xref: /aosp_15_r20/external/eigen/Eigen/src/Core/arch/MSA/MathFunctions.h (revision bf2c37156dfe67e5dfebd6d394bad8b2ab5804d4)
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 ([email protected])
6 // Copyright (C) 2016 Gael Guennebaud <[email protected]>
7 //
8 // Copyright (C) 2018 Wave Computing, Inc.
9 // Written by:
10 //   Chris Larsen
11 //   Alexey Frunze ([email protected])
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 
17 /* The sin, cos, exp, and log functions of this file come from
18  * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/
19  */
20 
21 /* The tanh function of this file is an adaptation of
22  * template<typename T> T generic_fast_tanh_float(const T&)
23  * from MathFunctionsImpl.h.
24  */
25 
26 #ifndef EIGEN_MATH_FUNCTIONS_MSA_H
27 #define EIGEN_MATH_FUNCTIONS_MSA_H
28 
29 namespace Eigen {
30 
31 namespace internal {
32 
33 template <>
34 EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f
35 plog<Packet4f>(const Packet4f& _x) {
36   static _EIGEN_DECLARE_CONST_Packet4f(cephes_SQRTHF, 0.707106781186547524f);
37   static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p0, 7.0376836292e-2f);
38   static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p1, -1.1514610310e-1f);
39   static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p2, 1.1676998740e-1f);
40   static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p3, -1.2420140846e-1f);
41   static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p4, +1.4249322787e-1f);
42   static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p5, -1.6668057665e-1f);
43   static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p6, +2.0000714765e-1f);
44   static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p7, -2.4999993993e-1f);
45   static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p8, +3.3333331174e-1f);
46   static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q1, -2.12194440e-4f);
47   static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q2, 0.693359375f);
48   static _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
49   static _EIGEN_DECLARE_CONST_Packet4f(1, 1.0f);
50 
51   // Convert negative argument into NAN (quiet negative, to be specific).
52   Packet4f zero = (Packet4f)__builtin_msa_ldi_w(0);
53   Packet4i neg_mask = __builtin_msa_fclt_w(_x, zero);
54   Packet4i zero_mask = __builtin_msa_fceq_w(_x, zero);
55   Packet4f non_neg_x_or_nan = padd(_x, (Packet4f)neg_mask);  // Add 0.0 or NAN.
56   Packet4f x = non_neg_x_or_nan;
57 
58   // Extract exponent from x = mantissa * 2**exponent, where 1.0 <= mantissa < 2.0.
59   // N.B. the exponent is one less of what frexpf() would return.
60   Packet4i e_int = __builtin_msa_ftint_s_w(__builtin_msa_flog2_w(x));
61   // Multiply x by 2**(-exponent-1) to get 0.5 <= x < 1.0 as from frexpf().
62   x = __builtin_msa_fexp2_w(x, (Packet4i)__builtin_msa_nori_b((v16u8)e_int, 0));
63 
64   /*
65      if (x < SQRTHF) {
66        x = x + x - 1.0;
67      } else {
68        e += 1;
69        x = x - 1.0;
70      }
71   */
72   Packet4f xx = padd(x, x);
73   Packet4i ge_mask = __builtin_msa_fcle_w(p4f_cephes_SQRTHF, x);
74   e_int = psub(e_int, ge_mask);
75   x = (Packet4f)__builtin_msa_bsel_v((v16u8)ge_mask, (v16u8)xx, (v16u8)x);
76   x = psub(x, p4f_1);
77   Packet4f e = __builtin_msa_ffint_s_w(e_int);
78 
79   Packet4f x2 = pmul(x, x);
80   Packet4f x3 = pmul(x2, x);
81 
82   Packet4f y, y1, y2;
83   y = pmadd(p4f_cephes_log_p0, x, p4f_cephes_log_p1);
84   y1 = pmadd(p4f_cephes_log_p3, x, p4f_cephes_log_p4);
85   y2 = pmadd(p4f_cephes_log_p6, x, p4f_cephes_log_p7);
86   y = pmadd(y, x, p4f_cephes_log_p2);
87   y1 = pmadd(y1, x, p4f_cephes_log_p5);
88   y2 = pmadd(y2, x, p4f_cephes_log_p8);
89   y = pmadd(y, x3, y1);
90   y = pmadd(y, x3, y2);
91   y = pmul(y, x3);
92 
93   y = pmadd(e, p4f_cephes_log_q1, y);
94   x = __builtin_msa_fmsub_w(x, x2, p4f_half);
95   x = padd(x, y);
96   x = pmadd(e, p4f_cephes_log_q2, x);
97 
98   // x is now the logarithm result candidate. We still need to handle the
99   // extreme arguments of zero and positive infinity, though.
