src/gromacs/legacyheaders/gmx_math_x86_sse4_1_double.h

   1 /* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
   2  *
   3  *
   4  * This file is part of GROMACS.
   5  * Copyright (c) 2012-
   6  *
   7  * Written by the Gromacs development team under coordination of
   8  * David van der Spoel, Berk Hess, and Erik Lindahl.
   9  *
  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.
  14  *
  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
  17  *
  18  * And Hey:
  19  * Gnomes, ROck Monsters And Chili Sauce
  20  */
  21 #ifndef _gmx_math_x86_sse4_1_double_h_
  22 #define _gmx_math_x86_sse4_1_double_h_
  23
  24 #include <stdio.h>
  25 #include <math.h>
  26
  27 #include "gmx_x86_sse4_1.h"
  28
  29
  30
  31 #ifndef M_PI
  32 #  define M_PI 3.14159265358979323846264338327950288
  33 #endif
  34
  35 /************************
  36  *                      *
  37  * Simple math routines *
  38  *                      *
  39  ************************/
  40
  41 /* 1.0/sqrt(x) */
  42 static gmx_inline __m128d
  43 gmx_mm_invsqrt_pd(__m128d x)
  44 {
  45     const __m128d half  = _mm_set1_pd(0.5);
  46     const __m128d three = _mm_set1_pd(3.0);
  47
  48     /* Lookup instruction only exists in single precision, convert back and forth... */
  49     __m128d lu = _mm_cvtps_pd(_mm_rsqrt_ps( _mm_cvtpd_ps(x)));
  50
  51     lu = _mm_mul_pd(half, _mm_mul_pd(_mm_sub_pd(three, _mm_mul_pd(_mm_mul_pd(lu, lu), x)), lu));
  52     return _mm_mul_pd(half, _mm_mul_pd(_mm_sub_pd(three, _mm_mul_pd(_mm_mul_pd(lu, lu), x)), lu));
  53 }
  54
  55 /* 1.0/sqrt(x), done for a pair of arguments to improve throughput */
  56 static void
  57 gmx_mm_invsqrt_pair_pd(__m128d x1, __m128d x2, __m128d *invsqrt1, __m128d *invsqrt2)
  58 {
  59     const __m128d half   = _mm_set1_pd(0.5);
  60     const __m128d three  = _mm_set1_pd(3.0);
  61     const __m128  halff  = _mm_set1_ps(0.5f);
  62     const __m128  threef = _mm_set1_ps(3.0f);
  63
  64     __m128        xf, luf;
  65     __m128d       lu1, lu2;
  66
  67     /* Do first N-R step in float for 2x throughput */
  68     xf  = _mm_shuffle_ps(_mm_cvtpd_ps(x1), _mm_cvtpd_ps(x2), _MM_SHUFFLE(1, 0, 1, 0));
  69     luf = _mm_rsqrt_ps(xf);
  70     luf = _mm_mul_ps(halff, _mm_mul_ps(_mm_sub_ps(threef, _mm_mul_ps(_mm_mul_ps(luf, luf), xf)), luf));
  71
  72     lu2 = _mm_cvtps_pd(_mm_shuffle_ps(luf, luf, _MM_SHUFFLE(3, 2, 3, 2)));
  73     lu1 = _mm_cvtps_pd(luf);
  74
  75     *invsqrt1 = _mm_mul_pd(half, _mm_mul_pd(_mm_sub_pd(three, _mm_mul_pd(_mm_mul_pd(lu1, lu1), x1)), lu1));
  76     *invsqrt2 = _mm_mul_pd(half, _mm_mul_pd(_mm_sub_pd(three, _mm_mul_pd(_mm_mul_pd(lu2, lu2), x2)), lu2));
  77 }
  78
  79 /* sqrt(x) - Do NOT use this (but rather invsqrt) if you actually need 1.0/sqrt(x) */
  80 static gmx_inline __m128d
  81 gmx_mm_sqrt_pd(__m128d x)
  82 {
  83     __m128d mask;
  84     __m128d res;
  85
  86     mask = _mm_cmpeq_pd(x, _mm_setzero_pd());
  87     res  = _mm_andnot_pd(mask, gmx_mm_invsqrt_pd(x));
  88
  89     res  = _mm_mul_pd(x, res);
  90
  91     return res;
  92 }
  93
  94 /* 1.0/x */
  95 static gmx_inline __m128d
  96 gmx_mm_inv_pd(__m128d x)
  97 {
  98     const __m128d two  = _mm_set1_pd(2.0);
  99
 100     /* Lookup instruction only exists in single precision, convert back and forth... */
 101     __m128d lu = _mm_cvtps_pd(_mm_rcp_ps( _mm_cvtpd_ps(x)));
 102
 103     /* Perform two N-R steps for double precision */
 104     lu         = _mm_mul_pd(lu, _mm_sub_pd(two, _mm_mul_pd(x, lu)));
 105     return _mm_mul_pd(lu, _mm_sub_pd(two, _mm_mul_pd(x, lu)));
 106 }
 107
 108 static gmx_inline __m128d
 109 gmx_mm_abs_pd(__m128d x)
 110 {
 111     const __m128d signmask  = gmx_mm_castsi128_pd( _mm_set_epi32(0x7FFFFFFF, 0xFFFFFFFF, 0x7FFFFFFF, 0xFFFFFFFF) );
 112
 113     return _mm_and_pd(x, signmask);
 114 }
 115
 116
 117 /*
 118  * 2^x function.
 119  *
 120  * The 2^w term is calculated from a (6,0)-th order (no denominator) Minimax polynomia on the interval
 121  * [-0.5,0.5].
 122  *
 123  * The approximation on [-0.5,0.5] is a rational Padé approximation, 1+2*P(x^2)/(Q(x^2)-P(x^2)),
 124  * according to the same algorithm as used in the Cephes/netlib math routines.
 125  */
 126 static __m128d
 127 gmx_mm_exp2_pd(__m128d x)
 128 {
 129     /* Lower bound: We do not allow numbers that would lead to an IEEE fp representation exponent smaller than -126. */
 130     const __m128d arglimit = _mm_set1_pd(1022.0);
 131     const __m128i expbase  = _mm_set1_epi32(1023);
 132
 133     const __m128d P2       = _mm_set1_pd(2.30933477057345225087e-2);
 134     const __m128d P1       = _mm_set1_pd(2.02020656693165307700e1);
 135     const __m128d P0       = _mm_set1_pd(1.51390680115615096133e3);
 136     /* Q2 == 1.0 */
 137     const __m128d Q1       = _mm_set1_pd(2.33184211722314911771e2);
 138     const __m128d Q0       = _mm_set1_pd(4.36821166879210612817e3);
 139     const __m128d one      = _mm_set1_pd(1.0);
 140     const __m128d two      = _mm_set1_pd(2.0);
 141
 142     __m128d       valuemask;
 143     __m128i       iexppart;
 144     __m128d       fexppart;
 145     __m128d       intpart;
 146     __m128d       z, z2;
 147     __m128d       PolyP, PolyQ;
 148
 149     iexppart  = _mm_cvtpd_epi32(x);
 150     intpart   = _mm_round_pd(x, _MM_FROUND_TO_NEAREST_INT);
 151
 152     /* The two lowest elements of iexppart now contains 32-bit numbers with a correctly biased exponent.
 153      * To be able to shift it into the exponent for a double precision number we first need to
 154      * shuffle so that the lower half contains the first element, and the upper half the second.
 155      * This should really be done as a zero-extension, but since the next instructions will shift
 156      * the registers left by 52 bits it doesn't matter what we put there - it will be shifted out.
 157      * (thus we just use element 2 from iexppart).
 158      */
 159     iexppart  = _mm_shuffle_epi32(iexppart, _MM_SHUFFLE(2, 1, 2, 0));
 160
 161     /* Do the shift operation on the 64-bit registers */
 162     iexppart  = _mm_add_epi32(iexppart, expbase);
 163     iexppart  = _mm_slli_epi64(iexppart, 52);
 164
 165     valuemask = _mm_cmpge_pd(arglimit, gmx_mm_abs_pd(x));
 166     fexppart  = _mm_and_pd(valuemask, gmx_mm_castsi128_pd(iexppart));
 167
 168     z         = _mm_sub_pd(x, intpart);
 169     z2        = _mm_mul_pd(z, z);
 170
 171     PolyP     = _mm_mul_pd(P2, z2);
 172     PolyP     = _mm_add_pd(PolyP, P1);
 173     PolyQ     = _mm_add_pd(z2, Q1);
 174     PolyP     = _mm_mul_pd(PolyP, z2);
 175     PolyQ     = _mm_mul_pd(PolyQ, z2);
 176     PolyP     = _mm_add_pd(PolyP, P0);
 177     PolyQ     = _mm_add_pd(PolyQ, Q0);
 178     PolyP     = _mm_mul_pd(PolyP, z);
 179
 180     z         = _mm_mul_pd(PolyP, gmx_mm_inv_pd(_mm_sub_pd(PolyQ, PolyP)));
 181     z         = _mm_add_pd(one, _mm_mul_pd(two, z));
 182
 183     z         = _mm_mul_pd(z, fexppart);
 184
 185     return z;
 186 }
 187
 188 /* Exponential function. This could be calculated from 2^x as Exp(x)=2^(y), where y=log2(e)*x,
 189  * but there will then be a small rounding error since we lose some precision due to the
 190  * multiplication. This will then be magnified a lot by the exponential.
 191  *
 192  * Instead, we calculate the fractional part directly as a Padé approximation of
 193  * Exp(z) on [-0.5,0.5]. We use extended precision arithmetics to calculate the fraction
 194  * remaining after 2^y, which avoids the precision-loss.
 195  */
 196 static __m128d
 197 gmx_mm_exp_pd(__m128d exparg)
 198 {
 199     const __m128d argscale = _mm_set1_pd(1.4426950408889634073599);
 200     /* Lower bound: We do not allow numbers that would lead to an IEEE fp representation exponent smaller than -126. */
 201     const __m128d arglimit = _mm_set1_pd(1022.0);
 202     const __m128i expbase  = _mm_set1_epi32(1023);
 203
 204     const __m128d invargscale0  = _mm_set1_pd(6.93145751953125e-1);
 205     const __m128d invargscale1  = _mm_set1_pd(1.42860682030941723212e-6);
 206
 207     const __m128d P2       = _mm_set1_pd(1.26177193074810590878e-4);
 208     const __m128d P1       = _mm_set1_pd(3.02994407707441961300e-2);
 209     /* P0 == 1.0 */
 210     const __m128d Q3       = _mm_set1_pd(3.00198505138664455042E-6);
 211     const __m128d Q2       = _mm_set1_pd(2.52448340349684104192E-3);
 212     const __m128d Q1       = _mm_set1_pd(2.27265548208155028766E-1);
 213     /* Q0 == 2.0 */
 214     const __m128d one      = _mm_set1_pd(1.0);
 215     const __m128d two      = _mm_set1_pd(2.0);
 216
 217     __m128d       valuemask;
 218     __m128i       iexppart;
 219     __m128d       fexppart;
 220     __m128d       intpart;
 221     __m128d       x, z, z2;
 222     __m128d       PolyP, PolyQ;
 223
 224     x             = _mm_mul_pd(exparg, argscale);
 225
 226     iexppart  = _mm_cvtpd_epi32(x);
 227     intpart   = _mm_round_pd(x, _MM_FROUND_TO_NEAREST_INT);
 228
 229     /* The two lowest elements of iexppart now contains 32-bit numbers with a correctly biased exponent.
