src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/kernelutil_x86_avx_128_fma_double.h

   1 /*
   2  *                This source code is part of
   3  *
   4  *                 G   R   O   M   A   C   S
   5  *
   6  * Copyright (c) 2011-2012, The GROMACS Development Team
   7  *
   8  * Gromacs is a library for molecular simulation and trajectory analysis,
   9  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  10  * a full list of developers and information, check out http://www.gromacs.org
  11  *
  12  * This program is free software; you can redistribute it and/or modify it under
  13  * the terms of the GNU Lesser General Public License as published by the Free
  14  * Software Foundation; either version 2 of the License, or (at your option) any
  15  * later version.
  16  * As a special exception, you may use this file as part of a free software
  17  * library without restriction.  Specifically, if other files instantiate
  18  * templates or use macros or inline functions from this file, or you compile
  19  * this file and link it with other files to produce an executable, this
  20  * file does not by itself cause the resulting executable to be covered by
  21  * the GNU Lesser General Public License.
  22  *
  23  * In plain-speak: do not worry about classes/macros/templates either - only
  24  * changes to the library have to be LGPL, not an application linking with it.
  25  *
  26  * To help fund GROMACS development, we humbly ask that you cite
  27  * the papers people have written on it - you can find them on the website!
  28  */
  29 #ifndef _kernelutil_x86_avx_128_fma_double_h_
  30 #define _kernelutil_x86_avx_128_fma_double_h_
  31
  32 #include "gmx_x86_avx_128_fma.h"
  33
  34
  35 static int
  36 gmx_mm_any_lt(__m128d a, __m128d b)
  37 {
  38     return _mm_movemask_pd(_mm_cmplt_pd(a, b));
  39 }
  40
  41
  42 static gmx_inline __m128d
  43 gmx_mm_calc_rsq_pd(__m128d dx, __m128d dy, __m128d dz)
  44 {
  45     return _mm_macc_pd(dx, dx, _mm_macc_pd(dy, dy, _mm_mul_pd(dz, dz)));
  46 }
  47
  48 /* Normal sum of four ymm registers */
  49 #define gmx_mm_sum4_pd(t0, t1, t2, t3)  _mm_add_pd(_mm_add_pd(t0, t1), _mm_add_pd(t2, t3))
  50
  51
  52
  53 /* Load a double value from 1-2 places, merge into xmm register */
  54
  55
  56 static __m128d
  57 gmx_mm_load_2real_swizzle_pd(const double * gmx_restrict ptrA,
  58                              const double * gmx_restrict ptrB)
  59 {
  60     return _mm_unpacklo_pd(_mm_load_sd(ptrA), _mm_load_sd(ptrB));
  61 }
  62
  63 static __m128d
  64 gmx_mm_load_1real_pd(const double * gmx_restrict ptrA)
  65 {
  66     return _mm_load_sd(ptrA);
  67 }
  68
  69
  70 static void
  71 gmx_mm_store_2real_swizzle_pd(double * gmx_restrict ptrA,
  72                               double * gmx_restrict ptrB,
  73                               __m128d               xmm1)
  74 {
  75     __m128d t2;
  76
  77     t2       = _mm_unpackhi_pd(xmm1, xmm1);
  78     _mm_store_sd(ptrA, xmm1);
  79     _mm_store_sd(ptrB, t2);
  80 }
  81
  82 static void
  83 gmx_mm_store_1real_pd(double * gmx_restrict ptrA, __m128d xmm1)
  84 {
  85     _mm_store_sd(ptrA, xmm1);
  86 }
  87
  88
  89 /* Similar to store, but increments value in memory */
  90 static void
  91 gmx_mm_increment_2real_swizzle_pd(double * gmx_restrict ptrA,
