1 /* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
4 * This file is part of GROMACS.
7 * Written by the Gromacs development team under coordination of
8 * David van der Spoel, Berk Hess, and Erik Lindahl.
10 * This library is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public License
12 * as published by the Free Software Foundation; either version 2
13 * of the License, or (at your option) any later version.
15 * To help us fund GROMACS development, we humbly ask that you cite
16 * the research papers on the package. Check out http://www.gromacs.org
19 * Gnomes, ROck Monsters And Chili Sauce
21 #ifndef _gmx_math_x86_avx_128_fma_single_h_
22 #define _gmx_math_x86_avx_128_fma_single_h_
24 #include <immintrin.h> /* AVX */
25 #ifdef HAVE_X86INTRIN_H
26 #include <x86intrin.h> /* FMA */
29 #include <intrin.h> /* FMA MSVC */
34 #include "gmx_x86_avx_128_fma.h"
38 # define M_PI 3.14159265358979323846264338327950288
44 /************************
46 * Simple math routines *
48 ************************/
51 static gmx_inline __m128
52 gmx_mm_invsqrt_ps(__m128 x)
54 const __m128 half = _mm_set1_ps(0.5);
55 const __m128 one = _mm_set1_ps(1.0);
57 __m128 lu = _mm_rsqrt_ps(x);
59 return _mm_macc_ps(_mm_nmacc_ps(x, _mm_mul_ps(lu, lu), one), _mm_mul_ps(lu, half), lu);
62 /* sqrt(x) - Do NOT use this (but rather invsqrt) if you actually need 1.0/sqrt(x) */
63 static gmx_inline __m128
64 gmx_mm_sqrt_ps(__m128 x)
69 mask = _mm_cmp_ps(x, _mm_setzero_ps(), _CMP_EQ_OQ);
70 res = _mm_andnot_ps(mask, gmx_mm_invsqrt_ps(x));
72 res = _mm_mul_ps(x, res);
78 static gmx_inline __m128
79 gmx_mm_inv_ps(__m128 x)
81 const __m128 two = _mm_set1_ps(2.0);
83 __m128 lu = _mm_rcp_ps(x);
85 return _mm_mul_ps(lu, _mm_nmacc_ps(lu, x, two));
88 static gmx_inline __m128
89 gmx_mm_abs_ps(__m128 x)
91 const __m128 signmask = gmx_mm_castsi128_ps( _mm_set1_epi32(0x7FFFFFFF) );
93 return _mm_and_ps(x, signmask);
97 gmx_mm_log_ps(__m128 x)
99 /* Same algorithm as cephes library */
100 const __m128 expmask = gmx_mm_castsi128_ps( _mm_set_epi32(0x7F800000, 0x7F800000, 0x7F800000, 0x7F800000) );
101 const __m128i expbase_m1 = _mm_set1_epi32(127-1); /* We want non-IEEE format */
102 const __m128 half = _mm_set1_ps(0.5f);
103 const __m128 one = _mm_set1_ps(1.0f);
104 const __m128 invsq2 = _mm_set1_ps(1.0f/sqrt(2.0f));
105 const __m128 corr1 = _mm_set1_ps(-2.12194440e-4f);
106 const __m128 corr2 = _mm_set1_ps(0.693359375f);
108 const __m128 CA_1 = _mm_set1_ps(0.070376836292f);
109 const __m128 CB_0 = _mm_set1_ps(1.6714950086782716f);
110 const __m128 CB_1 = _mm_set1_ps(-2.452088066061482f);
111 const __m128 CC_0 = _mm_set1_ps(1.5220770854701728f);
112 const __m128 CC_1 = _mm_set1_ps(-1.3422238433233642f);
113 const __m128 CD_0 = _mm_set1_ps(1.386218787509749f);
114 const __m128 CD_1 = _mm_set1_ps(0.35075468953796346f);
115 const __m128 CE_0 = _mm_set1_ps(1.3429983063133937f);
116 const __m128 CE_1 = _mm_set1_ps(1.807420826584643f);
123 __m128 pA, pB, pC, pD, pE, tB, tC, tD, tE;
125 /* Separate x into exponent and mantissa, with a mantissa in the range [0.5..1[ (not IEEE754 standard!) */
126 fexp = _mm_and_ps(x, expmask);
127 iexp = gmx_mm_castps_si128(fexp);
128 iexp = _mm_srli_epi32(iexp, 23);
129 iexp = _mm_sub_epi32(iexp, expbase_m1);
131 x = _mm_andnot_ps(expmask, x);
132 x = _mm_or_ps(x, one);
133 x = _mm_mul_ps(x, half);
135 mask = _mm_cmp_ps(x, invsq2, _CMP_LT_OQ);
137 x = _mm_add_ps(x, _mm_and_ps(mask, x));
138 x = _mm_sub_ps(x, one);
139 iexp = _mm_add_epi32(iexp, gmx_mm_castps_si128(mask)); /* 0xFFFFFFFF = -1 as int */
141 x2 = _mm_mul_ps(x, x);
143 pA = _mm_mul_ps(CA_1, x);
145 pB = _mm_add_ps(x, CB_1);
146 pC = _mm_add_ps(x, CC_1);
147 pD = _mm_add_ps(x, CD_1);
148 pE = _mm_add_ps(x, CE_1);
150 pB = _mm_macc_ps(x, pB, CB_0);
151 pC = _mm_macc_ps(x, pC, CC_0);
152 pD = _mm_macc_ps(x, pD, CD_0);
153 pE = _mm_macc_ps(x, pE, CE_0);
155 pA = _mm_mul_ps(pA, pB);
156 pC = _mm_mul_ps(pC, pD);
157 pE = _mm_mul_ps(pE, x2);
158 pA = _mm_mul_ps(pA, pC);
159 y = _mm_mul_ps(pA, pE);
161 fexp = _mm_cvtepi32_ps(iexp);
162 y = _mm_macc_ps(fexp, corr1, y);
163 y = _mm_nmacc_ps(half, x2, y);
165 x2 = _mm_add_ps(x, y);
166 x2 = _mm_macc_ps(fexp, corr2, x2);
175 * The 2^w term is calculated from a (6,0)-th order (no denominator) Minimax polynomia on the interval
176 * [-0.5,0.5]. The coefficiencts of this was derived in Mathematica using the command:
178 * MiniMaxApproximation[(2^x), {x, {-0.5, 0.5}, 6, 0}, WorkingPrecision -> 15]
180 * The largest-magnitude exponent we can represent in IEEE single-precision binary format
181 * is 2^-126 for small numbers and 2^127 for large ones. To avoid wrap-around problems, we set the
182 * result to zero if the argument falls outside this range. For small numbers this is just fine, but
183 * for large numbers you could be fancy and return the smallest/largest IEEE single-precision
184 * number instead. That would take a few extra cycles and not really help, since something is
185 * wrong if you are using single precision to work with numbers that cannot really be represented
186 * in single precision.
