Eigen  5.0.1
 
Loading...
Searching...
No Matches
MathFunctions.h
1// This file is part of Eigen, a lightweight C++ template library
2// for linear algebra.
3//
4// Copyright (C) 2007 Julien Pommier
5// Copyright (C) 2014 Pedro Gonnet (pedro.gonnet@gmail.com)
6// Copyright (C) 2016 Gael Guennebaud <gael.guennebaud@inria.fr>
7//
8// Copyright (C) 2018 Wave Computing, Inc.
9// Written by:
10// Chris Larsen
11// Alexey Frunze (afrunze@wavecomp.com)
12//
13// This Source Code Form is subject to the terms of the Mozilla
14// Public License v. 2.0. If a copy of the MPL was not distributed
15// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
16// SPDX-License-Identifier: MPL-2.0
17
18/* The sin, cos, exp, and log functions of this file come from
19 * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/
20 */
21
22/* The tanh function of this file is an adaptation of
23 * template<typename T> T generic_fast_tanh_float(const T&)
24 * from MathFunctionsImpl.h.
25 */
26
27#ifndef EIGEN_MATH_FUNCTIONS_MSA_H
28#define EIGEN_MATH_FUNCTIONS_MSA_H
29
30// IWYU pragma: private
31#include "../../InternalHeaderCheck.h"
32
33namespace Eigen {
34
35namespace internal {
36
37template <>
38EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet4f plog<Packet4f>(const Packet4f& _x) {
39 static EIGEN_DECLARE_CONST_Packet4f(cephes_SQRTHF, 0.707106781186547524f);
40 static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p0, 7.0376836292e-2f);
41 static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p1, -1.1514610310e-1f);
42 static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p2, 1.1676998740e-1f);
43 static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p3, -1.2420140846e-1f);
44 static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p4, +1.4249322787e-1f);
45 static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p5, -1.6668057665e-1f);
46 static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p6, +2.0000714765e-1f);
47 static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p7, -2.4999993993e-1f);
48 static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p8, +3.3333331174e-1f);
49 static EIGEN_DECLARE_CONST_Packet4f(cephes_log_q1, -2.12194440e-4f);
50 static EIGEN_DECLARE_CONST_Packet4f(cephes_log_q2, 0.693359375f);
51 static EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
52 static EIGEN_DECLARE_CONST_Packet4f(1, 1.0f);
53
54 // Convert negative argument into NAN (quiet negative, to be specific).
55 Packet4f zero = (Packet4f)__builtin_msa_ldi_w(0);
56 Packet4i neg_mask = __builtin_msa_fclt_w(_x, zero);
57 Packet4i zero_mask = __builtin_msa_fceq_w(_x, zero);
58 Packet4f non_neg_x_or_nan = padd(_x, (Packet4f)neg_mask); // Add 0.0 or NAN.
59 Packet4f x = non_neg_x_or_nan;
60
61 // Extract exponent from x = mantissa * 2**exponent, where 1.0 <= mantissa < 2.0.
62 // N.B. the exponent is one less of what frexpf() would return.
63 Packet4i e_int = __builtin_msa_ftint_s_w(__builtin_msa_flog2_w(x));
64 // Multiply x by 2**(-exponent-1) to get 0.5 <= x < 1.0 as from frexpf().
65 x = __builtin_msa_fexp2_w(x, (Packet4i)__builtin_msa_nori_b((v16u8)e_int, 0));
66
67 /*
68 if (x < SQRTHF) {
69 x = x + x - 1.0;
70 } else {
71 e += 1;
72 x = x - 1.0;
73 }
74 */
75 Packet4f xx = padd(x, x);
76 Packet4i ge_mask = __builtin_msa_fcle_w(p4f_cephes_SQRTHF, x);
77 e_int = psub(e_int, ge_mask);
78 x = (Packet4f)__builtin_msa_bsel_v((v16u8)ge_mask, (v16u8)xx, (v16u8)x);
79 x = psub(x, p4f_1);
80 Packet4f e = __builtin_msa_ffint_s_w(e_int);
81
82 Packet4f x2 = pmul(x, x);
83 Packet4f x3 = pmul(x2, x);
84
85 Packet4f y, y1, y2;
86 y = pmadd(p4f_cephes_log_p0, x, p4f_cephes_log_p1);
87 y1 = pmadd(p4f_cephes_log_p3, x, p4f_cephes_log_p4);
88 y2 = pmadd(p4f_cephes_log_p6, x, p4f_cephes_log_p7);
89 y = pmadd(y, x, p4f_cephes_log_p2);
90 y1 = pmadd(y1, x, p4f_cephes_log_p5);
91 y2 = pmadd(y2, x, p4f_cephes_log_p8);
92 y = pmadd(y, x3, y1);
93 y = pmadd(y, x3, y2);
94 y = pmul(y, x3);
95
96 y = pmadd(e, p4f_cephes_log_q1, y);
97 x = __builtin_msa_fmsub_w(x, x2, p4f_half);
98 x = padd(x, y);
99 x = pmadd(e, p4f_cephes_log_q2, x);
100
101 // x is now the logarithm result candidate. We still need to handle the
102 // extreme arguments of zero and positive infinity, though.