  92                                   double * gmx_restrict ptrB, __m128d xmm1)
  93 {
  94     __m128d t1;
  95
  96     t1   = _mm_unpackhi_pd(xmm1, xmm1);
  97     xmm1 = _mm_add_sd(xmm1, _mm_load_sd(ptrA));
  98     t1   = _mm_add_sd(t1, _mm_load_sd(ptrB));
  99     _mm_store_sd(ptrA, xmm1);
 100     _mm_store_sd(ptrB, t1);
 101 }
 102
 103 static void
 104 gmx_mm_increment_1real_pd(double * gmx_restrict ptrA, __m128d xmm1)
 105 {
 106     __m128d tmp;
 107
 108     tmp = gmx_mm_load_1real_pd(ptrA);
 109     tmp = _mm_add_sd(tmp, xmm1);
 110     gmx_mm_store_1real_pd(ptrA, tmp);
 111 }
 112
 113
 114
 115 static gmx_inline void
 116 gmx_mm_load_2pair_swizzle_pd(const double * gmx_restrict p1,
 117                              const double * gmx_restrict p2,
 118                              __m128d * gmx_restrict      c6,
 119                              __m128d * gmx_restrict      c12)
 120 {
 121     __m128d t1, t2, t3;
 122
 123     /* The c6/c12 array should be aligned */
 124     t1   = _mm_loadu_pd(p1);
 125     t2   = _mm_loadu_pd(p2);
 126     *c6  = _mm_unpacklo_pd(t1, t2);
 127     *c12 = _mm_unpackhi_pd(t1, t2);
 128 }
 129
 130 static gmx_inline void
 131 gmx_mm_load_1pair_swizzle_pd(const double * gmx_restrict p1,
 132                              __m128d * gmx_restrict      c6,
 133                              __m128d * gmx_restrict      c12)
 134 {
 135     *c6     = _mm_load_sd(p1);
 136     *c12    = _mm_load_sd(p1+1);
 137 }
 138
 139
 140 static gmx_inline void
 141 gmx_mm_load_shift_and_1rvec_broadcast_pd(const double * gmx_restrict xyz_shift,
 142                                          const double * gmx_restrict xyz,
 143                                          __m128d * gmx_restrict      x1,
 144                                          __m128d * gmx_restrict      y1,
 145                                          __m128d * gmx_restrict      z1)
 146 {
 147     __m128d mem_xy, mem_z, mem_sxy, mem_sz;
 148
 149     mem_xy  = _mm_loadu_pd(xyz);
 150     mem_z   = _mm_load_sd(xyz+2);
 151     mem_sxy = _mm_loadu_pd(xyz_shift);
 152     mem_sz  = _mm_load_sd(xyz_shift+2);
 153
 154     mem_xy  = _mm_add_pd(mem_xy, mem_sxy);
 155     mem_z   = _mm_add_pd(mem_z, mem_sz);
 156
 157     *x1  = _mm_shuffle_pd(mem_xy, mem_xy, _MM_SHUFFLE2(0, 0));
 158     *y1  = _mm_shuffle_pd(mem_xy, mem_xy, _MM_SHUFFLE2(1, 1));
 159     *z1  = _mm_shuffle_pd(mem_z, mem_z, _MM_SHUFFLE2(0, 0));
 160 }
 161
 162
 163 static gmx_inline void
 164 gmx_mm_load_shift_and_3rvec_broadcast_pd(const double * gmx_restrict xyz_shift,
 165                                          const double * gmx_restrict xyz,
 166                                          __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 167                                          __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 168                                          __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3)
 169 {
 170     __m128d t1, t2, t3, t4, t5, sxy, sz, szx, syz;
 171
 172     t1  = _mm_loadu_pd(xyz);
 173     t2  = _mm_loadu_pd(xyz+2);
 174     t3  = _mm_loadu_pd(xyz+4);
 175     t4  = _mm_loadu_pd(xyz+6);
 176     t5  = _mm_load_sd(xyz+8);
 177
 178     sxy = _mm_loadu_pd(xyz_shift);
 179     sz  = _mm_load_sd(xyz_shift+2);
 180     szx = _mm_shuffle_pd(sz, sxy, _MM_SHUFFLE2(0, 0));
 181     syz = _mm_shuffle_pd(sxy, sz, _MM_SHUFFLE2(0, 1));
 182