 230      * To be able to shift it into the exponent for a double precision number we first need to
 231      * shuffle so that the lower half contains the first element, and the upper half the second.
 232      * This should really be done as a zero-extension, but since the next instructions will shift
 233      * the registers left by 52 bits it doesn't matter what we put there - it will be shifted out.
 234      * (thus we just use element 2 from iexppart).
 235      */
 236     iexppart  = _mm_shuffle_epi32(iexppart, _MM_SHUFFLE(2, 1, 2, 0));
 237
 238     /* Do the shift operation on the 64-bit registers */
 239     iexppart  = _mm_add_epi32(iexppart, expbase);
 240     iexppart  = _mm_slli_epi64(iexppart, 52);
 241
 242     valuemask = _mm_cmpge_pd(arglimit, gmx_mm_abs_pd(x));
 243     fexppart  = _mm_and_pd(valuemask, gmx_mm_castsi128_pd(iexppart));
 244
 245     z         = _mm_sub_pd(exparg, _mm_mul_pd(invargscale0, intpart));
 246     z         = _mm_sub_pd(z, _mm_mul_pd(invargscale1, intpart));
 247
 248     z2        = _mm_mul_pd(z, z);
 249
 250     PolyQ     = _mm_mul_pd(Q3, z2);
 251     PolyQ     = _mm_add_pd(PolyQ, Q2);
 252     PolyP     = _mm_mul_pd(P2, z2);
 253     PolyQ     = _mm_mul_pd(PolyQ, z2);
 254     PolyP     = _mm_add_pd(PolyP, P1);
 255     PolyQ     = _mm_add_pd(PolyQ, Q1);
 256     PolyP     = _mm_mul_pd(PolyP, z2);
 257     PolyQ     = _mm_mul_pd(PolyQ, z2);
 258     PolyP     = _mm_add_pd(PolyP, one);
 259     PolyQ     = _mm_add_pd(PolyQ, two);
 260
 261     PolyP     = _mm_mul_pd(PolyP, z);
 262
 263     z         = _mm_mul_pd(PolyP, gmx_mm_inv_pd(_mm_sub_pd(PolyQ, PolyP)));
 264     z         = _mm_add_pd(one, _mm_mul_pd(two, z));
 265
 266     z         = _mm_mul_pd(z, fexppart);
 267
 268     return z;
 269 }
 270
 271
 272
 273 static __m128d
 274 gmx_mm_log_pd(__m128d x)
 275 {
 276     /* Same algorithm as cephes library */
 277     const __m128d expmask    = gmx_mm_castsi128_pd( _mm_set_epi32(0x7FF00000, 0x00000000, 0x7FF00000, 0x00000000) );
 278
 279     const __m128i expbase_m1 = _mm_set1_epi32(1023-1); /* We want non-IEEE format */
 280
 281     const __m128d half       = _mm_set1_pd(0.5);
 282     const __m128d one        = _mm_set1_pd(1.0);
 283     const __m128d two        = _mm_set1_pd(2.0);
 284     const __m128d invsq2     = _mm_set1_pd(1.0/sqrt(2.0));
 285
 286     const __m128d corr1      = _mm_set1_pd(-2.121944400546905827679e-4);
 287     const __m128d corr2      = _mm_set1_pd(0.693359375);
 288
 289     const __m128d P5         = _mm_set1_pd(1.01875663804580931796e-4);
 290     const __m128d P4         = _mm_set1_pd(4.97494994976747001425e-1);
 291     const __m128d P3         = _mm_set1_pd(4.70579119878881725854e0);
 292     const __m128d P2         = _mm_set1_pd(1.44989225341610930846e1);
 293     const __m128d P1         = _mm_set1_pd(1.79368678507819816313e1);
 294     const __m128d P0         = _mm_set1_pd(7.70838733755885391666e0);
 295
 296     const __m128d Q4         = _mm_set1_pd(1.12873587189167450590e1);
 297     const __m128d Q3         = _mm_set1_pd(4.52279145837532221105e1);
 298     const __m128d Q2         = _mm_set1_pd(8.29875266912776603211e1);
 299     const __m128d Q1         = _mm_set1_pd(7.11544750618563894466e1);
 300     const __m128d Q0         = _mm_set1_pd(2.31251620126765340583e1);
 301
 302     const __m128d R2         = _mm_set1_pd(-7.89580278884799154124e-1);
 303     const __m128d R1         = _mm_set1_pd(1.63866645699558079767e1);
 304     const __m128d R0         = _mm_set1_pd(-6.41409952958715622951e1);
 305
 306     const __m128d S2         = _mm_set1_pd(-3.56722798256324312549E1);
 307     const __m128d S1         = _mm_set1_pd(3.12093766372244180303E2);
 308     const __m128d S0         = _mm_set1_pd(-7.69691943550460008604E2);
 309
 310     __m128d       fexp;
 311     __m128i       iexp;
 312
 313     __m128d       mask1, mask2;
 314     __m128d       corr, t1, t2, q;
 315     __m128d       zA, yA, xA, zB, yB, xB, z;
 316     __m128d       polyR, polyS;
 317     __m128d       polyP1, polyP2, polyQ1, polyQ2;
 318
 319     /* Separate x into exponent and mantissa, with a mantissa in the range [0.5..1[ (not IEEE754 standard!) */
 320     fexp   = _mm_and_pd(x, expmask);
 321     iexp   = gmx_mm_castpd_si128(fexp);
 322     iexp   = _mm_srli_epi64(iexp, 52);
 323     iexp   = _mm_sub_epi32(iexp, expbase_m1);
 324     iexp   = _mm_shuffle_epi32(iexp, _MM_SHUFFLE(1, 1, 2, 0) );
 325     fexp   = _mm_cvtepi32_pd(iexp);
 326
 327     x      = _mm_andnot_pd(expmask, x);
 328     x      = _mm_or_pd(x, one);
 329     x      = _mm_mul_pd(x, half);
 330
 331     mask1     = _mm_cmpgt_pd(gmx_mm_abs_pd(fexp), two);
 332     mask2     = _mm_cmplt_pd(x, invsq2);
 333
 334     fexp   = _mm_sub_pd(fexp, _mm_and_pd(mask2, one));
 335
 336     /* If mask1 is set ('A') */
 337     zA     = _mm_sub_pd(x, half);
 338     t1     = _mm_blendv_pd( zA, x, mask2 );
 339     zA     = _mm_sub_pd(t1, half);
 340     t2     = _mm_blendv_pd( x, zA, mask2 );
 341     yA     = _mm_mul_pd(half, _mm_add_pd(t2, one));
 342
 343     xA     = _mm_mul_pd(zA, gmx_mm_inv_pd(yA));
 344     zA     = _mm_mul_pd(xA, xA);
 345
 346     /* EVALUATE POLY */
 347     polyR  = _mm_mul_pd(R2, zA);
 348     polyR  = _mm_add_pd(polyR, R1);
 349     polyR  = _mm_mul_pd(polyR, zA);
 350     polyR  = _mm_add_pd(polyR, R0);
 351
 352     polyS  = _mm_add_pd(zA, S2);
 353     polyS  = _mm_mul_pd(polyS, zA);
 354     polyS  = _mm_add_pd(polyS, S1);
 355     polyS  = _mm_mul_pd(polyS, zA);
 356     polyS  = _mm_add_pd(polyS, S0);
 357
 358     q      = _mm_mul_pd(polyR, gmx_mm_inv_pd(polyS));
 359     zA     = _mm_mul_pd(_mm_mul_pd(xA, zA), q);
 360
 361     zA     = _mm_add_pd(zA, _mm_mul_pd(corr1, fexp));
 362     zA     = _mm_add_pd(zA, xA);
 363     zA     = _mm_add_pd(zA, _mm_mul_pd(corr2, fexp));
 364
 365     /* If mask1 is not set ('B') */
 366     corr   = _mm_and_pd(mask2, x);
 367     xB     = _mm_add_pd(x, corr);
 368     xB     = _mm_sub_pd(xB, one);
 369     zB     = _mm_mul_pd(xB, xB);
 370
 371     polyP1 = _mm_mul_pd(P5, zB);
 372     polyP2 = _mm_mul_pd(P4, zB);
 373     polyP1 = _mm_add_pd(polyP1, P3);
 374     polyP2 = _mm_add_pd(polyP2, P2);
 375     polyP1 = _mm_mul_pd(polyP1, zB);
 376     polyP2 = _mm_mul_pd(polyP2, zB);
 377     polyP1 = _mm_add_pd(polyP1, P1);
 378     polyP2 = _mm_add_pd(polyP2, P0);
 379     polyP1 = _mm_mul_pd(polyP1, xB);
 380     polyP1 = _mm_add_pd(polyP1, polyP2);
 381
 382     polyQ2 = _mm_mul_pd(Q4, zB);
 383     polyQ1 = _mm_add_pd(zB, Q3);
 384     polyQ2 = _mm_add_pd(polyQ2, Q2);
 385     polyQ1 = _mm_mul_pd(polyQ1, zB);
 386     polyQ2 = _mm_mul_pd(polyQ2, zB);
 387     polyQ1 = _mm_add_pd(polyQ1, Q1);
 388     polyQ2 = _mm_add_pd(polyQ2, Q0);
 389     polyQ1 = _mm_mul_pd(polyQ1, xB);
 390     polyQ1 = _mm_add_pd(polyQ1, polyQ2);
 391
 392     fexp   = _mm_and_pd(fexp, _mm_cmpneq_pd(fexp, _mm_setzero_pd()));
 393
 394     q      = _mm_mul_pd(polyP1, gmx_mm_inv_pd(polyQ1));
 395     yB     = _mm_mul_pd(_mm_mul_pd(xB, zB), q);
 396
 397     yB     = _mm_add_pd(yB, _mm_mul_pd(corr1, fexp));
 398     yB     = _mm_sub_pd(yB, _mm_mul_pd(half, zB));
 399     zB     = _mm_add_pd(xB, yB);
 400     zB     = _mm_add_pd(zB, _mm_mul_pd(corr2, fexp));
 401
 402     z      = _mm_blendv_pd( zB, zA, mask1 );
 403
 404     return z;
 405 }
 406
 407
 408 static __m128d
 409 gmx_mm_erf_pd(__m128d x)
 410 {
 411     /* Coefficients for minimax approximation of erf(x)=x*(CAoffset + P(x^2)/Q(x^2)) in range [-0.75,0.75] */
 412     const __m128d CAP4      = _mm_set1_pd(-0.431780540597889301512e-4);
 413     const __m128d CAP3      = _mm_set1_pd(-0.00578562306260059236059);
 414     const __m128d CAP2      = _mm_set1_pd(-0.028593586920219752446);