188 * The accuracy is at least 23 bits.
191 gmx_mm_exp2_ps(__m128 x)
193 /* Lower bound: We do not allow numbers that would lead to an IEEE fp representation exponent smaller than -126. */
194 const __m128 arglimit = _mm_set1_ps(126.0f);
196 const __m128i expbase = _mm_set1_epi32(127);
197 const __m128 CA6 = _mm_set1_ps(1.535336188319500E-004);
198 const __m128 CA5 = _mm_set1_ps(1.339887440266574E-003);
199 const __m128 CA4 = _mm_set1_ps(9.618437357674640E-003);
200 const __m128 CA3 = _mm_set1_ps(5.550332471162809E-002);
201 const __m128 CA2 = _mm_set1_ps(2.402264791363012E-001);
202 const __m128 CA1 = _mm_set1_ps(6.931472028550421E-001);
203 const __m128 CA0 = _mm_set1_ps(1.0f);
212 iexppart = _mm_cvtps_epi32(x);
213 intpart = _mm_round_ps(x, _MM_FROUND_TO_NEAREST_INT);
214 iexppart = _mm_slli_epi32(_mm_add_epi32(iexppart, expbase), 23);
215 valuemask = _mm_cmp_ps(arglimit, gmx_mm_abs_ps(x), _CMP_GE_OQ);
216 fexppart = _mm_and_ps(valuemask, gmx_mm_castsi128_ps(iexppart));
218 x = _mm_sub_ps(x, intpart);
219 x2 = _mm_mul_ps(x, x);
221 p0 = _mm_macc_ps(CA6, x2, CA4);
222 p1 = _mm_macc_ps(CA5, x2, CA3);
223 p0 = _mm_macc_ps(p0, x2, CA2);
224 p1 = _mm_macc_ps(p1, x2, CA1);
225 p0 = _mm_macc_ps(p0, x2, CA0);
226 p0 = _mm_macc_ps(p1, x, p0);
227 x = _mm_mul_ps(p0, fexppart);
233 /* Exponential function. This could be calculated from 2^x as Exp(x)=2^(y), where y=log2(e)*x,
234 * but there will then be a small rounding error since we lose some precision due to the
235 * multiplication. This will then be magnified a lot by the exponential.
237 * Instead, we calculate the fractional part directly as a minimax approximation of
238 * Exp(z) on [-0.5,0.5]. We use extended precision arithmetics to calculate the fraction
239 * remaining after 2^y, which avoids the precision-loss.
240 * The final result is correct to within 1 LSB over the entire argument range.
243 gmx_mm_exp_ps(__m128 x)
245 const __m128 argscale = _mm_set1_ps(1.44269504088896341f);
246 /* Lower bound: Disallow numbers that would lead to an IEEE fp exponent reaching +-127. */
247 const __m128 arglimit = _mm_set1_ps(126.0f);
248 const __m128i expbase = _mm_set1_epi32(127);
250 const __m128 invargscale0 = _mm_set1_ps(0.693359375f);
251 const __m128 invargscale1 = _mm_set1_ps(-2.12194440e-4f);
253 const __m128 CC5 = _mm_set1_ps(1.9875691500e-4f);
254 const __m128 CC4 = _mm_set1_ps(1.3981999507e-3f);
255 const __m128 CC3 = _mm_set1_ps(8.3334519073e-3f);
256 const __m128 CC2 = _mm_set1_ps(4.1665795894e-2f);
257 const __m128 CC1 = _mm_set1_ps(1.6666665459e-1f);
258 const __m128 CC0 = _mm_set1_ps(5.0000001201e-1f);
259 const __m128 one = _mm_set1_ps(1.0f);
268 y = _mm_mul_ps(x, argscale);
270 iexppart = _mm_cvtps_epi32(y);
271 intpart = _mm_round_ps(y, _MM_FROUND_TO_NEAREST_INT);
273 iexppart = _mm_slli_epi32(_mm_add_epi32(iexppart, expbase), 23);
274 valuemask = _mm_cmp_ps(arglimit, gmx_mm_abs_ps(y), _CMP_GE_OQ);
275 fexppart = _mm_and_ps(valuemask, gmx_mm_castsi128_ps(iexppart));
277 /* Extended precision arithmetics */
278 x = _mm_nmacc_ps(invargscale0, intpart, x);
279 x = _mm_nmacc_ps(invargscale1, intpart, x);
281 x2 = _mm_mul_ps(x, x);
283 p1 = _mm_macc_ps(CC5, x2, CC3);
284 p0 = _mm_macc_ps(CC4, x2, CC2);
285 p1 = _mm_macc_ps(p1, x2, CC1);
286 p0 = _mm_macc_ps(p0, x2, CC0);
287 p0 = _mm_macc_ps(p1, x, p0);
288 p0 = _mm_macc_ps(p0, x2, one);
290 x = _mm_add_ps(x, p0);
292 x = _mm_mul_ps(x, fexppart);
297 /* FULL precision. Only errors in LSB */
299 gmx_mm_erf_ps(__m128 x)
301 /* Coefficients for minimax approximation of erf(x)=x*P(x^2) in range [-1,1] */
302 const __m128 CA6 = _mm_set1_ps(7.853861353153693e-5f);