 183     t1  = _mm_add_pd(t1, sxy);
 184     t2  = _mm_add_pd(t2, szx);
 185     t3  = _mm_add_pd(t3, syz);
 186     t4  = _mm_add_pd(t4, sxy);
 187     t5  = _mm_add_sd(t5, sz);
 188
 189     *x1  = _mm_shuffle_pd(t1, t1, _MM_SHUFFLE2(0, 0));
 190     *y1  = _mm_shuffle_pd(t1, t1, _MM_SHUFFLE2(1, 1));
 191     *z1  = _mm_shuffle_pd(t2, t2, _MM_SHUFFLE2(0, 0));
 192     *x2  = _mm_shuffle_pd(t2, t2, _MM_SHUFFLE2(1, 1));
 193     *y2  = _mm_shuffle_pd(t3, t3, _MM_SHUFFLE2(0, 0));
 194     *z2  = _mm_shuffle_pd(t3, t3, _MM_SHUFFLE2(1, 1));
 195     *x3  = _mm_shuffle_pd(t4, t4, _MM_SHUFFLE2(0, 0));
 196     *y3  = _mm_shuffle_pd(t4, t4, _MM_SHUFFLE2(1, 1));
 197     *z3  = _mm_shuffle_pd(t5, t5, _MM_SHUFFLE2(0, 0));
 198 }
 199
 200
 201 static gmx_inline void
 202 gmx_mm_load_shift_and_4rvec_broadcast_pd(const double * gmx_restrict xyz_shift,
 203                                          const double * gmx_restrict xyz,
 204                                          __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 205                                          __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 206                                          __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3,
 207                                          __m128d * gmx_restrict x4, __m128d * gmx_restrict y4, __m128d * gmx_restrict z4)
 208 {
 209     __m128d t1, t2, t3, t4, t5, t6, sxy, sz, szx, syz;
 210
 211     t1  = _mm_loadu_pd(xyz);
 212     t2  = _mm_loadu_pd(xyz+2);
 213     t3  = _mm_loadu_pd(xyz+4);
 214     t4  = _mm_loadu_pd(xyz+6);
 215     t5  = _mm_loadu_pd(xyz+8);
 216     t6  = _mm_loadu_pd(xyz+10);
 217
 218     sxy = _mm_loadu_pd(xyz_shift);
 219     sz  = _mm_load_sd(xyz_shift+2);
 220     szx = _mm_shuffle_pd(sz, sxy, _MM_SHUFFLE2(0, 0));
 221     syz = _mm_shuffle_pd(sxy, sz, _MM_SHUFFLE2(0, 1));
 222
 223     t1  = _mm_add_pd(t1, sxy);
 224     t2  = _mm_add_pd(t2, szx);
 225     t3  = _mm_add_pd(t3, syz);
 226     t4  = _mm_add_pd(t4, sxy);
 227     t5  = _mm_add_pd(t5, szx);
 228     t6  = _mm_add_pd(t6, syz);
 229
 230     *x1  = _mm_shuffle_pd(t1, t1, _MM_SHUFFLE2(0, 0));
 231     *y1  = _mm_shuffle_pd(t1, t1, _MM_SHUFFLE2(1, 1));
 232     *z1  = _mm_shuffle_pd(t2, t2, _MM_SHUFFLE2(0, 0));
 233     *x2  = _mm_shuffle_pd(t2, t2, _MM_SHUFFLE2(1, 1));
 234     *y2  = _mm_shuffle_pd(t3, t3, _MM_SHUFFLE2(0, 0));
 235     *z2  = _mm_shuffle_pd(t3, t3, _MM_SHUFFLE2(1, 1));
 236     *x3  = _mm_shuffle_pd(t4, t4, _MM_SHUFFLE2(0, 0));
 237     *y3  = _mm_shuffle_pd(t4, t4, _MM_SHUFFLE2(1, 1));
 238     *z3  = _mm_shuffle_pd(t5, t5, _MM_SHUFFLE2(0, 0));
 239     *x4  = _mm_shuffle_pd(t5, t5, _MM_SHUFFLE2(1, 1));
 240     *y4  = _mm_shuffle_pd(t6, t6, _MM_SHUFFLE2(0, 0));
 241     *z4  = _mm_shuffle_pd(t6, t6, _MM_SHUFFLE2(1, 1));
 242 }
 243
 244
 245
 246 static gmx_inline void
 247 gmx_mm_load_1rvec_1ptr_swizzle_pd(const double * gmx_restrict p1,
 248                                   __m128d * gmx_restrict x, __m128d * gmx_restrict y, __m128d * gmx_restrict z)
 249 {
 250     *x            = _mm_load_sd(p1);
 251     *y            = _mm_load_sd(p1+1);
 252     *z            = _mm_load_sd(p1+2);
 253 }
 254
 255 static gmx_inline void
 256 gmx_mm_load_3rvec_1ptr_swizzle_pd(const double * gmx_restrict p1,