 415     const __m128d CAP1      = _mm_set1_pd(-0.315924962948621698209);
 416     const __m128d CAP0      = _mm_set1_pd(0.14952975608477029151);
 417
 418     const __m128d CAQ5      = _mm_set1_pd(-0.374089300177174709737e-5);
 419     const __m128d CAQ4      = _mm_set1_pd(0.00015126584532155383535);
 420     const __m128d CAQ3      = _mm_set1_pd(0.00536692680669480725423);
 421     const __m128d CAQ2      = _mm_set1_pd(0.0668686825594046122636);
 422     const __m128d CAQ1      = _mm_set1_pd(0.402604990869284362773);
 423     /* CAQ0 == 1.0 */
 424     const __m128d CAoffset  = _mm_set1_pd(0.9788494110107421875);
 425
 426     /* Coefficients for minimax approximation of erfc(x)=exp(-x^2)*x*(P(x-1)/Q(x-1)) in range [1.0,4.5] */
 427     const __m128d CBP6      = _mm_set1_pd(2.49650423685462752497647637088e-10);
 428     const __m128d CBP5      = _mm_set1_pd(0.00119770193298159629350136085658);
 429     const __m128d CBP4      = _mm_set1_pd(0.0164944422378370965881008942733);
 430     const __m128d CBP3      = _mm_set1_pd(0.0984581468691775932063932439252);
 431     const __m128d CBP2      = _mm_set1_pd(0.317364595806937763843589437418);
 432     const __m128d CBP1      = _mm_set1_pd(0.554167062641455850932670067075);
 433     const __m128d CBP0      = _mm_set1_pd(0.427583576155807163756925301060);
 434     const __m128d CBQ7      = _mm_set1_pd(0.00212288829699830145976198384930);
 435     const __m128d CBQ6      = _mm_set1_pd(0.0334810979522685300554606393425);
 436     const __m128d CBQ5      = _mm_set1_pd(0.2361713785181450957579508850717);
 437     const __m128d CBQ4      = _mm_set1_pd(0.955364736493055670530981883072);
 438     const __m128d CBQ3      = _mm_set1_pd(2.36815675631420037315349279199);
 439     const __m128d CBQ2      = _mm_set1_pd(3.55261649184083035537184223542);
 440     const __m128d CBQ1      = _mm_set1_pd(2.93501136050160872574376997993);
 441     /* CBQ0 == 1.0 */
 442
 443     /* Coefficients for minimax approximation of erfc(x)=exp(-x^2)/x*(P(1/x)/Q(1/x)) in range [4.5,inf] */
 444     const __m128d CCP6      = _mm_set1_pd(-2.8175401114513378771);
 445     const __m128d CCP5      = _mm_set1_pd(-3.22729451764143718517);
 446     const __m128d CCP4      = _mm_set1_pd(-2.5518551727311523996);
 447     const __m128d CCP3      = _mm_set1_pd(-0.687717681153649930619);
 448     const __m128d CCP2      = _mm_set1_pd(-0.212652252872804219852);
 449     const __m128d CCP1      = _mm_set1_pd(0.0175389834052493308818);
 450     const __m128d CCP0      = _mm_set1_pd(0.00628057170626964891937);
 451
 452     const __m128d CCQ6      = _mm_set1_pd(5.48409182238641741584);
 453     const __m128d CCQ5      = _mm_set1_pd(13.5064170191802889145);
 454     const __m128d CCQ4      = _mm_set1_pd(22.9367376522880577224);
 455     const __m128d CCQ3      = _mm_set1_pd(15.930646027911794143);
 456     const __m128d CCQ2      = _mm_set1_pd(11.0567237927800161565);
 457     const __m128d CCQ1      = _mm_set1_pd(2.79257750980575282228);
 458     /* CCQ0 == 1.0 */
 459     const __m128d CCoffset  = _mm_set1_pd(0.5579090118408203125);
 460
 461     const __m128d one       = _mm_set1_pd(1.0);
 462     const __m128d two       = _mm_set1_pd(2.0);
 463
 464     const __m128d signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000, 0x00000000, 0x80000000, 0x00000000) );
 465
 466     __m128d       xabs, x2, x4, t, t2, w, w2;
 467     __m128d       PolyAP0, PolyAP1, PolyAQ0, PolyAQ1;
 468     __m128d       PolyBP0, PolyBP1, PolyBQ0, PolyBQ1;
 469     __m128d       PolyCP0, PolyCP1, PolyCQ0, PolyCQ1;
 470     __m128d       res_erf, res_erfcB, res_erfcC, res_erfc, res;
 471     __m128d       mask, expmx2;
 472
 473     /* Calculate erf() */
 474     xabs     = gmx_mm_abs_pd(x);
 475     x2       = _mm_mul_pd(x, x);
 476     x4       = _mm_mul_pd(x2, x2);
 477
 478     PolyAP0  = _mm_mul_pd(CAP4, x4);
 479     PolyAP1  = _mm_mul_pd(CAP3, x4);
 480     PolyAP0  = _mm_add_pd(PolyAP0, CAP2);
 481     PolyAP1  = _mm_add_pd(PolyAP1, CAP1);
 482     PolyAP0  = _mm_mul_pd(PolyAP0, x4);
 483     PolyAP1  = _mm_mul_pd(PolyAP1, x2);
 484     PolyAP0  = _mm_add_pd(PolyAP0, CAP0);
 485     PolyAP0  = _mm_add_pd(PolyAP0, PolyAP1);
 486
 487     PolyAQ1  = _mm_mul_pd(CAQ5, x4);
 488     PolyAQ0  = _mm_mul_pd(CAQ4, x4);
 489     PolyAQ1  = _mm_add_pd(PolyAQ1, CAQ3);
 490     PolyAQ0  = _mm_add_pd(PolyAQ0, CAQ2);
 491     PolyAQ1  = _mm_mul_pd(PolyAQ1, x4);
 492     PolyAQ0  = _mm_mul_pd(PolyAQ0, x4);
 493     PolyAQ1  = _mm_add_pd(PolyAQ1, CAQ1);
 494     PolyAQ0  = _mm_add_pd(PolyAQ0, one);
 495     PolyAQ1  = _mm_mul_pd(PolyAQ1, x2);
 496     PolyAQ0  = _mm_add_pd(PolyAQ0, PolyAQ1);
 497
 498     res_erf  = _mm_mul_pd(PolyAP0, gmx_mm_inv_pd(PolyAQ0));
 499     res_erf  = _mm_add_pd(CAoffset, res_erf);
 500     res_erf  = _mm_mul_pd(x, res_erf);
 501
 502     /* Calculate erfc() in range [1,4.5] */
 503     t       = _mm_sub_pd(xabs, one);
 504     t2      = _mm_mul_pd(t, t);
 505
 506     PolyBP0  = _mm_mul_pd(CBP6, t2);
 507     PolyBP1  = _mm_mul_pd(CBP5, t2);
 508     PolyBP0  = _mm_add_pd(PolyBP0, CBP4);
 509     PolyBP1  = _mm_add_pd(PolyBP1, CBP3);
 510     PolyBP0  = _mm_mul_pd(PolyBP0, t2);
 511     PolyBP1  = _mm_mul_pd(PolyBP1, t2);
 512     PolyBP0  = _mm_add_pd(PolyBP0, CBP2);
 513     PolyBP1  = _mm_add_pd(PolyBP1, CBP1);
 514     PolyBP0  = _mm_mul_pd(PolyBP0, t2);
 515     PolyBP1  = _mm_mul_pd(PolyBP1, t);
 516     PolyBP0  = _mm_add_pd(PolyBP0, CBP0);
 517     PolyBP0  = _mm_add_pd(PolyBP0, PolyBP1);
 518
 519     PolyBQ1 = _mm_mul_pd(CBQ7, t2);
 520     PolyBQ0 = _mm_mul_pd(CBQ6, t2);
 521     PolyBQ1 = _mm_add_pd(PolyBQ1, CBQ5);
 522     PolyBQ0 = _mm_add_pd(PolyBQ0, CBQ4);
 523     PolyBQ1 = _mm_mul_pd(PolyBQ1, t2);
 524     PolyBQ0 = _mm_mul_pd(PolyBQ0, t2);
 525     PolyBQ1 = _mm_add_pd(PolyBQ1, CBQ3);
 526     PolyBQ0 = _mm_add_pd(PolyBQ0, CBQ2);
 527     PolyBQ1 = _mm_mul_pd(PolyBQ1, t2);
 528     PolyBQ0 = _mm_mul_pd(PolyBQ0, t2);
 529     PolyBQ1 = _mm_add_pd(PolyBQ1, CBQ1);
 530     PolyBQ0 = _mm_add_pd(PolyBQ0, one);
 531     PolyBQ1 = _mm_mul_pd(PolyBQ1, t);
 532     PolyBQ0 = _mm_add_pd(PolyBQ0, PolyBQ1);
 533
 534     res_erfcB = _mm_mul_pd(PolyBP0, gmx_mm_inv_pd(PolyBQ0));
 535
 536     res_erfcB = _mm_mul_pd(res_erfcB, xabs);
 537
 538     /* Calculate erfc() in range [4.5,inf] */
 539     w       = gmx_mm_inv_pd(xabs);
 540     w2      = _mm_mul_pd(w, w);
 541
 542     PolyCP0  = _mm_mul_pd(CCP6, w2);
 543     PolyCP1  = _mm_mul_pd(CCP5, w2);
 544     PolyCP0  = _mm_add_pd(PolyCP0, CCP4);
 545     PolyCP1  = _mm_add_pd(PolyCP1, CCP3);
 546     PolyCP0  = _mm_mul_pd(PolyCP0, w2);
 547     PolyCP1  = _mm_mul_pd(PolyCP1, w2);
 548     PolyCP0  = _mm_add_pd(PolyCP0, CCP2);
 549     PolyCP1  = _mm_add_pd(PolyCP1, CCP1);
 550     PolyCP0  = _mm_mul_pd(PolyCP0, w2);
 551     PolyCP1  = _mm_mul_pd(PolyCP1, w);
 552     PolyCP0  = _mm_add_pd(PolyCP0, CCP0);
 553     PolyCP0  = _mm_add_pd(PolyCP0, PolyCP1);
 554
 555     PolyCQ0  = _mm_mul_pd(CCQ6, w2);
 556     PolyCQ1  = _mm_mul_pd(CCQ5, w2);
 557     PolyCQ0  = _mm_add_pd(PolyCQ0, CCQ4);
 558     PolyCQ1  = _mm_add_pd(PolyCQ1, CCQ3);
 559     PolyCQ0  = _mm_mul_pd(PolyCQ0, w2);
 560     PolyCQ1  = _mm_mul_pd(PolyCQ1, w2);
 561     PolyCQ0  = _mm_add_pd(PolyCQ0, CCQ2);
 562     PolyCQ1  = _mm_add_pd(PolyCQ1, CCQ1);
 563     PolyCQ0  = _mm_mul_pd(PolyCQ0, w2);
 564     PolyCQ1  = _mm_mul_pd(PolyCQ1, w);
 565     PolyCQ0  = _mm_add_pd(PolyCQ0, one);
 566     PolyCQ0  = _mm_add_pd(PolyCQ0, PolyCQ1);
 567
 568     expmx2   = gmx_mm_exp_pd( _mm_or_pd(signbit, x2) );
 569
 570     res_erfcC = _mm_mul_pd(PolyCP0, gmx_mm_inv_pd(PolyCQ0));
 571     res_erfcC = _mm_add_pd(res_erfcC, CCoffset);