303 const __m128 CA5 = _mm_set1_ps(-8.010193625184903e-4f);
304 const __m128 CA4 = _mm_set1_ps(5.188327685732524e-3f);
305 const __m128 CA3 = _mm_set1_ps(-2.685381193529856e-2f);
306 const __m128 CA2 = _mm_set1_ps(1.128358514861418e-1f);
307 const __m128 CA1 = _mm_set1_ps(-3.761262582423300e-1f);
308 const __m128 CA0 = _mm_set1_ps(1.128379165726710f);
309 /* Coefficients for minimax approximation of erfc(x)=Exp(-x^2)*P((1/(x-1))^2) in range [0.67,2] */
310 const __m128 CB9 = _mm_set1_ps(-0.0018629930017603923f);
311 const __m128 CB8 = _mm_set1_ps(0.003909821287598495f);
312 const __m128 CB7 = _mm_set1_ps(-0.0052094582210355615f);
313 const __m128 CB6 = _mm_set1_ps(0.005685614362160572f);
314 const __m128 CB5 = _mm_set1_ps(-0.0025367682853477272f);
315 const __m128 CB4 = _mm_set1_ps(-0.010199799682318782f);
316 const __m128 CB3 = _mm_set1_ps(0.04369575504816542f);
317 const __m128 CB2 = _mm_set1_ps(-0.11884063474674492f);
318 const __m128 CB1 = _mm_set1_ps(0.2732120154030589f);
319 const __m128 CB0 = _mm_set1_ps(0.42758357702025784f);
320 /* Coefficients for minimax approximation of erfc(x)=Exp(-x^2)*(1/x)*P((1/x)^2) in range [2,9.19] */
321 const __m128 CC10 = _mm_set1_ps(-0.0445555913112064f);
322 const __m128 CC9 = _mm_set1_ps(0.21376355144663348f);
323 const __m128 CC8 = _mm_set1_ps(-0.3473187200259257f);
324 const __m128 CC7 = _mm_set1_ps(0.016690861551248114f);
325 const __m128 CC6 = _mm_set1_ps(0.7560973182491192f);
326 const __m128 CC5 = _mm_set1_ps(-1.2137903600145787f);
327 const __m128 CC4 = _mm_set1_ps(0.8411872321232948f);
328 const __m128 CC3 = _mm_set1_ps(-0.08670413896296343f);
329 const __m128 CC2 = _mm_set1_ps(-0.27124782687240334f);
330 const __m128 CC1 = _mm_set1_ps(-0.0007502488047806069f);
331 const __m128 CC0 = _mm_set1_ps(0.5642114853803148f);
333 /* Coefficients for expansion of exp(x) in [0,0.1] */
334 /* CD0 and CD1 are both 1.0, so no need to declare them separately */
335 const __m128 CD2 = _mm_set1_ps(0.5000066608081202f);
336 const __m128 CD3 = _mm_set1_ps(0.1664795422874624f);
337 const __m128 CD4 = _mm_set1_ps(0.04379839977652482f);
339 const __m128 sieve = gmx_mm_castsi128_ps( _mm_set1_epi32(0xfffff000) );
340 const __m128 signbit = gmx_mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
341 const __m128 one = _mm_set1_ps(1.0f);
342 const __m128 two = _mm_set1_ps(2.0f);
345 __m128 z, q, t, t2, w, w2;
346 __m128 pA0, pA1, pB0, pB1, pC0, pC1;
348 __m128 res_erf, res_erfc, res;
351 /* Calculate erf() */
352 x2 = _mm_mul_ps(x, x);
353 x4 = _mm_mul_ps(x2, x2);
355 pA0 = _mm_macc_ps(CA6, x4, CA4);
356 pA1 = _mm_macc_ps(CA5, x4, CA3);
357 pA0 = _mm_macc_ps(pA0, x4, CA2);
358 pA1 = _mm_macc_ps(pA1, x4, CA1);
359 pA0 = _mm_mul_ps(pA0, x4);
360 pA0 = _mm_macc_ps(pA1, x2, pA0);
361 /* Constant term must come last for precision reasons */
362 pA0 = _mm_add_ps(pA0, CA0);
364 res_erf = _mm_mul_ps(x, pA0);
368 y = gmx_mm_abs_ps(x);
369 t = gmx_mm_inv_ps(y);
370 w = _mm_sub_ps(t, one);
371 t2 = _mm_mul_ps(t, t);
372 w2 = _mm_mul_ps(w, w);
374 * We cannot simply calculate exp(-x2) directly in single precision, since
375 * that will lose a couple of bits of precision due to the multiplication.
376 * Instead, we introduce x=z+w, where the last 12 bits of precision are in w.
377 * Then we get exp(-x2) = exp(-z2)*exp((z-x)*(z+x)).
379 * The only drawback with this is that it requires TWO separate exponential
380 * evaluations, which would be horrible performance-wise. However, the argument
381 * for the second exp() call is always small, so there we simply use a
382 * low-order minimax expansion on [0,0.1].