 257                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 258                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 259                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3)
 260 {
 261     *x1            = _mm_load_sd(p1);
 262     *y1            = _mm_load_sd(p1+1);
 263     *z1            = _mm_load_sd(p1+2);
 264     *x2            = _mm_load_sd(p1+3);
 265     *y2            = _mm_load_sd(p1+4);
 266     *z2            = _mm_load_sd(p1+5);
 267     *x3            = _mm_load_sd(p1+6);
 268     *y3            = _mm_load_sd(p1+7);
 269     *z3            = _mm_load_sd(p1+8);
 270 }
 271
 272 static gmx_inline void
 273 gmx_mm_load_4rvec_1ptr_swizzle_pd(const double * gmx_restrict p1,
 274                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 275                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 276                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3,
 277                                   __m128d * gmx_restrict x4, __m128d * gmx_restrict y4, __m128d * gmx_restrict z4)
 278 {
 279     *x1            = _mm_load_sd(p1);
 280     *y1            = _mm_load_sd(p1+1);
 281     *z1            = _mm_load_sd(p1+2);
 282     *x2            = _mm_load_sd(p1+3);
 283     *y2            = _mm_load_sd(p1+4);
 284     *z2            = _mm_load_sd(p1+5);
 285     *x3            = _mm_load_sd(p1+6);
 286     *y3            = _mm_load_sd(p1+7);
 287     *z3            = _mm_load_sd(p1+8);
 288     *x4            = _mm_load_sd(p1+9);
 289     *y4            = _mm_load_sd(p1+10);
 290     *z4            = _mm_load_sd(p1+11);
 291 }
 292
 293
 294 static gmx_inline void
 295 gmx_mm_load_1rvec_2ptr_swizzle_pd(const double * gmx_restrict ptrA,
 296                                   const double * gmx_restrict ptrB,
 297                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1)
 298 {
 299     __m128d t1, t2, t3, t4;
 300     t1           = _mm_loadu_pd(ptrA);
 301     t2           = _mm_loadu_pd(ptrB);
 302     t3           = _mm_load_sd(ptrA+2);
 303     t4           = _mm_load_sd(ptrB+2);
 304     GMX_MM_TRANSPOSE2_PD(t1, t2);
 305     *x1          = t1;
 306     *y1          = t2;
 307     *z1          = _mm_unpacklo_pd(t3, t4);
 308 }
 309
 310 static gmx_inline void
 311 gmx_mm_load_3rvec_2ptr_swizzle_pd(const double * gmx_restrict ptrA, const double * gmx_restrict ptrB,
 312                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 313                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 314                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3)
 315 {
 316     __m128d t1, t2, t3, t4, t5, t6, t7, t8, t9, t10;
 317     t1           = _mm_loadu_pd(ptrA);
 318     t2           = _mm_loadu_pd(ptrB);
 319     t3           = _mm_loadu_pd(ptrA+2);
 320     t4           = _mm_loadu_pd(ptrB+2);
 321     t5           = _mm_loadu_pd(ptrA+4);
 322     t6           = _mm_loadu_pd(ptrB+4);
 323     t7           = _mm_loadu_pd(ptrA+6);
 324     t8           = _mm_loadu_pd(ptrB+6);
 325     t9           = _mm_load_sd(ptrA+8);
 326     t10          = _mm_load_sd(ptrB+8);
 327     GMX_MM_TRANSPOSE2_PD(t1, t2);
 328     GMX_MM_TRANSPOSE2_PD(t3, t4);
 329     GMX_MM_TRANSPOSE2_PD(t5, t6);
 330     GMX_MM_TRANSPOSE2_PD(t7, t8);
 331     *x1          = t1;
 332     *y1          = t2;
 333     *z1          = t3;
 334     *x2          = t4;
 335     *y2          = t5;
 336     *z2          = t6;