100   // N.B. if the argument is +INFINITY, x is NAN because the polynomial terms
101   // contain infinities of both signs (see the coefficients and code above).
102   // INFINITY - INFINITY is NAN.
103 
104   // If the argument is +INFINITY, make it the new result candidate.
105   // To achieve that we choose the smaller of the result candidate and the
106   // argument.
107   // This is correct for all finite pairs of values (the logarithm is smaller
108   // than the argument).
109   // This is also correct in the special case when the argument is +INFINITY
110   // and the result candidate is NAN. This is because the fmin.df instruction
111   // prefers non-NANs to NANs.
112   x = __builtin_msa_fmin_w(x, non_neg_x_or_nan);
113 
114   // If the argument is zero (including -0.0), the result becomes -INFINITY.
115   Packet4i neg_infs = __builtin_msa_slli_w(zero_mask, 23);
116   x = (Packet4f)__builtin_msa_bsel_v((v16u8)zero_mask, (v16u8)x, (v16u8)neg_infs);
117 
118   return x;
119 }
120 
121 template <>
122 EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f
123 pexp<Packet4f>(const Packet4f& _x) {
124   // Limiting single-precision pexp's argument to [-128, +128] lets pexp
125   // reach 0 and INFINITY naturally.
126   static _EIGEN_DECLARE_CONST_Packet4f(exp_lo, -128.0f);
127   static _EIGEN_DECLARE_CONST_Packet4f(exp_hi, +128.0f);
128   static _EIGEN_DECLARE_CONST_Packet4f(cephes_LOG2EF, 1.44269504088896341f);
129   static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C1, 0.693359375f);
130   static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C2, -2.12194440e-4f);
131   static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p0, 1.9875691500e-4f);
132   static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p1, 1.3981999507e-3f);
133   static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p2, 8.3334519073e-3f);
134   static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p3, 4.1665795894e-2f);
135   static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459e-1f);
136   static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201e-1f);
137   static _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
138   static _EIGEN_DECLARE_CONST_Packet4f(1, 1.0f);
139 
140   Packet4f x = _x;
141 
142   // Clamp x.
143   x = (Packet4f)__builtin_msa_bsel_v((v16u8)__builtin_msa_fclt_w(x, p4f_exp_lo), (v16u8)x,
144                                      (v16u8)p4f_exp_lo);
145   x = (Packet4f)__builtin_msa_bsel_v((v16u8)__builtin_msa_fclt_w(p4f_exp_hi, x), (v16u8)x,
146                                      (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 
174 template <>
175 EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f
176 ptanh<Packet4f>(const Packet4f& _x) {
177   static _EIGEN_DECLARE_CONST_Packet4f(tanh_tiny, 1e-4f);
178   static _EIGEN_DECLARE_CONST_Packet4f(tanh_hi, 9.0f);
179   // The monomial coefficients of the numerator polynomial (odd).