385 z = _mm_and_ps(y, sieve);
386 q = _mm_mul_ps( _mm_sub_ps(z, y), _mm_add_ps(z, y) );
388 corr = _mm_macc_ps(CD4, q, CD3);
389 corr = _mm_macc_ps(corr, q, CD2);
390 corr = _mm_macc_ps(corr, q, one);
391 corr = _mm_macc_ps(corr, q, one);
393 expmx2 = gmx_mm_exp_ps( _mm_or_ps( signbit, _mm_mul_ps(z, z) ) );
394 expmx2 = _mm_mul_ps(expmx2, corr);
396 pB1 = _mm_macc_ps(CB9, w2, CB7);
397 pB0 = _mm_macc_ps(CB8, w2, CB6);
398 pB1 = _mm_macc_ps(pB1, w2, CB5);
399 pB0 = _mm_macc_ps(pB0, w2, CB4);
400 pB1 = _mm_macc_ps(pB1, w2, CB3);
401 pB0 = _mm_macc_ps(pB0, w2, CB2);
402 pB1 = _mm_macc_ps(pB1, w2, CB1);
403 pB0 = _mm_macc_ps(pB0, w2, CB0);
404 pB0 = _mm_macc_ps(pB1, w, pB0);
406 pC0 = _mm_macc_ps(CC10, t2, CC8);
407 pC1 = _mm_macc_ps(CC9, t2, CC7);
408 pC0 = _mm_macc_ps(pC0, t2, CC6);
409 pC1 = _mm_macc_ps(pC1, t2, CC5);
410 pC0 = _mm_macc_ps(pC0, t2, CC4);
411 pC1 = _mm_macc_ps(pC1, t2, CC3);
412 pC0 = _mm_macc_ps(pC0, t2, CC2);
413 pC1 = _mm_macc_ps(pC1, t2, CC1);
415 pC0 = _mm_macc_ps(pC0, t2, CC0);
416 pC0 = _mm_macc_ps(pC1, t, pC0);
417 pC0 = _mm_mul_ps(pC0, t);
419 /* SELECT pB0 or pC0 for erfc() */
420 mask = _mm_cmp_ps(two, y, _CMP_LT_OQ);
421 res_erfc = _mm_blendv_ps(pB0, pC0, mask);
422 res_erfc = _mm_mul_ps(res_erfc, expmx2);
424 /* erfc(x<0) = 2-erfc(|x|) */
425 mask = _mm_cmp_ps(x, _mm_setzero_ps(), _CMP_LT_OQ);
426 res_erfc = _mm_blendv_ps(res_erfc, _mm_sub_ps(two, res_erfc), mask);
428 /* Select erf() or erfc() */
429 mask = _mm_cmp_ps(y, _mm_set1_ps(0.75f), _CMP_LT_OQ);
430 res = _mm_blendv_ps(_mm_sub_ps(one, res_erfc), res_erf, mask);
436 /* FULL precision. Only errors in LSB */
438 gmx_mm_erfc_ps(__m128 x)
440 /* Coefficients for minimax approximation of erf(x)=x*P(x^2) in range [-1,1] */
441 const __m128 CA6 = _mm_set1_ps(7.853861353153693e-5f);
442 const __m128 CA5 = _mm_set1_ps(-8.010193625184903e-4f);
443 const __m128 CA4 = _mm_set1_ps(5.188327685732524e-3f);
444 const __m128 CA3 = _mm_set1_ps(-2.685381193529856e-2f);
445 const __m128 CA2 = _mm_set1_ps(1.128358514861418e-1f);
446 const __m128 CA1 = _mm_set1_ps(-3.761262582423300e-1f);
447 const __m128 CA0 = _mm_set1_ps(1.128379165726710f);
448 /* Coefficients for minimax approximation of erfc(x)=Exp(-x^2)*P((1/(x-1))^2) in range [0.67,2] */
449 const __m128 CB9 = _mm_set1_ps(-0.0018629930017603923f);
450 const __m128 CB8 = _mm_set1_ps(0.003909821287598495f);
451 const __m128 CB7 = _mm_set1_ps(-0.0052094582210355615f);
452 const __m128 CB6 = _mm_set1_ps(0.005685614362160572f);
453 const __m128 CB5 = _mm_set1_ps(-0.0025367682853477272f);
454 const __m128 CB4 = _mm_set1_ps(-0.010199799682318782f);
455 const __m128 CB3 = _mm_set1_ps(0.04369575504816542f);
456 const __m128 CB2 = _mm_set1_ps(-0.11884063474674492f);
457 const __m128 CB1 = _mm_set1_ps(0.2732120154030589f);
458 const __m128 CB0 = _mm_set1_ps(0.42758357702025784f);
459 /* Coefficients for minimax approximation of erfc(x)=Exp(-x^2)*(1/x)*P((1/x)^2) in range [2,9.19] */
460 const __m128 CC10 = _mm_set1_ps(-0.0445555913112064f);
461 const __m128 CC9 = _mm_set1_ps(0.21376355144663348f);
462 const __m128 CC8 = _mm_set1_ps(-0.3473187200259257f);
463 const __m128 CC7 = _mm_set1_ps(0.016690861551248114f);
464 const __m128 CC6 = _mm_set1_ps(0.7560973182491192f);
465 const __m128 CC5 = _mm_set1_ps(-1.2137903600145787f);
466 const __m128 CC4 = _mm_set1_ps(0.8411872321232948f);
467 const __m128 CC3 = _mm_set1_ps(-0.08670413896296343f);
468 const __m128 CC2 = _mm_set1_ps(-0.27124782687240334f);
469 const __m128 CC1 = _mm_set1_ps(-0.0007502488047806069f);
470 const __m128 CC0 = _mm_set1_ps(0.5642114853803148f);
472 /* Coefficients for expansion of exp(x) in [0,0.1] */
473 /* CD0 and CD1 are both 1.0, so no need to declare them separately */
474 const __m128 CD2 = _mm_set1_ps(0.5000066608081202f);
475 const __m128 CD3 = _mm_set1_ps(0.1664795422874624f);
476 const __m128 CD4 = _mm_set1_ps(0.04379839977652482f);
478 const __m128 sieve = gmx_mm_castsi128_ps( _mm_set1_epi32(0xfffff000) );
479 const __m128 signbit = gmx_mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
480 const __m128 one = _mm_set1_ps(1.0f);
481 const __m128 two = _mm_set1_ps(2.0f);
484 __m128 z, q, t, t2, w, w2;
485 __m128 pA0, pA1, pB0, pB1, pC0, pC1;
487 __m128 res_erf, res_erfc, res;
490 /* Calculate erf() */
491 x2 = _mm_mul_ps(x, x);
492 x4 = _mm_mul_ps(x2, x2);
494 pA0 = _mm_macc_ps(CA6, x4, CA4);
495 pA1 = _mm_macc_ps(CA5, x4, CA3);
496 pA0 = _mm_macc_ps(pA0, x4, CA2);
497 pA1 = _mm_macc_ps(pA1, x4, CA1);
498 pA1 = _mm_mul_ps(pA1, x2);
499 pA0 = _mm_macc_ps(pA0, x4, pA1);
500 /* Constant term must come last for precision reasons */
501 pA0 = _mm_add_ps(pA0, CA0);
503 res_erf = _mm_mul_ps(x, pA0);
506 y = gmx_mm_abs_ps(x);
507 t = gmx_mm_inv_ps(y);
508 w = _mm_sub_ps(t, one);
509 t2 = _mm_mul_ps(t, t);
510 w2 = _mm_mul_ps(w, w);
512 * We cannot simply calculate exp(-x2) directly in single precision, since
513 * that will lose a couple of bits of precision due to the multiplication.