 337     *x3          = t7;
 338     *y3          = t8;
 339     *z3          = _mm_unpacklo_pd(t9, t10);
 340 }
 341
 342
 343 static gmx_inline void
 344 gmx_mm_load_4rvec_2ptr_swizzle_pd(const double * gmx_restrict ptrA, const double * gmx_restrict ptrB,
 345                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 346                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 347                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3,
 348                                   __m128d * gmx_restrict x4, __m128d * gmx_restrict y4, __m128d * gmx_restrict z4)
 349 {
 350     __m128d t1, t2, t3, t4, t5, t6;
 351     t1           = _mm_loadu_pd(ptrA);
 352     t2           = _mm_loadu_pd(ptrB);
 353     t3           = _mm_loadu_pd(ptrA+2);
 354     t4           = _mm_loadu_pd(ptrB+2);
 355     t5           = _mm_loadu_pd(ptrA+4);
 356     t6           = _mm_loadu_pd(ptrB+4);
 357     GMX_MM_TRANSPOSE2_PD(t1, t2);
 358     GMX_MM_TRANSPOSE2_PD(t3, t4);
 359     GMX_MM_TRANSPOSE2_PD(t5, t6);
 360     *x1          = t1;
 361     *y1          = t2;
 362     *z1          = t3;
 363     *x2          = t4;
 364     *y2          = t5;
 365     *z2          = t6;
 366     t1           = _mm_loadu_pd(ptrA+6);
 367     t2           = _mm_loadu_pd(ptrB+6);
 368     t3           = _mm_loadu_pd(ptrA+8);
 369     t4           = _mm_loadu_pd(ptrB+8);
 370     t5           = _mm_loadu_pd(ptrA+10);
 371     t6           = _mm_loadu_pd(ptrB+10);
 372     GMX_MM_TRANSPOSE2_PD(t1, t2);
 373     GMX_MM_TRANSPOSE2_PD(t3, t4);
 374     GMX_MM_TRANSPOSE2_PD(t5, t6);
 375     *x3          = t1;
 376     *y3          = t2;
 377     *z3          = t3;
 378     *x4          = t4;
 379     *y4          = t5;
 380     *z4          = t6;
 381 }
 382
 383
 384 /* Routines to decrement rvec in memory, typically use for j particle force updates */
 385 static void
 386 gmx_mm_decrement_1rvec_1ptr_swizzle_pd(double * gmx_restrict ptrA,
 387                                        __m128d x1, __m128d y1, __m128d z1)
 388 {
 389     __m128d t1, t2, t3;
 390
 391     t1           = _mm_load_sd(ptrA);
 392     t2           = _mm_load_sd(ptrA+1);
 393     t3           = _mm_load_sd(ptrA+2);
 394
 395     t1           = _mm_sub_sd(t1, x1);
 396     t2           = _mm_sub_sd(t2, y1);
 397     t3           = _mm_sub_sd(t3, z1);
 398     _mm_store_sd(ptrA, t1);
 399     _mm_store_sd(ptrA+1, t2);
 400     _mm_store_sd(ptrA+2, t3);
 401 }
 402
 403
 404 #if defined (_MSC_VER) && defined(_M_IX86)
 405 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 406 #define gmx_mm_decrement_3rvec_1ptr_swizzle_pd(ptrA, _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3) \
 407     { \
 408         __m128d _t1, _t2, _t3, _t4, _t5; \
 409         _t1          = _mm_loadu_pd(ptrA); \
 410         _t2          = _mm_loadu_pd(ptrA+2); \
 411         _t3          = _mm_loadu_pd(ptrA+4); \
 412         _t4          = _mm_loadu_pd(ptrA+6); \
 413         _t5          = _mm_load_sd(ptrA+8); \
 414         _x1          = _mm_unpacklo_pd(_x1, _y1); \
 415         _z1          = _mm_unpacklo_pd(_z1, _x2); \
 416         _y2          = _mm_unpacklo_pd(_y2, _z2); \
 417         _x3          = _mm_unpacklo_pd(_x3, _y3); \
 418         _t1          = _mm_sub_pd(_t1, _x1); \
 419         _t2          = _mm_sub_pd(_t2, _z1); \
 420         _t3          = _mm_sub_pd(_t3, _y2); \
