1 / (1 + 0.3275911 * x)
    let t_denom = df64_add(one, df64_scale_f32(ax, 0.3275911));
    let t = df64_div(one, t_denom);

    // Horner: a₁t + a₂t² + a₃t³ + a₄t⁴ + a₅t⁵
    let a1 = df64_from_f32(0.254829592);
    let a2 = df64_from_f32(-0.284496736);
    let a3 = df64_from_f32(1.421413741);
    let a4 = df64_from_f32(-1.453152027);
    let a5 = df64_from_f32(1.061405429);

    let t2 = df64_mul(t, t);
    let t3 = df64_mul(t2, t);
    let t4 = df64_mul(t3, t);
    let t5 = df64_mul(t4, t);

    var p = df64_mul(a5, t5);
    p = df64_add(df64_mul(a4, t4), p);
    p = df64_add(df64_mul(a3, t3), p);
    p = df64_add(df64_mul(a2, t2), p);
    p = df64_add(df64_mul(a1, t), p);

    // exp(-x²)
    let x2 = df64_mul(ax, ax);
    let exp_neg_x2 = exp_df64(df64_neg(x2));

    // erf(x) = 1 - p · exp(-x²)
    var result = df64_sub(one, df64_mul(p, exp_neg_x2));
    if negate {
        result = df64_neg(result);
    }
    return result;
}