514 * Instead, we introduce x=z+w, where the last 12 bits of precision are in w.
515 * Then we get exp(-x2) = exp(-z2)*exp((z-x)*(z+x)).
517 * The only drawback with this is that it requires TWO separate exponential
518 * evaluations, which would be horrible performance-wise. However, the argument
519 * for the second exp() call is always small, so there we simply use a
520 * low-order minimax expansion on [0,0.1].
523 z = _mm_and_ps(y, sieve);
524 q = _mm_mul_ps( _mm_sub_ps(z, y), _mm_add_ps(z, y) );
526 corr = _mm_macc_ps(CD4, q, CD3);
527 corr = _mm_macc_ps(corr, q, CD2);
528 corr = _mm_macc_ps(corr, q, one);
529 corr = _mm_macc_ps(corr, q, one);
531 expmx2 = gmx_mm_exp_ps( _mm_or_ps( signbit, _mm_mul_ps(z, z) ) );
532 expmx2 = _mm_mul_ps(expmx2, corr);
534 pB1 = _mm_macc_ps(CB9, w2, CB7);
535 pB0 = _mm_macc_ps(CB8, w2, CB6);
536 pB1 = _mm_macc_ps(pB1, w2, CB5);
537 pB0 = _mm_macc_ps(pB0, w2, CB4);
538 pB1 = _mm_macc_ps(pB1, w2, CB3);
539 pB0 = _mm_macc_ps(pB0, w2, CB2);
540 pB1 = _mm_macc_ps(pB1, w2, CB1);
541 pB0 = _mm_macc_ps(pB0, w2, CB0);
542 pB0 = _mm_macc_ps(pB1, w, pB0);
544 pC0 = _mm_macc_ps(CC10, t2, CC8);
545 pC1 = _mm_macc_ps(CC9, t2, CC7);
546 pC0 = _mm_macc_ps(pC0, t2, CC6);
547 pC1 = _mm_macc_ps(pC1, t2, CC5);
548 pC0 = _mm_macc_ps(pC0, t2, CC4);
549 pC1 = _mm_macc_ps(pC1, t2, CC3);
550 pC0 = _mm_macc_ps(pC0, t2, CC2);
551 pC1 = _mm_macc_ps(pC1, t2, CC1);
553 pC0 = _mm_macc_ps(pC0, t2, CC0);
554 pC0 = _mm_macc_ps(pC1, t, pC0);
555 pC0 = _mm_mul_ps(pC0, t);
557 /* SELECT pB0 or pC0 for erfc() */
558 mask = _mm_cmp_ps(two, y, _CMP_LT_OQ);
559 res_erfc = _mm_blendv_ps(pB0, pC0, mask);
560 res_erfc = _mm_mul_ps(res_erfc, expmx2);
562 /* erfc(x<0) = 2-erfc(|x|) */
563 mask = _mm_cmp_ps(x, _mm_setzero_ps(), _CMP_LT_OQ);
564 res_erfc = _mm_blendv_ps(res_erfc, _mm_sub_ps(two, res_erfc), mask);
566 /* Select erf() or erfc() */
567 mask = _mm_cmp_ps(y, _mm_set1_ps(0.75f), _CMP_LT_OQ);
568 res = _mm_blendv_ps(res_erfc, _mm_sub_ps(one, res_erf), mask);
574 /* Calculate the force correction due to PME analytically.
576 * This routine is meant to enable analytical evaluation of the
577 * direct-space PME electrostatic force to avoid tables.
579 * The direct-space potential should be Erfc(beta*r)/r, but there
580 * are some problems evaluating that:
582 * First, the error function is difficult (read: expensive) to
583 * approxmiate accurately for intermediate to large arguments, and
584 * this happens already in ranges of beta*r that occur in simulations.
585 * Second, we now try to avoid calculating potentials in Gromacs but
586 * use forces directly.
588 * We can simply things slight by noting that the PME part is really
589 * a correction to the normal Coulomb force since Erfc(z)=1-Erf(z), i.e.
591 * V= 1/r - Erf(beta*r)/r
593 * The first term we already have from the inverse square root, so
594 * that we can leave out of this routine.
596 * For pme tolerances of 1e-3 to 1e-8 and cutoffs of 0.5nm to 1.8nm,
597 * the argument beta*r will be in the range 0.15 to ~4. Use your
598 * favorite plotting program to realize how well-behaved Erf(z)/z is
601 * We approximate f(z)=erf(z)/z with a rational minimax polynomial.
602 * However, it turns out it is more efficient to approximate f(z)/z and
603 * then only use even powers. This is another minor optimization, since
604 * we actually WANT f(z)/z, because it is going to be multiplied by
605 * the vector between the two atoms to get the vectorial force. The
606 * fastest flops are the ones we can avoid calculating!
608 * So, here's how it should be used:
611 * 2. Multiply by beta^2, so you get z^2=beta^2*r^2.
612 * 3. Evaluate this routine with z^2 as the argument.
613 * 4. The return value is the expression:
617 * ------------ - --------
620 * 5. Multiply the entire expression by beta^3. This will get you
622 * beta^3*2*exp(-z^2) beta^3*erf(z)
623 * ------------------ - ---------------
626 * or, switching back to r (z=r*beta):
628 * 2*beta*exp(-r^2*beta^2) erf(r*beta)
629 * ----------------------- - -----------
633 * With a bit of math exercise you should be able to confirm that
634 * this is exactly D[Erf[beta*r]/r,r] divided by r another time.
636 * 6. Add the result to 1/r^3, multiply by the product of the charges,
637 * and you have your force (divided by r). A final multiplication
638 * with the vector connecting the two particles and you have your
639 * vectorial force to add to the particles.