103 // N.B. if the argument is +INFINITY, x is NAN because the polynomial terms
104 // contain infinities of both signs (see the coefficients and code above).
105 // INFINITY - INFINITY is NAN.
106
107 // If the argument is +INFINITY, make it the new result candidate.
108 // To achieve that we choose the smaller of the result candidate and the
109 // argument.
110 // This is correct for all finite pairs of values (the logarithm is smaller
111 // than the argument).
112 // This is also correct in the special case when the argument is +INFINITY
113 // and the result candidate is NAN. This is because the fmin.df instruction
114 // prefers non-NANs to NANs.
115 x = __builtin_msa_fmin_w(x, non_neg_x_or_nan);
116
117 // If the argument is zero (including -0.0), the result becomes -INFINITY.
118 Packet4i neg_infs = __builtin_msa_slli_w(zero_mask, 23);
119 x = (Packet4f)__builtin_msa_bsel_v((v16u8)zero_mask, (v16u8)x, (v16u8)neg_infs);
120
121 return x;
122}
123
124template <>
125EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet4f pexp<Packet4f>(const Packet4f& _x) {
126 // Limiting single-precision pexp's argument to [-128, +128] lets pexp
127 // reach 0 and INFINITY naturally.
128 static EIGEN_DECLARE_CONST_Packet4f(exp_lo, -128.0f);
129 static EIGEN_DECLARE_CONST_Packet4f(exp_hi, +128.0f);
130 static EIGEN_DECLARE_CONST_Packet4f(cephes_LOG2EF, 1.44269504088896341f);
131 static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C1, 0.693359375f);
132 static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C2, -2.12194440e-4f);
133 static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p0, 1.9875691500e-4f);
134 static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p1, 1.3981999507e-3f);
135 static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p2, 8.3334519073e-3f);
136 static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p3, 4.1665795894e-2f);
137 static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459e-1f);
138 static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201e-1f);
139 static EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
140 static EIGEN_DECLARE_CONST_Packet4f(1, 1.0f);
141
142 Packet4f x = _x;
143
144 // Clamp x.
145 x = (Packet4f)__builtin_msa_bsel_v((v16u8)__builtin_msa_fclt_w(x, p4f_exp_lo), (v16u8)x, (v16u8)p4f_exp_lo);
146 x = (Packet4f)__builtin_msa_bsel_v((v16u8)__builtin_msa_fclt_w(p4f_exp_hi, x), (v16u8)x, (v16u8)p4f_exp_hi);
147
148 // Round to nearest integer by adding 0.5 (with x's sign) and truncating.
149 Packet4f x2_add = (Packet4f)__builtin_msa_binsli_w((v4u32)p4f_half, (v4u32)x, 0);
150 Packet4f x2 = pmadd(x, p4f_cephes_LOG2EF, x2_add);
151 Packet4i x2_int = __builtin_msa_ftrunc_s_w(x2);
152 Packet4f x2_int_f = __builtin_msa_ffint_s_w(x2_int);
153
154 x = __builtin_msa_fmsub_w(x, x2_int_f, p4f_cephes_exp_C1);
155 x = __builtin_msa_fmsub_w(x, x2_int_f, p4f_cephes_exp_C2);
156
157 Packet4f z = pmul(x, x);
158
159 Packet4f y = p4f_cephes_exp_p0;
160 y = pmadd(y, x, p4f_cephes_exp_p1);
161 y = pmadd(y, x, p4f_cephes_exp_p2);
162 y = pmadd(y, x, p4f_cephes_exp_p3);
163 y = pmadd(y, x, p4f_cephes_exp_p4);
164 y = pmadd(y, x, p4f_cephes_exp_p5);
165 y = pmadd(y, z, x);
166 y = padd(y, p4f_1);
167
168 // y *= 2**exponent.