 421         _t4          = _mm_sub_pd(_t4, _x3); \
 422         _t5          = _mm_sub_sd(_t5, _z3); \
 423         _mm_storeu_pd(ptrA, _t1); \
 424         _mm_storeu_pd(ptrA+2, _t2); \
 425         _mm_storeu_pd(ptrA+4, _t3); \
 426         _mm_storeu_pd(ptrA+6, _t4); \
 427         _mm_store_sd(ptrA+8, _t5); \
 428     }
 429 #else
 430 /* Real function for sane compilers */
 431 static void
 432 gmx_mm_decrement_3rvec_1ptr_swizzle_pd(double * gmx_restrict ptrA,
 433                                        __m128d x1, __m128d y1, __m128d z1,
 434                                        __m128d x2, __m128d y2, __m128d z2,
 435                                        __m128d x3, __m128d y3, __m128d z3)
 436 {
 437     __m128d t1, t2, t3, t4, t5;
 438
 439     t1          = _mm_loadu_pd(ptrA);
 440     t2          = _mm_loadu_pd(ptrA+2);
 441     t3          = _mm_loadu_pd(ptrA+4);
 442     t4          = _mm_loadu_pd(ptrA+6);
 443     t5          = _mm_load_sd(ptrA+8);
 444
 445     x1          = _mm_unpacklo_pd(x1, y1);
 446     z1          = _mm_unpacklo_pd(z1, x2);
 447     y2          = _mm_unpacklo_pd(y2, z2);
 448     x3          = _mm_unpacklo_pd(x3, y3);
 449     /* nothing to be done for z3 */
 450
 451     t1          = _mm_sub_pd(t1, x1);
 452     t2          = _mm_sub_pd(t2, z1);
 453     t3          = _mm_sub_pd(t3, y2);
 454     t4          = _mm_sub_pd(t4, x3);
 455     t5          = _mm_sub_sd(t5, z3);
 456     _mm_storeu_pd(ptrA, t1);
 457     _mm_storeu_pd(ptrA+2, t2);
 458     _mm_storeu_pd(ptrA+4, t3);
 459     _mm_storeu_pd(ptrA+6, t4);
 460     _mm_store_sd(ptrA+8, t5);
 461 }
 462 #endif
 463
 464
 465 #if defined (_MSC_VER) && defined(_M_IX86)
 466 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 467 #define gmx_mm_decrement_4rvec_1ptr_swizzle_pd(ptrA, _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3, _x4, _y4, _z4) \
 468     { \
 469         __m128d _t1, _t2, _t3, _t4, _t5, _t6; \
 470         _t1          = _mm_loadu_pd(ptrA); \
 471         _t2          = _mm_loadu_pd(ptrA+2); \
 472         _t3          = _mm_loadu_pd(ptrA+4); \
 473         _t4          = _mm_loadu_pd(ptrA+6); \
 474         _t5          = _mm_loadu_pd(ptrA+8); \
 475         _t6          = _mm_loadu_pd(ptrA+10); \
 476         _x1          = _mm_unpacklo_pd(_x1, _y1); \
 477         _z1          = _mm_unpacklo_pd(_z1, _x2); \
 478         _y2          = _mm_unpacklo_pd(_y2, _z2); \
 479         _x3          = _mm_unpacklo_pd(_x3, _y3); \
 480         _z3          = _mm_unpacklo_pd(_z3, _x4); \
 481         _y4          = _mm_unpacklo_pd(_y4, _z4); \
 482         _mm_storeu_pd(ptrA,    _mm_sub_pd( _t1, _x1 )); \
 483         _mm_storeu_pd(ptrA+2,  _mm_sub_pd( _t2, _z1 )); \
 484         _mm_storeu_pd(ptrA+4,  _mm_sub_pd( _t3, _y2 )); \
 485         _mm_storeu_pd(ptrA+6,  _mm_sub_pd( _t4, _x3 )); \
 486         _mm_storeu_pd(ptrA+8,  _mm_sub_pd( _t5, _z3 )); \
 487         _mm_storeu_pd(ptrA+10, _mm_sub_pd( _t6, _y4 )); \
 488     }
 489 #else
 490 /* Real function for sane compilers */
 491 static void
 492 gmx_mm_decrement_4rvec_1ptr_swizzle_pd(double * gmx_restrict ptrA,
 493                                        __m128d x1, __m128d y1, __m128d z1,
 494                                        __m128d x2, __m128d y2, __m128d z2,
 495                                        __m128d x3, __m128d y3, __m128d z3,