643 gmx_mm_pmecorrF_ps(__m128 z2)
645 const __m128 FN6 = _mm_set1_ps(-1.7357322914161492954e-8f);
646 const __m128 FN5 = _mm_set1_ps(1.4703624142580877519e-6f);
647 const __m128 FN4 = _mm_set1_ps(-0.000053401640219807709149f);
648 const __m128 FN3 = _mm_set1_ps(0.0010054721316683106153f);
649 const __m128 FN2 = _mm_set1_ps(-0.019278317264888380590f);
650 const __m128 FN1 = _mm_set1_ps(0.069670166153766424023f);
651 const __m128 FN0 = _mm_set1_ps(-0.75225204789749321333f);
653 const __m128 FD4 = _mm_set1_ps(0.0011193462567257629232f);
654 const __m128 FD3 = _mm_set1_ps(0.014866955030185295499f);
655 const __m128 FD2 = _mm_set1_ps(0.11583842382862377919f);
656 const __m128 FD1 = _mm_set1_ps(0.50736591960530292870f);
657 const __m128 FD0 = _mm_set1_ps(1.0f);
660 __m128 polyFN0, polyFN1, polyFD0, polyFD1;
662 z4 = _mm_mul_ps(z2, z2);
664 polyFD0 = _mm_macc_ps(FD4, z4, FD2);
665 polyFD1 = _mm_macc_ps(FD3, z4, FD1);
666 polyFD0 = _mm_macc_ps(polyFD0, z4, FD0);
667 polyFD0 = _mm_macc_ps(polyFD1, z2, polyFD0);
669 polyFD0 = gmx_mm_inv_ps(polyFD0);
671 polyFN0 = _mm_macc_ps(FN6, z4, FN4);
672 polyFN1 = _mm_macc_ps(FN5, z4, FN3);
673 polyFN0 = _mm_macc_ps(polyFN0, z4, FN2);
674 polyFN1 = _mm_macc_ps(polyFN1, z4, FN1);
675 polyFN0 = _mm_macc_ps(polyFN0, z4, FN0);
676 polyFN0 = _mm_macc_ps(polyFN1, z2, polyFN0);
678 return _mm_mul_ps(polyFN0, polyFD0);
684 /* Calculate the potential correction due to PME analytically.
686 * See gmx_mm256_pmecorrF_ps() for details about the approximation.
688 * This routine calculates Erf(z)/z, although you should provide z^2
689 * as the input argument.
691 * Here's how it should be used:
694 * 2. Multiply by beta^2, so you get z^2=beta^2*r^2.
695 * 3. Evaluate this routine with z^2 as the argument.
696 * 4. The return value is the expression:
703 * 5. Multiply the entire expression by beta and switching back to r (z=r*beta):
709 * 6. Add the result to 1/r, multiply by the product of the charges,
710 * and you have your potential.
713 gmx_mm_pmecorrV_ps(__m128 z2)
715 const __m128 VN6 = _mm_set1_ps(1.9296833005951166339e-8f);
716 const __m128 VN5 = _mm_set1_ps(-1.4213390571557850962e-6f);
717 const __m128 VN4 = _mm_set1_ps(0.000041603292906656984871f);
718 const __m128 VN3 = _mm_set1_ps(-0.00013134036773265025626f);
719 const __m128 VN2 = _mm_set1_ps(0.038657983986041781264f);
720 const __m128 VN1 = _mm_set1_ps(0.11285044772717598220f);
721 const __m128 VN0 = _mm_set1_ps(1.1283802385263030286f);
723 const __m128 VD3 = _mm_set1_ps(0.0066752224023576045451f);
724 const __m128 VD2 = _mm_set1_ps(0.078647795836373922256f);
725 const __m128 VD1 = _mm_set1_ps(0.43336185284710920150f);
726 const __m128 VD0 = _mm_set1_ps(1.0f);
729 __m128 polyVN0, polyVN1, polyVD0, polyVD1;
731 z4 = _mm_mul_ps(z2, z2);
733 polyVD1 = _mm_macc_ps(VD3, z4, VD1);
734 polyVD0 = _mm_macc_ps(VD2, z4, VD0);
735 polyVD0 = _mm_macc_ps(polyVD1, z2, polyVD0);
737 polyVD0 = gmx_mm_inv_ps(polyVD0);
739 polyVN0 = _mm_macc_ps(VN6, z4, VN4);
740 polyVN1 = _mm_macc_ps(VN5, z4, VN3);
741 polyVN0 = _mm_macc_ps(polyVN0, z4, VN2);
742 polyVN1 = _mm_macc_ps(polyVN1, z4, VN1);
743 polyVN0 = _mm_macc_ps(polyVN0, z4, VN0);
744 polyVN0 = _mm_macc_ps(polyVN1, z2, polyVN0);
746 return _mm_mul_ps(polyVN0, polyVD0);
752 gmx_mm_sincos_ps(__m128 x,
756 const __m128 two_over_pi = _mm_set1_ps(2.0/M_PI);
757 const __m128 half = _mm_set1_ps(0.5);
758 const __m128 one = _mm_set1_ps(1.0);
760 const __m128i izero = _mm_set1_epi32(0);
761 const __m128i ione = _mm_set1_epi32(1);
762 const __m128i itwo = _mm_set1_epi32(2);
763 const __m128i ithree = _mm_set1_epi32(3);
764 const __m128 signbit = gmx_mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
766 const __m128 CA1 = _mm_set1_ps(1.5703125f);
767 const __m128 CA2 = _mm_set1_ps(4.837512969970703125e-4f);
768 const __m128 CA3 = _mm_set1_ps(7.54978995489188216e-8f);