169 y = __builtin_msa_fexp2_w(y, x2_int);
170
171 return y;
172}
173
174template <>
175EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet4f ptanh<Packet4f>(const Packet4f& _x) {
176 static EIGEN_DECLARE_CONST_Packet4f(tanh_tiny, 1e-4f);
177 static EIGEN_DECLARE_CONST_Packet4f(tanh_hi, 9.0f);
178 // The monomial coefficients of the numerator polynomial (odd).
179 static EIGEN_DECLARE_CONST_Packet4f(alpha_1, 4.89352455891786e-3f);
180 static EIGEN_DECLARE_CONST_Packet4f(alpha_3, 6.37261928875436e-4f);
181 static EIGEN_DECLARE_CONST_Packet4f(alpha_5, 1.48572235717979e-5f);
182 static EIGEN_DECLARE_CONST_Packet4f(alpha_7, 5.12229709037114e-8f);
183 static EIGEN_DECLARE_CONST_Packet4f(alpha_9, -8.60467152213735e-11f);
184 static EIGEN_DECLARE_CONST_Packet4f(alpha_11, 2.00018790482477e-13f);
185 static EIGEN_DECLARE_CONST_Packet4f(alpha_13, -2.76076847742355e-16f);
186 // The monomial coefficients of the denominator polynomial (even).
187 static EIGEN_DECLARE_CONST_Packet4f(beta_0, 4.89352518554385e-3f);
188 static EIGEN_DECLARE_CONST_Packet4f(beta_2, 2.26843463243900e-3f);
189 static EIGEN_DECLARE_CONST_Packet4f(beta_4, 1.18534705686654e-4f);
190 static EIGEN_DECLARE_CONST_Packet4f(beta_6, 1.19825839466702e-6f);
191
192 Packet4f x = pabs(_x);
193 Packet4i tiny_mask = __builtin_msa_fclt_w(x, p4f_tanh_tiny);
194
195 // Clamp the inputs to the range [-9, 9] since anything outside
196 // this range is -/+1.0f in single-precision.
197 x = (Packet4f)__builtin_msa_bsel_v((v16u8)__builtin_msa_fclt_w(p4f_tanh_hi, x), (v16u8)x, (v16u8)p4f_tanh_hi);
198
199 // Since the polynomials are odd/even, we need x**2.
200 Packet4f x2 = pmul(x, x);
201
202 // Evaluate the numerator polynomial p.
203 Packet4f p = pmadd(x2, p4f_alpha_13, p4f_alpha_11);
204 p = pmadd(x2, p, p4f_alpha_9);
205 p = pmadd(x2, p, p4f_alpha_7);
206 p = pmadd(x2, p, p4f_alpha_5);
207 p = pmadd(x2, p, p4f_alpha_3);
208 p = pmadd(x2, p, p4f_alpha_1);
209 p = pmul(x, p);
210
211 // Evaluate the denominator polynomial q.
212 Packet4f q = pmadd(x2, p4f_beta_6, p4f_beta_4);
213 q = pmadd(x2, q, p4f_beta_2);
214 q = pmadd(x2, q, p4f_beta_0);
215
216 // Divide the numerator by the denominator.
217 p = pdiv(p, q);
218
219 // Reinstate the sign.
220 p = (Packet4f)__builtin_msa_binsli_w((v4u32)p, (v4u32)_x, 0);
221
222 // When the argument is very small in magnitude it's more accurate to just return it.
223 p = (Packet4f)__builtin_msa_bsel_v((v16u8)tiny_mask, (v16u8)p, (v16u8)_x);
224
225 return p;
226}
227
228template <bool sine>
229Packet4f psincos_inner_msa_float(const Packet4f& _x) {
230 static EIGEN_DECLARE_CONST_Packet4f(sincos_max_arg, 13176795.0f); // Approx. (2**24) / (4/Pi).