180   static _EIGEN_DECLARE_CONST_Packet4f(alpha_1, 4.89352455891786e-3f);
181   static _EIGEN_DECLARE_CONST_Packet4f(alpha_3, 6.37261928875436e-4f);
182   static _EIGEN_DECLARE_CONST_Packet4f(alpha_5, 1.48572235717979e-5f);
183   static _EIGEN_DECLARE_CONST_Packet4f(alpha_7, 5.12229709037114e-8f);
184   static _EIGEN_DECLARE_CONST_Packet4f(alpha_9, -8.60467152213735e-11f);
185   static _EIGEN_DECLARE_CONST_Packet4f(alpha_11, 2.00018790482477e-13f);
186   static _EIGEN_DECLARE_CONST_Packet4f(alpha_13, -2.76076847742355e-16f);
187   // The monomial coefficients of the denominator polynomial (even).
188   static _EIGEN_DECLARE_CONST_Packet4f(beta_0, 4.89352518554385e-3f);
189   static _EIGEN_DECLARE_CONST_Packet4f(beta_2, 2.26843463243900e-3f);
190   static _EIGEN_DECLARE_CONST_Packet4f(beta_4, 1.18534705686654e-4f);
191   static _EIGEN_DECLARE_CONST_Packet4f(beta_6, 1.19825839466702e-6f);
192 
193   Packet4f x = pabs(_x);
194   Packet4i tiny_mask = __builtin_msa_fclt_w(x, p4f_tanh_tiny);
195 
196   // Clamp the inputs to the range [-9, 9] since anything outside
197   // this range is -/+1.0f in single-precision.
198   x = (Packet4f)__builtin_msa_bsel_v((v16u8)__builtin_msa_fclt_w(p4f_tanh_hi, x), (v16u8)x,
199                                      (v16u8)p4f_tanh_hi);
200 
201   // Since the polynomials are odd/even, we need x**2.
202   Packet4f x2 = pmul(x, x);
203 
204   // Evaluate the numerator polynomial p.
205   Packet4f p = pmadd(x2, p4f_alpha_13, p4f_alpha_11);
206   p = pmadd(x2, p, p4f_alpha_9);
207   p = pmadd(x2, p, p4f_alpha_7);
208   p = pmadd(x2, p, p4f_alpha_5);
209   p = pmadd(x2, p, p4f_alpha_3);
210   p = pmadd(x2, p, p4f_alpha_1);
211   p = pmul(x, p);
212 
213   // Evaluate the denominator polynomial q.
214   Packet4f q = pmadd(x2, p4f_beta_6, p4f_beta_4);
215   q = pmadd(x2, q, p4f_beta_2);
216   q = pmadd(x2, q, p4f_beta_0);
217 
218   // Divide the numerator by the denominator.
219   p = pdiv(p, q);
220 
221   // Reinstate the sign.
222   p = (Packet4f)__builtin_msa_binsli_w((v4u32)p, (v4u32)_x, 0);
223 
224   // When the argument is very small in magnitude it's more accurate to just return it.
225   p = (Packet4f)__builtin_msa_bsel_v((v16u8)tiny_mask, (v16u8)p, (v16u8)_x);
226 
227   return p;
228 }
229 
230 template <bool sine>
psincos_inner_msa_float(const Packet4f & _x)231 Packet4f psincos_inner_msa_float(const Packet4f& _x) {
232   static _EIGEN_DECLARE_CONST_Packet4f(sincos_max_arg, 13176795.0f);  // Approx. (2**24) / (4/Pi).
233   static _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP1, -0.78515625f);
234   static _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP2, -2.4187564849853515625e-4f);
235   static _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP3, -3.77489497744594108e-8f);
236   static _EIGEN_DECLARE_CONST_Packet4f(sincof_p0, -1.9515295891e-4f);
237   static _EIGEN_DECLARE_CONST_Packet4f(sincof_p1, 8.3321608736e-3f);
238   static _EIGEN_DECLARE_CONST_Packet4f(sincof_p2, -1.6666654611e-1f);
239   static _EIGEN_DECLARE_CONST_Packet4f(coscof_p0, 2.443315711809948e-5f);
240   static _EIGEN_DECLARE_CONST_Packet4f(coscof_p1, -1.388731625493765e-3f);
241   static _EIGEN_DECLARE_CONST_Packet4f(coscof_p2, 4.166664568298827e-2f);
242   static _EIGEN_DECLARE_CONST_Packet4f(cephes_FOPI, 1.27323954473516f);  // 4/Pi.