770 const __m128 CC0 = _mm_set1_ps(-0.0013602249f);
771 const __m128 CC1 = _mm_set1_ps(0.0416566950f);
772 const __m128 CC2 = _mm_set1_ps(-0.4999990225f);
773 const __m128 CS0 = _mm_set1_ps(-0.0001950727f);
774 const __m128 CS1 = _mm_set1_ps(0.0083320758f);
775 const __m128 CS2 = _mm_set1_ps(-0.1666665247f);
780 __m128i offset_sin, offset_cos;
782 __m128 mask_sin, mask_cos;
783 __m128 tmp_sin, tmp_cos;
785 y = _mm_mul_ps(x, two_over_pi);
786 y = _mm_add_ps(y, _mm_or_ps(_mm_and_ps(y, signbit), half));
788 iz = _mm_cvttps_epi32(y);
789 z = _mm_round_ps(y, _MM_FROUND_TO_ZERO);
791 offset_sin = _mm_and_si128(iz, ithree);
792 offset_cos = _mm_add_epi32(iz, ione);
794 /* Extended precision arithmethic to achieve full precision */
795 y = _mm_nmacc_ps(z, CA1, x);
796 y = _mm_nmacc_ps(z, CA2, y);
797 y = _mm_nmacc_ps(z, CA3, y);
799 y2 = _mm_mul_ps(y, y);
801 tmp1 = _mm_macc_ps(CC0, y2, CC1);
802 tmp2 = _mm_macc_ps(CS0, y2, CS1);
803 tmp1 = _mm_macc_ps(tmp1, y2, CC2);
804 tmp2 = _mm_macc_ps(tmp2, y2, CS2);
806 tmp1 = _mm_macc_ps(tmp1, y2, one);
808 tmp2 = _mm_macc_ps(tmp2, _mm_mul_ps(y, y2), y);
810 mask_sin = gmx_mm_castsi128_ps(_mm_cmpeq_epi32( _mm_and_si128(offset_sin, ione), izero));
811 mask_cos = gmx_mm_castsi128_ps(_mm_cmpeq_epi32( _mm_and_si128(offset_cos, ione), izero));
813 tmp_sin = _mm_blendv_ps(tmp1, tmp2, mask_sin);
814 tmp_cos = _mm_blendv_ps(tmp1, tmp2, mask_cos);
816 mask_sin = gmx_mm_castsi128_ps(_mm_cmpeq_epi32( _mm_and_si128(offset_sin, itwo), izero));
817 mask_cos = gmx_mm_castsi128_ps(_mm_cmpeq_epi32( _mm_and_si128(offset_cos, itwo), izero));
819 tmp1 = _mm_xor_ps(signbit, tmp_sin);
820 tmp2 = _mm_xor_ps(signbit, tmp_cos);
822 *sinval = _mm_blendv_ps(tmp1, tmp_sin, mask_sin);
823 *cosval = _mm_blendv_ps(tmp2, tmp_cos, mask_cos);
829 * IMPORTANT: Do NOT call both sin & cos if you need both results, since each of them
830 * will then call the sincos() routine and waste a factor 2 in performance!
833 gmx_mm_sin_ps(__m128 x)
836 gmx_mm_sincos_ps(x, &s, &c);
841 * IMPORTANT: Do NOT call both sin & cos if you need both results, since each of them
842 * will then call the sincos() routine and waste a factor 2 in performance!
845 gmx_mm_cos_ps(__m128 x)
848 gmx_mm_sincos_ps(x, &s, &c);
854 gmx_mm_tan_ps(__m128 x)
856 __m128 sinval, cosval;
859 gmx_mm_sincos_ps(x, &sinval, &cosval);
861 tanval = _mm_mul_ps(sinval, gmx_mm_inv_ps(cosval));
868 gmx_mm_asin_ps(__m128 x)
870 /* Same algorithm as cephes library */
871 const __m128 signmask = gmx_mm_castsi128_ps( _mm_set1_epi32(0x7FFFFFFF) );
872 const __m128 limitlow = _mm_set1_ps(1e-4f);
873 const __m128 half = _mm_set1_ps(0.5f);
874 const __m128 one = _mm_set1_ps(1.0f);
875 const __m128 halfpi = _mm_set1_ps(M_PI/2.0f);
877 const __m128 CC5 = _mm_set1_ps(4.2163199048E-2f);
878 const __m128 CC4 = _mm_set1_ps(2.4181311049E-2f);
879 const __m128 CC3 = _mm_set1_ps(4.5470025998E-2f);
880 const __m128 CC2 = _mm_set1_ps(7.4953002686E-2f);
881 const __m128 CC1 = _mm_set1_ps(1.6666752422E-1f);
886 __m128 z, z1, z2, q, q1, q2;
889 sign = _mm_andnot_ps(signmask, x);
890 xabs = _mm_and_ps(x, signmask);
892 mask = _mm_cmp_ps(xabs, half, _CMP_GT_OQ);
894 z1 = _mm_mul_ps(half, _mm_sub_ps(one, xabs));
895 q1 = _mm_mul_ps(z1, gmx_mm_invsqrt_ps(z1));
896 q1 = _mm_andnot_ps(_mm_cmp_ps(xabs, one, _CMP_EQ_OQ), q1);
899 z2 = _mm_mul_ps(q2, q2);
901 z = _mm_or_ps( _mm_and_ps(mask, z1), _mm_andnot_ps(mask, z2) );
902 q = _mm_or_ps( _mm_and_ps(mask, q1), _mm_andnot_ps(mask, q2) );
904 z2 = _mm_mul_ps(z, z);
906 pA = _mm_macc_ps(CC5, z2, CC3);
907 pB = _mm_macc_ps(CC4, z2, CC2);
909 pA = _mm_macc_ps(pA, z2, CC1);
910 pA = _mm_mul_ps(pA, z);
912 z = _mm_macc_ps(pB, z2, pA);
914 z = _mm_macc_ps(z, q, q);
916 q2 = _mm_sub_ps(halfpi, z);
917 q2 = _mm_sub_ps(q2, z);
919 z = _mm_or_ps( _mm_and_ps(mask, q2), _mm_andnot_ps(mask, z) );
921 mask = _mm_cmp_ps(xabs, limitlow, _CMP_GT_OQ);