231 static EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP1, -0.78515625f);
232 static EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP2, -2.4187564849853515625e-4f);
233 static EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP3, -3.77489497744594108e-8f);
234 static EIGEN_DECLARE_CONST_Packet4f(sincof_p0, -1.9515295891e-4f);
235 static EIGEN_DECLARE_CONST_Packet4f(sincof_p1, 8.3321608736e-3f);
236 static EIGEN_DECLARE_CONST_Packet4f(sincof_p2, -1.6666654611e-1f);
237 static EIGEN_DECLARE_CONST_Packet4f(coscof_p0, 2.443315711809948e-5f);
238 static EIGEN_DECLARE_CONST_Packet4f(coscof_p1, -1.388731625493765e-3f);
239 static EIGEN_DECLARE_CONST_Packet4f(coscof_p2, 4.166664568298827e-2f);
240 static EIGEN_DECLARE_CONST_Packet4f(cephes_FOPI, 1.27323954473516f); // 4/Pi.
241 static EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
242 static EIGEN_DECLARE_CONST_Packet4f(1, 1.0f);
243
244 Packet4f x = pabs(_x);
245
246 // Translate infinite arguments into NANs.
247 Packet4f zero_or_nan_if_inf = psub(_x, _x);
248 x = padd(x, zero_or_nan_if_inf);
249 // Prevent sin/cos from generating values larger than 1.0 in magnitude
250 // for very large arguments by setting x to 0.0.
251 Packet4i small_or_nan_mask = __builtin_msa_fcult_w(x, p4f_sincos_max_arg);
252 x = pand(x, (Packet4f)small_or_nan_mask);
253
254 // Scale x by 4/Pi to find x's octant.
255 Packet4f y = pmul(x, p4f_cephes_FOPI);
256 // Get the octant. We'll reduce x by this number of octants or by one more than it.
257 Packet4i y_int = __builtin_msa_ftrunc_s_w(y);
258 // x's from even-numbered octants will translate to octant 0: [0, +Pi/4].
259 // x's from odd-numbered octants will translate to octant -1: [-Pi/4, 0].
260 // Adjustment for odd-numbered octants: octant = (octant + 1) & (~1).
261 Packet4i y_int1 = __builtin_msa_addvi_w(y_int, 1);
262 Packet4i y_int2 = (Packet4i)__builtin_msa_bclri_w((Packet4ui)y_int1, 0); // bclri = bit-clear
263 y = __builtin_msa_ffint_s_w(y_int2);
264
265 // Compute the sign to apply to the polynomial.
266 Packet4i sign_mask = sine ? pxor(__builtin_msa_slli_w(y_int1, 29), (Packet4i)_x)
267 : __builtin_msa_slli_w(__builtin_msa_addvi_w(y_int, 3), 29);
268
269 // Get the polynomial selection mask.
270 // We'll calculate both (sin and cos) polynomials and then select from the two.
271 Packet4i poly_mask = __builtin_msa_ceqi_w(__builtin_msa_slli_w(y_int2, 30), 0);
272
273 // Reduce x by y octants to get: -Pi/4 <= x <= +Pi/4.
274 // The magic pass: "Extended precision modular arithmetic"
275 // x = ((x - y * DP1) - y * DP2) - y * DP3
276 Packet4f tmp1 = pmul(y, p4f_minus_cephes_DP1);
277 Packet4f tmp2 = pmul(y, p4f_minus_cephes_DP2);
278 Packet4f tmp3 = pmul(y, p4f_minus_cephes_DP3);
279 x = padd(x, tmp1);
280 x = padd(x, tmp2);
281 x = padd(x, tmp3);
282
283 // Evaluate the cos(x) polynomial.
284 y = p4f_coscof_p0;
285 Packet4f z = pmul(x, x);
286 y = pmadd(y, z, p4f_coscof_p1);
287 y = pmadd(y, z, p4f_coscof_p2);
288 y = pmul(y, z);
289 y = pmul(y, z);
290 y = __builtin_msa_fmsub_w(y, z, p4f_half);
291 y = padd(y, p4f_1);
292
293 // Evaluate the sin(x) polynomial.
294 Packet4f y2 = p4f_sincof_p0;
295 y2 = pmadd(y2, z, p4f_sincof_p1);
296 y2 = pmadd(y2, z, p4f_sincof_p2);
297 y2 = pmul(y2, z);
298 y2 = pmadd(y2, x, x);
299
300 // Select the correct result from the two polynomials.