 496                                        __m128d x4, __m128d y4, __m128d z4)
 497 {
 498     __m128d t1, t2, t3, t4, t5, t6;
 499
 500     t1          = _mm_loadu_pd(ptrA);
 501     t2          = _mm_loadu_pd(ptrA+2);
 502     t3          = _mm_loadu_pd(ptrA+4);
 503     t4          = _mm_loadu_pd(ptrA+6);
 504     t5          = _mm_loadu_pd(ptrA+8);
 505     t6          = _mm_loadu_pd(ptrA+10);
 506
 507     x1          = _mm_unpacklo_pd(x1, y1);
 508     z1          = _mm_unpacklo_pd(z1, x2);
 509     y2          = _mm_unpacklo_pd(y2, z2);
 510     x3          = _mm_unpacklo_pd(x3, y3);
 511     z3          = _mm_unpacklo_pd(z3, x4);
 512     y4          = _mm_unpacklo_pd(y4, z4);
 513
 514     _mm_storeu_pd(ptrA,    _mm_sub_pd( t1, x1 ));
 515     _mm_storeu_pd(ptrA+2,  _mm_sub_pd( t2, z1 ));
 516     _mm_storeu_pd(ptrA+4,  _mm_sub_pd( t3, y2 ));
 517     _mm_storeu_pd(ptrA+6,  _mm_sub_pd( t4, x3 ));
 518     _mm_storeu_pd(ptrA+8,  _mm_sub_pd( t5, z3 ));
 519     _mm_storeu_pd(ptrA+10, _mm_sub_pd( t6, y4 ));
 520 }
 521 #endif
 522
 523
 524 static void
 525 gmx_mm_decrement_1rvec_2ptr_swizzle_pd(double * gmx_restrict ptrA, double * gmx_restrict ptrB,
 526                                        __m128d x1, __m128d y1, __m128d z1)
 527 {
 528     __m128d t1, t2, t3, t4, t5, t6, t7;
 529
 530     t1          = _mm_loadu_pd(ptrA);
 531     t2          = _mm_load_sd(ptrA+2);
 532     t3          = _mm_loadu_pd(ptrB);
 533     t4          = _mm_load_sd(ptrB+2);
 534
 535     t5          = _mm_unpacklo_pd(x1, y1);
 536     t6          = _mm_unpackhi_pd(x1, y1);
 537     t7          = _mm_unpackhi_pd(z1, z1);
 538
 539     t1          = _mm_sub_pd(t1, t5);
 540     t2          = _mm_sub_sd(t2, z1);
 541
 542     t3          = _mm_sub_pd(t3, t6);
 543     t4          = _mm_sub_sd(t4, t7);
 544
 545     _mm_storeu_pd(ptrA, t1);
 546     _mm_store_sd(ptrA+2, t2);
 547     _mm_storeu_pd(ptrB, t3);
 548     _mm_store_sd(ptrB+2, t4);
 549 }
 550
 551
 552 #if defined (_MSC_VER) && defined(_M_IX86)
 553 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 554 #define gmx_mm_decrement_3rvec_2ptr_swizzle_pd(ptrA, ptrB, _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3) \
 555     { \
 556         __m128d _t1, _t2, _t3, _t4, _t5, _t6, _t7, _t8, _t9, _t10; \
 557         __m128d _tA, _tB, _tC, _tD, _tE, _tF, _tG, _tH, _tI; \
 558         _t1          = _mm_loadu_pd(ptrA); \
 559         _t2          = _mm_loadu_pd(ptrA+2); \
 560         _t3          = _mm_loadu_pd(ptrA+4); \
 561         _t4          = _mm_loadu_pd(ptrA+6); \
 562         _t5          = _mm_load_sd(ptrA+8); \
 563         _t6          = _mm_loadu_pd(ptrB); \
 564         _t7          = _mm_loadu_pd(ptrB+2); \
 565         _t8          = _mm_loadu_pd(ptrB+4); \
 566         _t9          = _mm_loadu_pd(ptrB+6); \
 567         _t10         = _mm_load_sd(ptrB+8); \
 568         _tA          = _mm_unpacklo_pd(_x1, _y1); \
 569         _tB          = _mm_unpackhi_pd(_x1, _y1); \
 570         _tC          = _mm_unpacklo_pd(_z1, _x2); \
 571         _tD          = _mm_unpackhi_pd(_z1, _x2); \
 572         _tE          = _mm_unpacklo_pd(_y2, _z2); \
 573         _tF          = _mm_unpackhi_pd(_y2, _z2); \
 574         _tG          = _mm_unpacklo_pd(_x3, _y3); \
 575         _tH          = _mm_unpackhi_pd(_x3, _y3); \
 576         _tI          = _mm_unpackhi_pd(_z3, _z3); \