 572     res_erfcC = _mm_mul_pd(res_erfcC, w);
 573
 574     mask     = _mm_cmpgt_pd(xabs, _mm_set1_pd(4.5));
 575     res_erfc = _mm_blendv_pd(res_erfcB, res_erfcC, mask);
 576
 577     res_erfc = _mm_mul_pd(res_erfc, expmx2);
 578
 579     /* erfc(x<0) = 2-erfc(|x|) */
 580     mask     = _mm_cmplt_pd(x, _mm_setzero_pd());
 581     res_erfc = _mm_blendv_pd(res_erfc, _mm_sub_pd(two, res_erfc), mask);
 582
 583     /* Select erf() or erfc() */
 584     mask = _mm_cmplt_pd(xabs, one);
 585     res  = _mm_blendv_pd(_mm_sub_pd(one, res_erfc), res_erf, mask);
 586
 587     return res;
 588 }
 589
 590
 591 static __m128d
 592 gmx_mm_erfc_pd(__m128d x)
 593 {
 594     /* Coefficients for minimax approximation of erf(x)=x*(CAoffset + P(x^2)/Q(x^2)) in range [-0.75,0.75] */
 595     const __m128d CAP4      = _mm_set1_pd(-0.431780540597889301512e-4);
 596     const __m128d CAP3      = _mm_set1_pd(-0.00578562306260059236059);
 597     const __m128d CAP2      = _mm_set1_pd(-0.028593586920219752446);
 598     const __m128d CAP1      = _mm_set1_pd(-0.315924962948621698209);
 599     const __m128d CAP0      = _mm_set1_pd(0.14952975608477029151);
 600
 601     const __m128d CAQ5      = _mm_set1_pd(-0.374089300177174709737e-5);
 602     const __m128d CAQ4      = _mm_set1_pd(0.00015126584532155383535);
 603     const __m128d CAQ3      = _mm_set1_pd(0.00536692680669480725423);
 604     const __m128d CAQ2      = _mm_set1_pd(0.0668686825594046122636);
 605     const __m128d CAQ1      = _mm_set1_pd(0.402604990869284362773);
 606     /* CAQ0 == 1.0 */
 607     const __m128d CAoffset  = _mm_set1_pd(0.9788494110107421875);
 608
 609     /* Coefficients for minimax approximation of erfc(x)=exp(-x^2)*x*(P(x-1)/Q(x-1)) in range [1.0,4.5] */
 610     const __m128d CBP6      = _mm_set1_pd(2.49650423685462752497647637088e-10);
 611     const __m128d CBP5      = _mm_set1_pd(0.00119770193298159629350136085658);
 612     const __m128d CBP4      = _mm_set1_pd(0.0164944422378370965881008942733);
 613     const __m128d CBP3      = _mm_set1_pd(0.0984581468691775932063932439252);
 614     const __m128d CBP2      = _mm_set1_pd(0.317364595806937763843589437418);
 615     const __m128d CBP1      = _mm_set1_pd(0.554167062641455850932670067075);
 616     const __m128d CBP0      = _mm_set1_pd(0.427583576155807163756925301060);
 617     const __m128d CBQ7      = _mm_set1_pd(0.00212288829699830145976198384930);
 618     const __m128d CBQ6      = _mm_set1_pd(0.0334810979522685300554606393425);
 619     const __m128d CBQ5      = _mm_set1_pd(0.2361713785181450957579508850717);
 620     const __m128d CBQ4      = _mm_set1_pd(0.955364736493055670530981883072);
 621     const __m128d CBQ3      = _mm_set1_pd(2.36815675631420037315349279199);
 622     const __m128d CBQ2      = _mm_set1_pd(3.55261649184083035537184223542);
 623     const __m128d CBQ1      = _mm_set1_pd(2.93501136050160872574376997993);
 624     /* CBQ0 == 1.0 */
 625
 626     /* Coefficients for minimax approximation of erfc(x)=exp(-x^2)/x*(P(1/x)/Q(1/x)) in range [4.5,inf] */
 627     const __m128d CCP6      = _mm_set1_pd(-2.8175401114513378771);
 628     const __m128d CCP5      = _mm_set1_pd(-3.22729451764143718517);
 629     const __m128d CCP4      = _mm_set1_pd(-2.5518551727311523996);
 630     const __m128d CCP3      = _mm_set1_pd(-0.687717681153649930619);
 631     const __m128d CCP2      = _mm_set1_pd(-0.212652252872804219852);
 632     const __m128d CCP1      = _mm_set1_pd(0.0175389834052493308818);
 633     const __m128d CCP0      = _mm_set1_pd(0.00628057170626964891937);
 634
 635     const __m128d CCQ6      = _mm_set1_pd(5.48409182238641741584);
 636     const __m128d CCQ5      = _mm_set1_pd(13.5064170191802889145);
 637     const __m128d CCQ4      = _mm_set1_pd(22.9367376522880577224);
 638     const __m128d CCQ3      = _mm_set1_pd(15.930646027911794143);
 639     const __m128d CCQ2      = _mm_set1_pd(11.0567237927800161565);
 640     const __m128d CCQ1      = _mm_set1_pd(2.79257750980575282228);
 641     /* CCQ0 == 1.0 */
 642     const __m128d CCoffset  = _mm_set1_pd(0.5579090118408203125);
 643
 644     const __m128d one       = _mm_set1_pd(1.0);
 645     const __m128d two       = _mm_set1_pd(2.0);
 646
 647     const __m128d signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000, 0x00000000, 0x80000000, 0x00000000) );
 648
 649     __m128d       xabs, x2, x4, t, t2, w, w2;
 650     __m128d       PolyAP0, PolyAP1, PolyAQ0, PolyAQ1;
 651     __m128d       PolyBP0, PolyBP1, PolyBQ0, PolyBQ1;
 652     __m128d       PolyCP0, PolyCP1, PolyCQ0, PolyCQ1;
 653     __m128d       res_erf, res_erfcB, res_erfcC, res_erfc, res;
 654     __m128d       mask, expmx2;
 655
 656     /* Calculate erf() */
 657     xabs     = gmx_mm_abs_pd(x);
 658     x2       = _mm_mul_pd(x, x);
 659     x4       = _mm_mul_pd(x2, x2);
 660
 661     PolyAP0  = _mm_mul_pd(CAP4, x4);
 662     PolyAP1  = _mm_mul_pd(CAP3, x4);
 663     PolyAP0  = _mm_add_pd(PolyAP0, CAP2);
 664     PolyAP1  = _mm_add_pd(PolyAP1, CAP1);
 665     PolyAP0  = _mm_mul_pd(PolyAP0, x4);
 666     PolyAP1  = _mm_mul_pd(PolyAP1, x2);
 667     PolyAP0  = _mm_add_pd(PolyAP0, CAP0);
 668     PolyAP0  = _mm_add_pd(PolyAP0, PolyAP1);
 669
 670     PolyAQ1  = _mm_mul_pd(CAQ5, x4);
 671     PolyAQ0  = _mm_mul_pd(CAQ4, x4);
 672     PolyAQ1  = _mm_add_pd(PolyAQ1, CAQ3);
 673     PolyAQ0  = _mm_add_pd(PolyAQ0, CAQ2);
 674     PolyAQ1  = _mm_mul_pd(PolyAQ1, x4);
 675     PolyAQ0  = _mm_mul_pd(PolyAQ0, x4);
 676     PolyAQ1  = _mm_add_pd(PolyAQ1, CAQ1);
 677     PolyAQ0  = _mm_add_pd(PolyAQ0, one);
 678     PolyAQ1  = _mm_mul_pd(PolyAQ1, x2);
 679     PolyAQ0  = _mm_add_pd(PolyAQ0, PolyAQ1);
 680
 681     res_erf  = _mm_mul_pd(PolyAP0, gmx_mm_inv_pd(PolyAQ0));
 682     res_erf  = _mm_add_pd(CAoffset, res_erf);
 683     res_erf  = _mm_mul_pd(x, res_erf);
 684
 685     /* Calculate erfc() in range [1,4.5] */
 686     t       = _mm_sub_pd(xabs, one);
 687     t2      = _mm_mul_pd(t, t);
 688
 689     PolyBP0  = _mm_mul_pd(CBP6, t2);
 690     PolyBP1  = _mm_mul_pd(CBP5, t2);
 691     PolyBP0  = _mm_add_pd(PolyBP0, CBP4);
 692     PolyBP1  = _mm_add_pd(PolyBP1, CBP3);
 693     PolyBP0  = _mm_mul_pd(PolyBP0, t2);
 694     PolyBP1  = _mm_mul_pd(PolyBP1, t2);
 695     PolyBP0  = _mm_add_pd(PolyBP0, CBP2);
 696     PolyBP1  = _mm_add_pd(PolyBP1, CBP1);
 697     PolyBP0  = _mm_mul_pd(PolyBP0, t2);
 698     PolyBP1  = _mm_mul_pd(PolyBP1, t);
 699     PolyBP0  = _mm_add_pd(PolyBP0, CBP0);
 700     PolyBP0  = _mm_add_pd(PolyBP0, PolyBP1);
 701
 702     PolyBQ1 = _mm_mul_pd(CBQ7, t2);
 703     PolyBQ0 = _mm_mul_pd(CBQ6, t2);
 704     PolyBQ1 = _mm_add_pd(PolyBQ1, CBQ5);
 705     PolyBQ0 = _mm_add_pd(PolyBQ0, CBQ4);
 706     PolyBQ1 = _mm_mul_pd(PolyBQ1, t2);
 707     PolyBQ0 = _mm_mul_pd(PolyBQ0, t2);
 708     PolyBQ1 = _mm_add_pd(PolyBQ1, CBQ3);
 709     PolyBQ0 = _mm_add_pd(PolyBQ0, CBQ2);
 710     PolyBQ1 = _mm_mul_pd(PolyBQ1, t2);
 711     PolyBQ0 = _mm_mul_pd(PolyBQ0, t2);
 712     PolyBQ1 = _mm_add_pd(PolyBQ1, CBQ1);
 713     PolyBQ0 = _mm_add_pd(PolyBQ0, one);
 714     PolyBQ1 = _mm_mul_pd(PolyBQ1, t);
 715     PolyBQ0 = _mm_add_pd(PolyBQ0, PolyBQ1);
 716
 717     res_erfcB = _mm_mul_pd(PolyBP0, gmx_mm_inv_pd(PolyBQ0));
 718
 719     res_erfcB = _mm_mul_pd(res_erfcB, xabs);
 720
 721     /* Calculate erfc() in range [4.5,inf] */
 722     w       = gmx_mm_inv_pd(xabs);
 723     w2      = _mm_mul_pd(w, w);
 724
 725     PolyCP0  = _mm_mul_pd(CCP6, w2);
 726     PolyCP1  = _mm_mul_pd(CCP5, w2);
 727     PolyCP0  = _mm_add_pd(PolyCP0, CCP4);
 728     PolyCP1  = _mm_add_pd(PolyCP1, CCP3);
 729     PolyCP0  = _mm_mul_pd(PolyCP0, w2);
 730     PolyCP1  = _mm_mul_pd(PolyCP1, w2);
 731     PolyCP0  = _mm_add_pd(PolyCP0, CCP2);