243   static _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
244   static _EIGEN_DECLARE_CONST_Packet4f(1, 1.0f);
245 
246   Packet4f x = pabs(_x);
247 
248   // Translate infinite arguments into NANs.
249   Packet4f zero_or_nan_if_inf = psub(_x, _x);
250   x = padd(x, zero_or_nan_if_inf);
251   // Prevent sin/cos from generating values larger than 1.0 in magnitude
252   // for very large arguments by setting x to 0.0.
253   Packet4i small_or_nan_mask = __builtin_msa_fcult_w(x, p4f_sincos_max_arg);
254   x = pand(x, (Packet4f)small_or_nan_mask);
255 
256   // Scale x by 4/Pi to find x's octant.
257   Packet4f y = pmul(x, p4f_cephes_FOPI);
258   // Get the octant. We'll reduce x by this number of octants or by one more than it.
259   Packet4i y_int = __builtin_msa_ftrunc_s_w(y);
260   // x's from even-numbered octants will translate to octant 0: [0, +Pi/4].
261   // x's from odd-numbered octants will translate to octant -1: [-Pi/4, 0].
262   // Adjustment for odd-numbered octants: octant = (octant + 1) & (~1).
263   Packet4i y_int1 = __builtin_msa_addvi_w(y_int, 1);
264   Packet4i y_int2 = (Packet4i)__builtin_msa_bclri_w((Packet4ui)y_int1, 0); // bclri = bit-clear
265   y = __builtin_msa_ffint_s_w(y_int2);
266 
267   // Compute the sign to apply to the polynomial.
268   Packet4i sign_mask = sine ? pxor(__builtin_msa_slli_w(y_int1, 29), (Packet4i)_x)
269                             : __builtin_msa_slli_w(__builtin_msa_addvi_w(y_int, 3), 29);
270 
271   // Get the polynomial selection mask.
272   // We'll calculate both (sin and cos) polynomials and then select from the two.
273   Packet4i poly_mask = __builtin_msa_ceqi_w(__builtin_msa_slli_w(y_int2, 30), 0);
274 
275   // Reduce x by y octants to get: -Pi/4 <= x <= +Pi/4.
276   // The magic pass: "Extended precision modular arithmetic"
277   // x = ((x - y * DP1) - y * DP2) - y * DP3
278   Packet4f tmp1 = pmul(y, p4f_minus_cephes_DP1);
279   Packet4f tmp2 = pmul(y, p4f_minus_cephes_DP2);
280   Packet4f tmp3 = pmul(y, p4f_minus_cephes_DP3);
281   x = padd(x, tmp1);
282   x = padd(x, tmp2);
283   x = padd(x, tmp3);
284 
285   // Evaluate the cos(x) polynomial.
286   y = p4f_coscof_p0;
287   Packet4f z = pmul(x, x);
288   y = pmadd(y, z, p4f_coscof_p1);
289   y = pmadd(y, z, p4f_coscof_p2);
290   y = pmul(y, z);
291   y = pmul(y, z);
292   y = __builtin_msa_fmsub_w(y, z, p4f_half);
293   y = padd(y, p4f_1);
294 
295   // Evaluate the sin(x) polynomial.
296   Packet4f y2 = p4f_sincof_p0;
297   y2 = pmadd(y2, z, p4f_sincof_p1);
298   y2 = pmadd(y2, z, p4f_sincof_p2);
299   y2 = pmul(y2, z);
300   y2 = pmadd(y2, x, x);
301 
302   // Select the correct result from the two polynomials.
303   y = sine ? (Packet4f)__builtin_msa_bsel_v((v16u8)poly_mask, (v16u8)y, (v16u8)y2)
304            : (Packet4f)__builtin_msa_bsel_v((v16u8)poly_mask, (v16u8)y2, (v16u8)y);
305 
306   // Update the sign.