922 z = _mm_or_ps( _mm_and_ps(mask, z), _mm_andnot_ps(mask, xabs) );
924 z = _mm_xor_ps(z, sign);
931 gmx_mm_acos_ps(__m128 x)
933 const __m128 signmask = gmx_mm_castsi128_ps( _mm_set1_epi32(0x7FFFFFFF) );
934 const __m128 one_ps = _mm_set1_ps(1.0f);
935 const __m128 half_ps = _mm_set1_ps(0.5f);
936 const __m128 pi_ps = _mm_set1_ps(M_PI);
937 const __m128 halfpi_ps = _mm_set1_ps(M_PI/2.0f);
942 __m128 z, z1, z2, z3;
944 xabs = _mm_and_ps(x, signmask);
945 mask1 = _mm_cmp_ps(xabs, half_ps, _CMP_GT_OQ);
946 mask2 = _mm_cmp_ps(x, _mm_setzero_ps(), _CMP_GT_OQ);
948 z = _mm_mul_ps(half_ps, _mm_sub_ps(one_ps, xabs));
949 z = _mm_mul_ps(z, gmx_mm_invsqrt_ps(z));
950 z = _mm_andnot_ps(_mm_cmp_ps(xabs, one_ps, _CMP_EQ_OQ), z);
952 z = _mm_blendv_ps(x, z, mask1);
953 z = gmx_mm_asin_ps(z);
955 z2 = _mm_add_ps(z, z);
956 z1 = _mm_sub_ps(pi_ps, z2);
957 z3 = _mm_sub_ps(halfpi_ps, z);
959 z = _mm_blendv_ps(z1, z2, mask2);
960 z = _mm_blendv_ps(z3, z, mask1);
967 gmx_mm_atan_ps(__m128 x)
969 /* Same algorithm as cephes library */
970 const __m128 signmask = gmx_mm_castsi128_ps( _mm_set1_epi32(0x7FFFFFFF) );
971 const __m128 limit1 = _mm_set1_ps(0.414213562373095f);
972 const __m128 limit2 = _mm_set1_ps(2.414213562373095f);
973 const __m128 quarterpi = _mm_set1_ps(0.785398163397448f);
974 const __m128 halfpi = _mm_set1_ps(1.570796326794896f);
975 const __m128 mone = _mm_set1_ps(-1.0f);
976 const __m128 CC3 = _mm_set1_ps(-3.33329491539E-1f);
977 const __m128 CC5 = _mm_set1_ps(1.99777106478E-1f);
978 const __m128 CC7 = _mm_set1_ps(-1.38776856032E-1);
979 const __m128 CC9 = _mm_set1_ps(8.05374449538e-2f);
987 sign = _mm_andnot_ps(signmask, x);
988 x = _mm_and_ps(x, signmask);
990 mask1 = _mm_cmp_ps(x, limit1, _CMP_GT_OQ);
991 mask2 = _mm_cmp_ps(x, limit2, _CMP_GT_OQ);
993 z1 = _mm_mul_ps(_mm_add_ps(x, mone), gmx_mm_inv_ps(_mm_sub_ps(x, mone)));
994 z2 = _mm_mul_ps(mone, gmx_mm_inv_ps(x));
996 y = _mm_and_ps(mask1, quarterpi);
997 y = _mm_blendv_ps(y, halfpi, mask2);
999 x = _mm_blendv_ps(x, z1, mask1);
1000 x = _mm_blendv_ps(x, z2, mask2);
1002 x2 = _mm_mul_ps(x, x);
1003 x4 = _mm_mul_ps(x2, x2);
1005 sum1 = _mm_macc_ps(CC9, x4, CC5);
1006 sum2 = _mm_macc_ps(CC7, x4, CC3);
1007 sum1 = _mm_mul_ps(sum1, x4);
1008 sum1 = _mm_macc_ps(sum2, x2, sum1);
1010 sum1 = _mm_sub_ps(sum1, mone);
1011 y = _mm_macc_ps(sum1, x, y);
1013 y = _mm_xor_ps(y, sign);
1020 gmx_mm_atan2_ps(__m128 y, __m128 x)
1022 const __m128 pi = _mm_set1_ps(M_PI);
1023 const __m128 minuspi = _mm_set1_ps(-M_PI);
1024 const __m128 halfpi = _mm_set1_ps(M_PI/2.0);
1025 const __m128 minushalfpi = _mm_set1_ps(-M_PI/2.0);
1027 __m128 z, z1, z3, z4;
1029 __m128 maskx_lt, maskx_eq;
1030 __m128 masky_lt, masky_eq;
1031 __m128 mask1, mask2, mask3, mask4, maskall;
1033 maskx_lt = _mm_cmp_ps(x, _mm_setzero_ps(), _CMP_LT_OQ);
1034 masky_lt = _mm_cmp_ps(y, _mm_setzero_ps(), _CMP_LT_OQ);
1035 maskx_eq = _mm_cmp_ps(x, _mm_setzero_ps(), _CMP_EQ_OQ);
1036 masky_eq = _mm_cmp_ps(y, _mm_setzero_ps(), _CMP_EQ_OQ);
1038 z = _mm_mul_ps(y, gmx_mm_inv_ps(x));
1039 z = gmx_mm_atan_ps(z);
1041 mask1 = _mm_and_ps(maskx_eq, masky_lt);
1042 mask2 = _mm_andnot_ps(maskx_lt, masky_eq);
1043 mask3 = _mm_andnot_ps( _mm_or_ps(masky_lt, masky_eq), maskx_eq);
1044 mask4 = _mm_and_ps(masky_eq, maskx_lt);
1046 maskall = _mm_or_ps( _mm_or_ps(mask1, mask2), _mm_or_ps(mask3, mask4) );
1048 z = _mm_andnot_ps(maskall, z);
1049 z1 = _mm_and_ps(mask1, minushalfpi);
1050 z3 = _mm_and_ps(mask3, halfpi);
1051 z4 = _mm_and_ps(mask4, pi);
1053 z = _mm_or_ps( _mm_or_ps(z, z1), _mm_or_ps(z3, z4) );
1055 mask1 = _mm_andnot_ps(masky_lt, maskx_lt);
1056 mask2 = _mm_and_ps(maskx_lt, masky_lt);
1058 w = _mm_or_ps( _mm_and_ps(mask1, pi), _mm_and_ps(mask2, minuspi) );
1059 w = _mm_andnot_ps(maskall, w);
1061 z = _mm_add_ps(z, w);
1068 #endif /* _gmx_math_x86_avx_128_fma_single_h_ */