 732     PolyCP1  = _mm_add_pd(PolyCP1, CCP1);
 733     PolyCP0  = _mm_mul_pd(PolyCP0, w2);
 734     PolyCP1  = _mm_mul_pd(PolyCP1, w);
 735     PolyCP0  = _mm_add_pd(PolyCP0, CCP0);
 736     PolyCP0  = _mm_add_pd(PolyCP0, PolyCP1);
 737
 738     PolyCQ0  = _mm_mul_pd(CCQ6, w2);
 739     PolyCQ1  = _mm_mul_pd(CCQ5, w2);
 740     PolyCQ0  = _mm_add_pd(PolyCQ0, CCQ4);
 741     PolyCQ1  = _mm_add_pd(PolyCQ1, CCQ3);
 742     PolyCQ0  = _mm_mul_pd(PolyCQ0, w2);
 743     PolyCQ1  = _mm_mul_pd(PolyCQ1, w2);
 744     PolyCQ0  = _mm_add_pd(PolyCQ0, CCQ2);
 745     PolyCQ1  = _mm_add_pd(PolyCQ1, CCQ1);
 746     PolyCQ0  = _mm_mul_pd(PolyCQ0, w2);
 747     PolyCQ1  = _mm_mul_pd(PolyCQ1, w);
 748     PolyCQ0  = _mm_add_pd(PolyCQ0, one);
 749     PolyCQ0  = _mm_add_pd(PolyCQ0, PolyCQ1);
 750
 751     expmx2   = gmx_mm_exp_pd( _mm_or_pd(signbit, x2) );
 752
 753     res_erfcC = _mm_mul_pd(PolyCP0, gmx_mm_inv_pd(PolyCQ0));
 754     res_erfcC = _mm_add_pd(res_erfcC, CCoffset);
 755     res_erfcC = _mm_mul_pd(res_erfcC, w);
 756
 757     mask     = _mm_cmpgt_pd(xabs, _mm_set1_pd(4.5));
 758     res_erfc = _mm_blendv_pd(res_erfcB, res_erfcC, mask);
 759
 760     res_erfc = _mm_mul_pd(res_erfc, expmx2);
 761
 762     /* erfc(x<0) = 2-erfc(|x|) */
 763     mask     = _mm_cmplt_pd(x, _mm_setzero_pd());
 764     res_erfc = _mm_blendv_pd(res_erfc, _mm_sub_pd(two, res_erfc), mask);
 765
 766     /* Select erf() or erfc() */
 767     mask = _mm_cmplt_pd(xabs, one);
 768     res  = _mm_blendv_pd(res_erfc, _mm_sub_pd(one, res_erf), mask);
 769
 770     return res;
 771 }
 772
 773
 774 /* Calculate the force correction due to PME analytically.
 775  *
 776  * This routine is meant to enable analytical evaluation of the
 777  * direct-space PME electrostatic force to avoid tables.
 778  *
 779  * The direct-space potential should be Erfc(beta*r)/r, but there
 780  * are some problems evaluating that:
 781  *
 782  * First, the error function is difficult (read: expensive) to
 783  * approxmiate accurately for intermediate to large arguments, and
 784  * this happens already in ranges of beta*r that occur in simulations.
 785  * Second, we now try to avoid calculating potentials in Gromacs but
 786  * use forces directly.
 787  *
 788  * We can simply things slight by noting that the PME part is really
 789  * a correction to the normal Coulomb force since Erfc(z)=1-Erf(z), i.e.
 790  *
 791  * V= 1/r - Erf(beta*r)/r
 792  *
 793  * The first term we already have from the inverse square root, so
 794  * that we can leave out of this routine.
 795  *
 796  * For pme tolerances of 1e-3 to 1e-8 and cutoffs of 0.5nm to 1.8nm,
 797  * the argument beta*r will be in the range 0.15 to ~4. Use your
 798  * favorite plotting program to realize how well-behaved Erf(z)/z is
 799  * in this range!
 800  *
 801  * We approximate f(z)=erf(z)/z with a rational minimax polynomial.
 802  * However, it turns out it is more efficient to approximate f(z)/z and
 803  * then only use even powers. This is another minor optimization, since
 804  * we actually WANT f(z)/z, because it is going to be multiplied by
 805  * the vector between the two atoms to get the vectorial force. The
 806  * fastest flops are the ones we can avoid calculating!
 807  *
 808  * So, here's how it should be used:
 809  *
 810  * 1. Calculate r^2.
 811  * 2. Multiply by beta^2, so you get z^2=beta^2*r^2.
 812  * 3. Evaluate this routine with z^2 as the argument.
 813  * 4. The return value is the expression:
 814  *
 815  *
 816  *       2*exp(-z^2)     erf(z)
 817  *       ------------ - --------
 818  *       sqrt(Pi)*z^2      z^3
 819  *
 820  * 5. Multiply the entire expression by beta^3. This will get you
 821  *
 822  *       beta^3*2*exp(-z^2)     beta^3*erf(z)
 823  *       ------------------  - ---------------
 824  *          sqrt(Pi)*z^2            z^3
 825  *
 826  *    or, switching back to r (z=r*beta):
 827  *
 828  *       2*beta*exp(-r^2*beta^2)   erf(r*beta)
 829  *       ----------------------- - -----------
 830  *            sqrt(Pi)*r^2            r^3
 831  *
 832  *
 833  *    With a bit of math exercise you should be able to confirm that
 834  *    this is exactly D[Erf[beta*r]/r,r] divided by r another time.
 835  *
 836  * 6. Add the result to 1/r^3, multiply by the product of the charges,
 837  *    and you have your force (divided by r). A final multiplication
 838  *    with the vector connecting the two particles and you have your
 839  *    vectorial force to add to the particles.
 840  *
 841  */
 842 static __m128d
 843 gmx_mm_pmecorrF_pd(__m128d z2)
 844 {
 845     const __m128d  FN10     = _mm_set1_pd(-8.0072854618360083154e-14);
 846     const __m128d  FN9      = _mm_set1_pd(1.1859116242260148027e-11);
 847     const __m128d  FN8      = _mm_set1_pd(-8.1490406329798423616e-10);
 848     const __m128d  FN7      = _mm_set1_pd(3.4404793543907847655e-8);
 849     const __m128d  FN6      = _mm_set1_pd(-9.9471420832602741006e-7);
 850     const __m128d  FN5      = _mm_set1_pd(0.000020740315999115847456);
 851     const __m128d  FN4      = _mm_set1_pd(-0.00031991745139313364005);
 852     const __m128d  FN3      = _mm_set1_pd(0.0035074449373659008203);
 853     const __m128d  FN2      = _mm_set1_pd(-0.031750380176100813405);
 854     const __m128d  FN1      = _mm_set1_pd(0.13884101728898463426);
 855     const __m128d  FN0      = _mm_set1_pd(-0.75225277815249618847);
 856
 857     const __m128d  FD5      = _mm_set1_pd(0.000016009278224355026701);
 858     const __m128d  FD4      = _mm_set1_pd(0.00051055686934806966046);
 859     const __m128d  FD3      = _mm_set1_pd(0.0081803507497974289008);
 860     const __m128d  FD2      = _mm_set1_pd(0.077181146026670287235);
 861     const __m128d  FD1      = _mm_set1_pd(0.41543303143712535988);
 862     const __m128d  FD0      = _mm_set1_pd(1.0);
 863
 864     __m128d        z4;
 865     __m128d        polyFN0, polyFN1, polyFD0, polyFD1;
 866
 867     z4             = _mm_mul_pd(z2, z2);
 868
 869     polyFD1        = _mm_mul_pd(FD5, z4);
 870     polyFD0        = _mm_mul_pd(FD4, z4);
 871     polyFD1        = _mm_add_pd(polyFD1, FD3);
 872     polyFD0        = _mm_add_pd(polyFD0, FD2);
 873     polyFD1        = _mm_mul_pd(polyFD1, z4);
 874     polyFD0        = _mm_mul_pd(polyFD0, z4);
 875     polyFD1        = _mm_add_pd(polyFD1, FD1);
 876     polyFD0        = _mm_add_pd(polyFD0, FD0);
 877     polyFD1        = _mm_mul_pd(polyFD1, z2);
 878     polyFD0        = _mm_add_pd(polyFD0, polyFD1);
 879
 880     polyFD0        = gmx_mm_inv_pd(polyFD0);
 881
 882     polyFN0        = _mm_mul_pd(FN10, z4);
 883     polyFN1        = _mm_mul_pd(FN9, z4);
 884     polyFN0        = _mm_add_pd(polyFN0, FN8);
 885     polyFN1        = _mm_add_pd(polyFN1, FN7);
 886     polyFN0        = _mm_mul_pd(polyFN0, z4);
 887     polyFN1        = _mm_mul_pd(polyFN1, z4);
 888     polyFN0        = _mm_add_pd(polyFN0, FN6);
 889     polyFN1        = _mm_add_pd(polyFN1, FN5);
 890     polyFN0        = _mm_mul_pd(polyFN0, z4);
 891     polyFN1        = _mm_mul_pd(polyFN1, z4);
 892     polyFN0        = _mm_add_pd(polyFN0, FN4);
 893     polyFN1        = _mm_add_pd(polyFN1, FN3);
 894     polyFN0        = _mm_mul_pd(polyFN0, z4);
 895     polyFN1        = _mm_mul_pd(polyFN1, z4);
 896     polyFN0        = _mm_add_pd(polyFN0, FN2);
 897     polyFN1        = _mm_add_pd(polyFN1, FN1);
 898     polyFN0        = _mm_mul_pd(polyFN0, z4);
 899     polyFN1        = _mm_mul_pd(polyFN1, z2);
 900     polyFN0        = _mm_add_pd(polyFN0, FN0);
 901     polyFN0        = _mm_add_pd(polyFN0, polyFN1);
 902
 903     return _mm_mul_pd(polyFN0, polyFD0);
 904 }
 905
 906
 907
 908
 909 /* Calculate the potential correction due to PME analytically.