307   sign_mask = pxor(sign_mask, (Packet4i)y);
308   y = (Packet4f)__builtin_msa_binsli_w((v4u32)y, (v4u32)sign_mask, 0); // binsli = bit-insert-left
309   return y;
310 }
311 
312 template <>
313 EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f
314 psin<Packet4f>(const Packet4f& x) {
315   return psincos_inner_msa_float</* sine */ true>(x);
316 }
317 
318 template <>
319 EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f
320 pcos<Packet4f>(const Packet4f& x) {
321   return psincos_inner_msa_float</* sine */ false>(x);
322 }
323 
324 template <>
325 EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet2d
326 pexp<Packet2d>(const Packet2d& _x) {
327   // Limiting double-precision pexp's argument to [-1024, +1024] lets pexp
328   // reach 0 and INFINITY naturally.
329   static _EIGEN_DECLARE_CONST_Packet2d(exp_lo, -1024.0);
330   static _EIGEN_DECLARE_CONST_Packet2d(exp_hi, +1024.0);
331   static _EIGEN_DECLARE_CONST_Packet2d(cephes_LOG2EF, 1.4426950408889634073599);
332   static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125);
333   static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6);
334   static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p0, 1.26177193074810590878e-4);
335   static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p1, 3.02994407707441961300e-2);
336   static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p2, 9.99999999999999999910e-1);
337   static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q0, 3.00198505138664455042e-6);
338   static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q1, 2.52448340349684104192e-3);
339   static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q2, 2.27265548208155028766e-1);
340   static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0);
341   static _EIGEN_DECLARE_CONST_Packet2d(half, 0.5);
342   static _EIGEN_DECLARE_CONST_Packet2d(1, 1.0);
343   static _EIGEN_DECLARE_CONST_Packet2d(2, 2.0);
344 
345   Packet2d x = _x;
346 
347   // Clamp x.
348   x = (Packet2d)__builtin_msa_bsel_v((v16u8)__builtin_msa_fclt_d(x, p2d_exp_lo), (v16u8)x,
349                                      (v16u8)p2d_exp_lo);
350   x = (Packet2d)__builtin_msa_bsel_v((v16u8)__builtin_msa_fclt_d(p2d_exp_hi, x), (v16u8)x,
351                                      (v16u8)p2d_exp_hi);
352 
353   // Round to nearest integer by adding 0.5 (with x's sign) and truncating.
354   Packet2d x2_add = (Packet2d)__builtin_msa_binsli_d((v2u64)p2d_half, (v2u64)x, 0);
355   Packet2d x2 = pmadd(x, p2d_cephes_LOG2EF, x2_add);
356   Packet2l x2_long = __builtin_msa_ftrunc_s_d(x2);
357   Packet2d x2_long_d = __builtin_msa_ffint_s_d(x2_long);
358 
359   x = __builtin_msa_fmsub_d(x, x2_long_d, p2d_cephes_exp_C1);
360   x = __builtin_msa_fmsub_d(x, x2_long_d, p2d_cephes_exp_C2);
361 
362   x2 = pmul(x, x);
363 
364   Packet2d px = p2d_cephes_exp_p0;
365   px = pmadd(px, x2, p2d_cephes_exp_p1);
366   px = pmadd(px, x2, p2d_cephes_exp_p2);
367   px = pmul(px, x);
368 
369   Packet2d qx = p2d_cephes_exp_q0;
370   qx = pmadd(qx, x2, p2d_cephes_exp_q1);
371   qx = pmadd(qx, x2, p2d_cephes_exp_q2);
372   qx = pmadd(qx, x2, p2d_cephes_exp_q3);
373 
374   x = pdiv(px, psub(qx, px));
375   x = pmadd(p2d_2, x, p2d_1);
376 
377   // x *= 2**exponent.
378   x = __builtin_msa_fexp2_d(x, x2_long);
379 
380   return x;
381 }
382 
383 }  // end namespace internal
384 
385 }  // end namespace Eigen
386 
387 #endif  // EIGEN_MATH_FUNCTIONS_MSA_H
388