301 y = sine ? (Packet4f)__builtin_msa_bsel_v((v16u8)poly_mask, (v16u8)y, (v16u8)y2)
302 : (Packet4f)__builtin_msa_bsel_v((v16u8)poly_mask, (v16u8)y2, (v16u8)y);
303
304 // Update the sign.
305 sign_mask = pxor(sign_mask, (Packet4i)y);
306 y = (Packet4f)__builtin_msa_binsli_w((v4u32)y, (v4u32)sign_mask, 0); // binsli = bit-insert-left
307 return y;
308}
309
310template <>
311EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet4f psin<Packet4f>(const Packet4f& x) {
312 return psincos_inner_msa_float</* sine */ true>(x);
313}
314
315template <>
316EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet4f pcos<Packet4f>(const Packet4f& x) {
317 return psincos_inner_msa_float</* sine */ false>(x);
318}
319
320template <>
321EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet2d pexp<Packet2d>(const Packet2d& _x) {
322 // Limiting double-precision pexp's argument to [-1024, +1024] lets pexp
323 // reach 0 and INFINITY naturally.
324 static EIGEN_DECLARE_CONST_Packet2d(exp_lo, -1024.0);
325 static EIGEN_DECLARE_CONST_Packet2d(exp_hi, +1024.0);
326 static EIGEN_DECLARE_CONST_Packet2d(cephes_LOG2EF, 1.4426950408889634073599);
327 static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125);
328 static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6);
329 static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p0, 1.26177193074810590878e-4);
330 static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p1, 3.02994407707441961300e-2);
331 static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p2, 9.99999999999999999910e-1);
332 static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q0, 3.00198505138664455042e-6);
333 static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q1, 2.52448340349684104192e-3);
334 static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q2, 2.27265548208155028766e-1);
335 static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0);
336 static EIGEN_DECLARE_CONST_Packet2d(half, 0.5);
337 static EIGEN_DECLARE_CONST_Packet2d(1, 1.0);
338 static EIGEN_DECLARE_CONST_Packet2d(2, 2.0);
339
340 Packet2d x = _x;
341
342 // Clamp x.
343 x = (Packet2d)__builtin_msa_bsel_v((v16u8)__builtin_msa_fclt_d(x, p2d_exp_lo), (v16u8)x, (v16u8)p2d_exp_lo);
344 x = (Packet2d)__builtin_msa_bsel_v((v16u8)__builtin_msa_fclt_d(p2d_exp_hi, x), (v16u8)x, (v16u8)p2d_exp_hi);
345
346 // Round to nearest integer by adding 0.5 (with x's sign) and truncating.
347 Packet2d x2_add = (Packet2d)__builtin_msa_binsli_d((v2u64)p2d_half, (v2u64)x, 0);
348 Packet2d x2 = pmadd(x, p2d_cephes_LOG2EF, x2_add);
349 Packet2l x2_long = __builtin_msa_ftrunc_s_d(x2);
350 Packet2d x2_long_d = __builtin_msa_ffint_s_d(x2_long);
351
352 x = __builtin_msa_fmsub_d(x, x2_long_d, p2d_cephes_exp_C1);
353 x = __builtin_msa_fmsub_d(x, x2_long_d, p2d_cephes_exp_C2);
354
355 x2 = pmul(x, x);
356
357 Packet2d px = p2d_cephes_exp_p0;
358 px = pmadd(px, x2, p2d_cephes_exp_p1);
359 px = pmadd(px, x2, p2d_cephes_exp_p2);
360 px = pmul(px, x);
361
362 Packet2d qx = p2d_cephes_exp_q0;
363 qx = pmadd(qx, x2, p2d_cephes_exp_q1);
364 qx = pmadd(qx, x2, p2d_cephes_exp_q2);
365 qx = pmadd(qx, x2, p2d_cephes_exp_q3);
366
367 x = pdiv(px, psub(qx, px));
368 x = pmadd(p2d_2, x, p2d_1);
369
370 // x *= 2**exponent.
371 x = __builtin_msa_fexp2_d(x, x2_long);
372
373 return x;
374}
375
376} // end namespace internal
377
378} // end namespace Eigen
379
380#endif // EIGEN_MATH_FUNCTIONS_MSA_H