 910  *
 911  * See gmx_mm256_pmecorrF_ps() for details about the approximation.
 912  *
 913  * This routine calculates Erf(z)/z, although you should provide z^2
 914  * as the input argument.
 915  *
 916  * Here's how it should be used:
 917  *
 918  * 1. Calculate r^2.
 919  * 2. Multiply by beta^2, so you get z^2=beta^2*r^2.
 920  * 3. Evaluate this routine with z^2 as the argument.
 921  * 4. The return value is the expression:
 922  *
 923  *
 924  *        erf(z)
 925  *       --------
 926  *          z
 927  *
 928  * 5. Multiply the entire expression by beta and switching back to r (z=r*beta):
 929  *
 930  *       erf(r*beta)
 931  *       -----------
 932  *           r
 933  *
 934  * 6. Subtract the result from 1/r, multiply by the product of the charges,
 935  *    and you have your potential.
 936  *
 937  */
 938 static __m128d
 939 gmx_mm_pmecorrV_pd(__m128d z2)
 940 {
 941     const __m128d  VN9      = _mm_set1_pd(-9.3723776169321855475e-13);
 942     const __m128d  VN8      = _mm_set1_pd(1.2280156762674215741e-10);
 943     const __m128d  VN7      = _mm_set1_pd(-7.3562157912251309487e-9);
 944     const __m128d  VN6      = _mm_set1_pd(2.6215886208032517509e-7);
 945     const __m128d  VN5      = _mm_set1_pd(-4.9532491651265819499e-6);
 946     const __m128d  VN4      = _mm_set1_pd(0.00025907400778966060389);
 947     const __m128d  VN3      = _mm_set1_pd(0.0010585044856156469792);
 948     const __m128d  VN2      = _mm_set1_pd(0.045247661136833092885);
 949     const __m128d  VN1      = _mm_set1_pd(0.11643931522926034421);
 950     const __m128d  VN0      = _mm_set1_pd(1.1283791671726767970);
 951
 952     const __m128d  VD5      = _mm_set1_pd(0.000021784709867336150342);
 953     const __m128d  VD4      = _mm_set1_pd(0.00064293662010911388448);
 954     const __m128d  VD3      = _mm_set1_pd(0.0096311444822588683504);
 955     const __m128d  VD2      = _mm_set1_pd(0.085608012351550627051);
 956     const __m128d  VD1      = _mm_set1_pd(0.43652499166614811084);
 957     const __m128d  VD0      = _mm_set1_pd(1.0);
 958
 959     __m128d        z4;
 960     __m128d        polyVN0, polyVN1, polyVD0, polyVD1;
 961
 962     z4             = _mm_mul_pd(z2, z2);
 963
 964     polyVD1        = _mm_mul_pd(VD5, z4);
 965     polyVD0        = _mm_mul_pd(VD4, z4);
 966     polyVD1        = _mm_add_pd(polyVD1, VD3);
 967     polyVD0        = _mm_add_pd(polyVD0, VD2);
 968     polyVD1        = _mm_mul_pd(polyVD1, z4);
 969     polyVD0        = _mm_mul_pd(polyVD0, z4);
 970     polyVD1        = _mm_add_pd(polyVD1, VD1);
 971     polyVD0        = _mm_add_pd(polyVD0, VD0);
 972     polyVD1        = _mm_mul_pd(polyVD1, z2);
 973     polyVD0        = _mm_add_pd(polyVD0, polyVD1);
 974
 975     polyVD0        = gmx_mm_inv_pd(polyVD0);
 976
 977     polyVN1        = _mm_mul_pd(VN9, z4);
 978     polyVN0        = _mm_mul_pd(VN8, z4);
 979     polyVN1        = _mm_add_pd(polyVN1, VN7);
 980     polyVN0        = _mm_add_pd(polyVN0, VN6);
 981     polyVN1        = _mm_mul_pd(polyVN1, z4);
 982     polyVN0        = _mm_mul_pd(polyVN0, z4);
 983     polyVN1        = _mm_add_pd(polyVN1, VN5);
 984     polyVN0        = _mm_add_pd(polyVN0, VN4);
 985     polyVN1        = _mm_mul_pd(polyVN1, z4);
 986     polyVN0        = _mm_mul_pd(polyVN0, z4);
 987     polyVN1        = _mm_add_pd(polyVN1, VN3);
 988     polyVN0        = _mm_add_pd(polyVN0, VN2);
 989     polyVN1        = _mm_mul_pd(polyVN1, z4);
 990     polyVN0        = _mm_mul_pd(polyVN0, z4);
 991     polyVN1        = _mm_add_pd(polyVN1, VN1);
 992     polyVN0        = _mm_add_pd(polyVN0, VN0);
 993     polyVN1        = _mm_mul_pd(polyVN1, z2);
 994     polyVN0        = _mm_add_pd(polyVN0, polyVN1);
 995
 996     return _mm_mul_pd(polyVN0, polyVD0);
 997 }
 998
 999
1000 static int
1001 gmx_mm_sincos_pd(__m128d  x,
1002                  __m128d *sinval,
1003                  __m128d *cosval)
1004 {
1005 #ifdef _MSC_VER
1006     __declspec(align(16))
1007     const double sintable[34] =
1008     {
1009         1.00000000000000000e+00, 0.00000000000000000e+00,
1010         9.95184726672196929e-01, 9.80171403295606036e-02,
1011         9.80785280403230431e-01, 1.95090322016128248e-01,
1012         9.56940335732208824e-01, 2.90284677254462331e-01,
1013         9.23879532511286738e-01, 3.82683432365089782e-01,
1014         8.81921264348355050e-01, 4.71396736825997642e-01,
1015         8.31469612302545236e-01, 5.55570233019602178e-01,
1016         7.73010453362736993e-01, 6.34393284163645488e-01,
1017         7.07106781186547573e-01, 7.07106781186547462e-01,
1018         6.34393284163645599e-01, 7.73010453362736882e-01,
1019         5.55570233019602289e-01, 8.31469612302545125e-01,
1020         4.71396736825997809e-01, 8.81921264348354939e-01,
1021         3.82683432365089837e-01, 9.23879532511286738e-01,
1022         2.90284677254462276e-01, 9.56940335732208935e-01,
1023         1.95090322016128304e-01, 9.80785280403230431e-01,
1024         9.80171403295607702e-02, 9.95184726672196818e-01,
1025         0.0, 1.00000000000000000e+00
1026     };
1027 #else
1028     const __m128d sintable[17] =
1029     {
1030         _mm_set_pd( 0.0, 1.0 ),
1031         _mm_set_pd( sin(  1.0 * (M_PI/2.0) / 16.0), cos(  1.0 * (M_PI/2.0) / 16.0) ),
1032         _mm_set_pd( sin(  2.0 * (M_PI/2.0) / 16.0), cos(  2.0 * (M_PI/2.0) / 16.0) ),
1033         _mm_set_pd( sin(  3.0 * (M_PI/2.0) / 16.0), cos(  3.0 * (M_PI/2.0) / 16.0) ),
1034         _mm_set_pd( sin(  4.0 * (M_PI/2.0) / 16.0), cos(  4.0 * (M_PI/2.0) / 16.0) ),
1035         _mm_set_pd( sin(  5.0 * (M_PI/2.0) / 16.0), cos(  5.0 * (M_PI/2.0) / 16.0) ),
1036         _mm_set_pd( sin(  6.0 * (M_PI/2.0) / 16.0), cos(  6.0 * (M_PI/2.0) / 16.0) ),
1037         _mm_set_pd( sin(  7.0 * (M_PI/2.0) / 16.0), cos(  7.0 * (M_PI/2.0) / 16.0) ),
1038         _mm_set_pd( sin(  8.0 * (M_PI/2.0) / 16.0), cos(  8.0 * (M_PI/2.0) / 16.0) ),
1039         _mm_set_pd( sin(  9.0 * (M_PI/2.0) / 16.0), cos(  9.0 * (M_PI/2.0) / 16.0) ),
1040         _mm_set_pd( sin( 10.0 * (M_PI/2.0) / 16.0), cos( 10.0 * (M_PI/2.0) / 16.0) ),
1041         _mm_set_pd( sin( 11.0 * (M_PI/2.0) / 16.0), cos( 11.0 * (M_PI/2.0) / 16.0) ),
1042         _mm_set_pd( sin( 12.0 * (M_PI/2.0) / 16.0), cos( 12.0 * (M_PI/2.0) / 16.0) ),
1043         _mm_set_pd( sin( 13.0 * (M_PI/2.0) / 16.0), cos( 13.0 * (M_PI/2.0) / 16.0) ),
1044         _mm_set_pd( sin( 14.0 * (M_PI/2.0) / 16.0), cos( 14.0 * (M_PI/2.0) / 16.0) ),
1045         _mm_set_pd( sin( 15.0 * (M_PI/2.0) / 16.0), cos( 15.0 * (M_PI/2.0) / 16.0) ),
1046         _mm_set_pd(  1.0, 0.0 )
1047     };
1048 #endif
1049
1050     const __m128d signmask       = gmx_mm_castsi128_pd( _mm_set_epi32(0x7FFFFFFF, 0xFFFFFFFF, 0x7FFFFFFF, 0xFFFFFFFF) );
1051
1052     const __m128d tabscale      = _mm_set1_pd(32.0/M_PI);
1053     const __m128d invtabscale0  = _mm_set1_pd(9.81747508049011230469e-02);
1054     const __m128d invtabscale1  = _mm_set1_pd(1.96197799156550576057e-08);
1055     const __m128i ione          = _mm_set1_epi32(1);
1056     const __m128i i32           = _mm_set1_epi32(32);
1057     const __m128i i16           = _mm_set1_epi32(16);
1058     const __m128i tabmask       = _mm_set1_epi32(0x3F);
1059     const __m128d sinP7         = _mm_set1_pd(-1.0/5040.0);
1060     const __m128d sinP5         = _mm_set1_pd(1.0/120.0);
1061     const __m128d sinP3         = _mm_set1_pd(-1.0/6.0);
1062     const __m128d sinP1         = _mm_set1_pd(1.0);
1063
1064     const __m128d cosP6         = _mm_set1_pd(-1.0/720.0);
1065     const __m128d cosP4         = _mm_set1_pd(1.0/24.0);
1066     const __m128d cosP2         = _mm_set1_pd(-1.0/2.0);
1067     const __m128d cosP0         = _mm_set1_pd(1.0);
1068
1069     __m128d       scalex;
1070     __m128i       tabidx, corridx;
1071     __m128d       xabs, z, z2, polySin, polyCos;
1072     __m128d       xpoint;
1073     __m128d       ypoint0, ypoint1;
1074
1075     __m128d       sinpoint, cospoint;
1076     __m128d       xsign, ssign, csign;
1077     __m128i       imask, sswapsign, cswapsign;
1078
1079     xsign    = _mm_andnot_pd(signmask, x);
1080     xabs     = _mm_and_pd(x, signmask);
1081
1082     scalex   = _mm_mul_pd(tabscale, xabs);
1083     tabidx   = _mm_cvtpd_epi32(scalex);
1084
1085     xpoint   = _mm_round_pd(scalex, _MM_FROUND_TO_NEAREST_INT);
1086
1087     /* Extended precision arithmetics */
1088     z        = _mm_sub_pd(xabs, _mm_mul_pd(invtabscale0, xpoint));
1089     z        = _mm_sub_pd(z, _mm_mul_pd(invtabscale1, xpoint));
1090
1091     /* Range reduction to 0..2*Pi */
1092     tabidx   = _mm_and_si128(tabidx, tabmask);
1093
1094     /* tabidx is now in range [0,..,64] */
1095     imask     = _mm_cmpgt_epi32(tabidx, i32);
1096     sswapsign = imask;
1097     cswapsign = imask;
1098     corridx   = _mm_and_si128(imask, i32);
1099     tabidx    = _mm_sub_epi32(tabidx, corridx);
1100
1101     /* tabidx is now in range [0..32] */
1102     imask     = _mm_cmpgt_epi32(tabidx, i16);
1103     cswapsign = _mm_xor_si128(cswapsign, imask);
1104     corridx   = _mm_sub_epi32(i32, tabidx);
1105     tabidx    = _mm_blendv_epi8(tabidx, corridx, imask);
1106     /* tabidx is now in range [0..16] */
1107     ssign     = _mm_cvtepi32_pd( _mm_or_si128( sswapsign, ione ) );
1108     csign     = _mm_cvtepi32_pd( _mm_or_si128( cswapsign, ione ) );
1109
1110 #ifdef _MSC_VER
1111     ypoint0  = _mm_load_pd(sintable + 2*_mm_extract_epi32(tabidx, 0));
1112     ypoint1  = _mm_load_pd(sintable + 2*_mm_extract_epi32(tabidx, 1));
1113 #else
1114     ypoint0  = sintable[_mm_extract_epi32(tabidx, 0)];
1115     ypoint1  = sintable[_mm_extract_epi32(tabidx, 1)];
1116 #endif
1117     sinpoint = _mm_unpackhi_pd(ypoint0, ypoint1);
1118     cospoint = _mm_unpacklo_pd(ypoint0, ypoint1);
1119
1120     sinpoint = _mm_mul_pd(sinpoint, ssign);
1121     cospoint = _mm_mul_pd(cospoint, csign);
1122
1123     z2       = _mm_mul_pd(z, z);
1124
1125     polySin  = _mm_mul_pd(sinP7, z2);
1126     polySin  = _mm_add_pd(polySin, sinP5);
1127     polySin  = _mm_mul_pd(polySin, z2);
1128     polySin  = _mm_add_pd(polySin, sinP3);
1129     polySin  = _mm_mul_pd(polySin, z2);
1130     polySin  = _mm_add_pd(polySin, sinP1);
1131     polySin  = _mm_mul_pd(polySin, z);
1132
1133     polyCos  = _mm_mul_pd(cosP6, z2);
1134     polyCos  = _mm_add_pd(polyCos, cosP4);
1135     polyCos  = _mm_mul_pd(polyCos, z2);
1136     polyCos  = _mm_add_pd(polyCos, cosP2);
1137     polyCos  = _mm_mul_pd(polyCos, z2);
1138     polyCos  = _mm_add_pd(polyCos, cosP0);
1139
1140     *sinval  = _mm_xor_pd(_mm_add_pd( _mm_mul_pd(sinpoint, polyCos), _mm_mul_pd(cospoint, polySin) ), xsign);
1141     *cosval  = _mm_sub_pd( _mm_mul_pd(cospoint, polyCos), _mm_mul_pd(sinpoint, polySin) );
1142
1143     return 0;
1144 }
1145
1146 /*
1147  * IMPORTANT: Do NOT call both sin & cos if you need both results, since each of them
1148  * will then call the sincos() routine and waste a factor 2 in performance!
1149  */
1150 static __m128d
1151 gmx_mm_sin_pd(__m128d x)
1152 {
1153     __m128d s, c;
1154     gmx_mm_sincos_pd(x, &s, &c);
1155     return s;
1156 }
1157
1158 /*
1159  * IMPORTANT: Do NOT call both sin & cos if you need both results, since each of them
1160  * will then call the sincos() routine and waste a factor 2 in performance!
1161  */
1162 static __m128d
1163 gmx_mm_cos_pd(__m128d x)
1164 {
1165     __m128d s, c;
1166     gmx_mm_sincos_pd(x, &s, &c);
1167     return c;
1168 }
1169
1170
1171
1172 static __m128d
1173 gmx_mm_tan_pd(__m128d x)
1174 {
1175     __m128d sinval, cosval;
1176     __m128d tanval;
1177
1178     gmx_mm_sincos_pd(x, &sinval, &cosval);
1179
1180     tanval = _mm_mul_pd(sinval, gmx_mm_inv_pd(cosval));
1181
1182     return tanval;
1183 }
1184
1185
1186
1187 static __m128d
1188 gmx_mm_asin_pd(__m128d x)
1189 {
1190     /* Same algorithm as cephes library */
1191     const __m128d signmask  = gmx_mm_castsi128_pd( _mm_set_epi32(0x7FFFFFFF, 0xFFFFFFFF, 0x7FFFFFFF, 0xFFFFFFFF) );
1192     const __m128d limit1    = _mm_set1_pd(0.625);
1193     const __m128d limit2    = _mm_set1_pd(1e-8);
1194     const __m128d one       = _mm_set1_pd(1.0);
1195     const __m128d quarterpi = _mm_set1_pd(M_PI/4.0);
1196     const __m128d morebits  = _mm_set1_pd(6.123233995736765886130e-17);
1197
1198     const __m128d P5        = _mm_set1_pd(4.253011369004428248960e-3);
1199     const __m128d P4        = _mm_set1_pd(-6.019598008014123785661e-1);
1200     const __m128d P3        = _mm_set1_pd(5.444622390564711410273e0);
1201     const __m128d P2        = _mm_set1_pd(-1.626247967210700244449e1);
1202     const __m128d P1        = _mm_set1_pd(1.956261983317594739197e1);
1203     const __m128d P0        = _mm_set1_pd(-8.198089802484824371615e0);
1204
1205     const __m128d Q4        = _mm_set1_pd(-1.474091372988853791896e1);
1206     const __m128d Q3        = _mm_set1_pd(7.049610280856842141659e1);
1207     const __m128d Q2        = _mm_set1_pd(-1.471791292232726029859e2);
1208     const __m128d Q1        = _mm_set1_pd(1.395105614657485689735e2);
1209     const __m128d Q0        = _mm_set1_pd(-4.918853881490881290097e1);
1210
1211     const __m128d R4        = _mm_set1_pd(2.967721961301243206100e-3);
1212     const __m128d R3        = _mm_set1_pd(-5.634242780008963776856e-1);
1213     const __m128d R2        = _mm_set1_pd(6.968710824104713396794e0);
1214     const __m128d R1        = _mm_set1_pd(-2.556901049652824852289e1);
1215     const __m128d R0        = _mm_set1_pd(2.853665548261061424989e1);
1216
1217     const __m128d S3        = _mm_set1_pd(-2.194779531642920639778e1);
1218     const __m128d S2        = _mm_set1_pd(1.470656354026814941758e2);
1219     const __m128d S1        = _mm_set1_pd(-3.838770957603691357202e2);
1220     const __m128d S0        = _mm_set1_pd(3.424398657913078477438e2);
1221
1222     __m128d       sign;
1223     __m128d       mask;
1224     __m128d       xabs;
1225     __m128d       zz, ww, z, q, w, zz2, ww2;
1226     __m128d       PA, PB;
1227     __m128d       QA, QB;
1228     __m128d       RA, RB;
1229     __m128d       SA, SB;
1230     __m128d       nom, denom;
1231
1232     sign  = _mm_andnot_pd(signmask, x);
1233     xabs  = _mm_and_pd(x, signmask);
1234
1235     mask  = _mm_cmpgt_pd(xabs, limit1);
1236
1237     zz    = _mm_sub_pd(one, xabs);
1238     ww    = _mm_mul_pd(xabs, xabs);
1239     zz2   = _mm_mul_pd(zz, zz);
1240     ww2   = _mm_mul_pd(ww, ww);
1241
1242     /* R */
1243     RA    = _mm_mul_pd(R4, zz2);
1244     RB    = _mm_mul_pd(R3, zz2);
1245     RA    = _mm_add_pd(RA, R2);
1246     RB    = _mm_add_pd(RB, R1);
1247     RA    = _mm_mul_pd(RA, zz2);
1248     RB    = _mm_mul_pd(RB, zz);
1249     RA    = _mm_add_pd(RA, R0);
1250     RA    = _mm_add_pd(RA, RB);
1251
1252     /* S, SA = zz2 */
1253     SB    = _mm_mul_pd(S3, zz2);
1254     SA    = _mm_add_pd(zz2, S2);
1255     SB    = _mm_add_pd(SB, S1);
1256     SA    = _mm_mul_pd(SA, zz2);
1257     SB    = _mm_mul_pd(SB, zz);
1258     SA    = _mm_add_pd(SA, S0);
1259     SA    = _mm_add_pd(SA, SB);
1260
1261     /* P */
1262     PA    = _mm_mul_pd(P5, ww2);
1263     PB    = _mm_mul_pd(P4, ww2);
1264     PA    = _mm_add_pd(PA, P3);
1265     PB    = _mm_add_pd(PB, P2);
1266     PA    = _mm_mul_pd(PA, ww2);
1267     PB    = _mm_mul_pd(PB, ww2);
1268     PA    = _mm_add_pd(PA, P1);
1269     PB    = _mm_add_pd(PB, P0);
1270     PA    = _mm_mul_pd(PA, ww);
1271     PA    = _mm_add_pd(PA, PB);
1272
1273     /* Q, QA = ww2 */
1274     QB    = _mm_mul_pd(Q4, ww2);
1275     QA    = _mm_add_pd(ww2, Q3);
1276     QB    = _mm_add_pd(QB, Q2);
1277     QA    = _mm_mul_pd(QA, ww2);
1278     QB    = _mm_mul_pd(QB, ww2);
1279     QA    = _mm_add_pd(QA, Q1);
1280     QB    = _mm_add_pd(QB, Q0);
1281     QA    = _mm_mul_pd(QA, ww);
1282     QA    = _mm_add_pd(QA, QB);
1283
1284     RA    = _mm_mul_pd(RA, zz);
1285     PA    = _mm_mul_pd(PA, ww);
1286
1287     nom   = _mm_blendv_pd( PA, RA, mask );
1288     denom = _mm_blendv_pd( QA, SA, mask );
1289
1290     q     = _mm_mul_pd( nom, gmx_mm_inv_pd(denom) );
1291
1292     zz    = _mm_add_pd(zz, zz);
1293     zz    = gmx_mm_sqrt_pd(zz);
1294     z     = _mm_sub_pd(quarterpi, zz);
1295     zz    = _mm_mul_pd(zz, q);
1296     zz    = _mm_sub_pd(zz, morebits);
1297     z     = _mm_sub_pd(z, zz);
1298     z     = _mm_add_pd(z, quarterpi);
1299
1300     w     = _mm_mul_pd(xabs, q);
1301     w     = _mm_add_pd(w, xabs);
1302
1303     z     = _mm_blendv_pd( w, z, mask );
1304
1305     mask  = _mm_cmpgt_pd(xabs, limit2);
1306     z     = _mm_blendv_pd( xabs, z, mask );
1307
1308     z = _mm_xor_pd(z, sign);
1309
1310     return z;
1311 }
1312
1313
1314 static __m128d
1315 gmx_mm_acos_pd(__m128d x)
1316 {
1317     const __m128d one        = _mm_set1_pd(1.0);
1318     const __m128d half       = _mm_set1_pd(0.5);
1319     const __m128d quarterpi0 = _mm_set1_pd(7.85398163397448309616e-1);
1320     const __m128d quarterpi1 = _mm_set1_pd(6.123233995736765886130e-17);
1321
1322
1323     __m128d mask1;
1324
1325     __m128d z, z1, z2;
1326
1327     mask1 = _mm_cmpgt_pd(x, half);
1328     z1    = _mm_mul_pd(half, _mm_sub_pd(one, x));
1329     z1    = gmx_mm_sqrt_pd(z1);
1330     z     = _mm_blendv_pd( x, z1, mask1 );
1331
1332     z     = gmx_mm_asin_pd(z);
1333
1334     z1    = _mm_add_pd(z, z);
1335
1336     z2    = _mm_sub_pd(quarterpi0, z);
1337     z2    = _mm_add_pd(z2, quarterpi1);
1338     z2    = _mm_add_pd(z2, quarterpi0);
1339
1340     z     = _mm_blendv_pd(z2, z1, mask1);
1341
1342     return z;
1343 }
1344
1345 static __m128d
1346 gmx_mm_atan_pd(__m128d x)
1347 {
1348     /* Same algorithm as cephes library */
1349     const __m128d signmask  = gmx_mm_castsi128_pd( _mm_set_epi32(0x7FFFFFFF, 0xFFFFFFFF, 0x7FFFFFFF, 0xFFFFFFFF) );
1350     const __m128d limit1    = _mm_set1_pd(0.66);
1351     const __m128d limit2    = _mm_set1_pd(2.41421356237309504880);
1352     const __m128d quarterpi = _mm_set1_pd(M_PI/4.0);
1353     const __m128d halfpi    = _mm_set1_pd(M_PI/2.0);
1354     const __m128d mone      = _mm_set1_pd(-1.0);
1355     const __m128d morebits1 = _mm_set1_pd(0.5*6.123233995736765886130E-17);
1356     const __m128d morebits2 = _mm_set1_pd(6.123233995736765886130E-17);
1357
1358     const __m128d P4        = _mm_set1_pd(-8.750608600031904122785E-1);
1359     const __m128d P3        = _mm_set1_pd(-1.615753718733365076637E1);
1360     const __m128d P2        = _mm_set1_pd(-7.500855792314704667340E1);
1361     const __m128d P1        = _mm_set1_pd(-1.228866684490136173410E2);
1362     const __m128d P0        = _mm_set1_pd(-6.485021904942025371773E1);
1363
1364     const __m128d Q4        = _mm_set1_pd(2.485846490142306297962E1);
1365     const __m128d Q3        = _mm_set1_pd(1.650270098316988542046E2);
1366     const __m128d Q2        = _mm_set1_pd(4.328810604912902668951E2);
1367     const __m128d Q1        = _mm_set1_pd(4.853903996359136964868E2);
1368     const __m128d Q0        = _mm_set1_pd(1.945506571482613964425E2);
1369
1370     __m128d       sign;
1371     __m128d       mask1, mask2;
1372     __m128d       y, t1, t2;
1373     __m128d       z, z2;
1374     __m128d       P_A, P_B, Q_A, Q_B;
1375
1376     sign   = _mm_andnot_pd(signmask, x);
1377     x      = _mm_and_pd(x, signmask);
1378
1379     mask1  = _mm_cmpgt_pd(x, limit1);
1380     mask2  = _mm_cmpgt_pd(x, limit2);
1381
1382     t1     = _mm_mul_pd(_mm_add_pd(x, mone), gmx_mm_inv_pd(_mm_sub_pd(x, mone)));
1383     t2     = _mm_mul_pd(mone, gmx_mm_inv_pd(x));
1384
1385     y      = _mm_and_pd(mask1, quarterpi);
1386     y      = _mm_or_pd( _mm_and_pd(mask2, halfpi), _mm_andnot_pd(mask2, y) );
1387
1388     x      = _mm_or_pd( _mm_and_pd(mask1, t1), _mm_andnot_pd(mask1, x) );
1389     x      = _mm_or_pd( _mm_and_pd(mask2, t2), _mm_andnot_pd(mask2, x) );
1390
1391     z      = _mm_mul_pd(x, x);
1392     z2     = _mm_mul_pd(z, z);
1393
1394     P_A    = _mm_mul_pd(P4, z2);
1395     P_B    = _mm_mul_pd(P3, z2);
1396     P_A    = _mm_add_pd(P_A, P2);
1397     P_B    = _mm_add_pd(P_B, P1);
1398     P_A    = _mm_mul_pd(P_A, z2);
1399     P_B    = _mm_mul_pd(P_B, z);
1400     P_A    = _mm_add_pd(P_A, P0);
1401     P_A    = _mm_add_pd(P_A, P_B);
1402
1403     /* Q_A = z2 */
1404     Q_B    = _mm_mul_pd(Q4, z2);
1405     Q_A    = _mm_add_pd(z2, Q3);
1406     Q_B    = _mm_add_pd(Q_B, Q2);
1407     Q_A    = _mm_mul_pd(Q_A, z2);
1408     Q_B    = _mm_mul_pd(Q_B, z2);
1409     Q_A    = _mm_add_pd(Q_A, Q1);
1410     Q_B    = _mm_add_pd(Q_B, Q0);
1411     Q_A    = _mm_mul_pd(Q_A, z);
1412     Q_A    = _mm_add_pd(Q_A, Q_B);
1413
1414     z      = _mm_mul_pd(z, P_A);
1415     z      = _mm_mul_pd(z, gmx_mm_inv_pd(Q_A));
1416     z      = _mm_mul_pd(z, x);
1417     z      = _mm_add_pd(z, x);
1418
1419     t1     = _mm_and_pd(mask1, morebits1);
1420     t1     = _mm_or_pd( _mm_and_pd(mask2, morebits2), _mm_andnot_pd(mask2, t1) );
1421
1422     z      = _mm_add_pd(z, t1);
1423     y      = _mm_add_pd(y, z);
1424
1425     y      = _mm_xor_pd(y, sign);
1426
1427     return y;
1428 }
1429
1430
1431 static __m128d
1432 gmx_mm_atan2_pd(__m128d y, __m128d x)
1433 {
1434     const __m128d pi          = _mm_set1_pd(M_PI);
1435     const __m128d minuspi     = _mm_set1_pd(-M_PI);
1436     const __m128d halfpi      = _mm_set1_pd(M_PI/2.0);
1437     const __m128d minushalfpi = _mm_set1_pd(-M_PI/2.0);
1438
1439     __m128d       z, z1, z3, z4;
1440     __m128d       w;
1441     __m128d       maskx_lt, maskx_eq;
1442     __m128d       masky_lt, masky_eq;
1443     __m128d       mask1, mask2, mask3, mask4, maskall;
1444
1445     maskx_lt  = _mm_cmplt_pd(x, _mm_setzero_pd());
1446     masky_lt  = _mm_cmplt_pd(y, _mm_setzero_pd());
1447     maskx_eq  = _mm_cmpeq_pd(x, _mm_setzero_pd());
1448     masky_eq  = _mm_cmpeq_pd(y, _mm_setzero_pd());
1449
1450     z         = _mm_mul_pd(y, gmx_mm_inv_pd(x));
1451     z         = gmx_mm_atan_pd(z);
1452
1453     mask1     = _mm_and_pd(maskx_eq, masky_lt);
1454     mask2     = _mm_andnot_pd(maskx_lt, masky_eq);
1455     mask3     = _mm_andnot_pd( _mm_or_pd(masky_lt, masky_eq), maskx_eq);
1456     mask4     = _mm_and_pd(masky_eq, maskx_lt);
1457
1458     maskall   = _mm_or_pd( _mm_or_pd(mask1, mask2), _mm_or_pd(mask3, mask4) );
1459
1460     z         = _mm_andnot_pd(maskall, z);
1461     z1        = _mm_and_pd(mask1, minushalfpi);
1462     z3        = _mm_and_pd(mask3, halfpi);
1463     z4        = _mm_and_pd(mask4, pi);
1464
1465     z         = _mm_or_pd( _mm_or_pd(z, z1), _mm_or_pd(z3, z4) );
1466
1467     w         = _mm_blendv_pd(pi, minuspi, masky_lt);
1468     w         = _mm_and_pd(w, maskx_lt);
1469
1470     w         = _mm_andnot_pd(maskall, w);
1471
1472     z         = _mm_add_pd(z, w);
1473
1474     return z;
1475 }
1476
1477 #endif /*_gmx_math_x86_sse4_1_double_h_ */