 577         _t1          = _mm_sub_pd(_t1, _tA); \
 578         _t2          = _mm_sub_pd(_t2, _tC); \
 579         _t3          = _mm_sub_pd(_t3, _tE); \
 580         _t4          = _mm_sub_pd(_t4, _tG); \
 581         _t5          = _mm_sub_sd(_t5, _z3); \
 582         _t6          = _mm_sub_pd(_t6, _tB); \
 583         _t7          = _mm_sub_pd(_t7, _tD); \
 584         _t8          = _mm_sub_pd(_t8, _tF); \
 585         _t9          = _mm_sub_pd(_t9, _tH); \
 586         _t10         = _mm_sub_sd(_t10, _tI); \
 587         _mm_storeu_pd(ptrA, _t1); \
 588         _mm_storeu_pd(ptrA+2, _t2); \
 589         _mm_storeu_pd(ptrA+4, _t3); \
 590         _mm_storeu_pd(ptrA+6, _t4); \
 591         _mm_store_sd(ptrA+8, _t5); \
 592         _mm_storeu_pd(ptrB, _t6); \
 593         _mm_storeu_pd(ptrB+2, _t7); \
 594         _mm_storeu_pd(ptrB+4, _t8); \
 595         _mm_storeu_pd(ptrB+6, _t9); \
 596         _mm_store_sd(ptrB+8, _t10); \
 597     }
 598 #else
 599 /* Real function for sane compilers */
 600 static void
 601 gmx_mm_decrement_3rvec_2ptr_swizzle_pd(double * gmx_restrict ptrA, double * gmx_restrict ptrB,
 602                                        __m128d x1, __m128d y1, __m128d z1,
 603                                        __m128d x2, __m128d y2, __m128d z2,
 604                                        __m128d x3, __m128d y3, __m128d z3)
 605 {
 606     __m128d t1, t2, t3, t4, t5, t6, t7, t8, t9, t10;
 607     __m128d tA, tB, tC, tD, tE, tF, tG, tH, tI;
 608
 609     t1          = _mm_loadu_pd(ptrA);
 610     t2          = _mm_loadu_pd(ptrA+2);
 611     t3          = _mm_loadu_pd(ptrA+4);
 612     t4          = _mm_loadu_pd(ptrA+6);
 613     t5          = _mm_load_sd(ptrA+8);
 614     t6          = _mm_loadu_pd(ptrB);
 615     t7          = _mm_loadu_pd(ptrB+2);
 616     t8          = _mm_loadu_pd(ptrB+4);
 617     t9          = _mm_loadu_pd(ptrB+6);
 618     t10         = _mm_load_sd(ptrB+8);
 619
 620     tA          = _mm_unpacklo_pd(x1, y1);
 621     tB          = _mm_unpackhi_pd(x1, y1);
 622     tC          = _mm_unpacklo_pd(z1, x2);
 623     tD          = _mm_unpackhi_pd(z1, x2);
 624     tE          = _mm_unpacklo_pd(y2, z2);
 625     tF          = _mm_unpackhi_pd(y2, z2);
 626     tG          = _mm_unpacklo_pd(x3, y3);
 627     tH          = _mm_unpackhi_pd(x3, y3);
 628     tI          = _mm_unpackhi_pd(z3, z3);
 629
 630     t1          = _mm_sub_pd(t1, tA);
 631     t2          = _mm_sub_pd(t2, tC);
 632     t3          = _mm_sub_pd(t3, tE);
 633     t4          = _mm_sub_pd(t4, tG);
 634     t5          = _mm_sub_sd(t5, z3);
 635
 636     t6          = _mm_sub_pd(t6, tB);
 637     t7          = _mm_sub_pd(t7, tD);
 638     t8          = _mm_sub_pd(t8, tF);
 639     t9          = _mm_sub_pd(t9, tH);
 640     t10         = _mm_sub_sd(t10, tI);
 641
 642     _mm_storeu_pd(ptrA, t1);
 643     _mm_storeu_pd(ptrA+2, t2);
 644     _mm_storeu_pd(ptrA+4, t3);
 645     _mm_storeu_pd(ptrA+6, t4);
 646     _mm_store_sd(ptrA+8, t5);
 647     _mm_storeu_pd(ptrB, t6);
 648     _mm_storeu_pd(ptrB+2, t7);
 649     _mm_storeu_pd(ptrB+4, t8);
 650     _mm_storeu_pd(ptrB+6, t9);
 651     _mm_store_sd(ptrB+8, t10);
 652 }
 653 #endif
 654
 655
 656 #if defined (_MSC_VER) && defined(_M_IX86)
 657 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 658 #define gmx_mm_decrement_4rvec_2ptr_swizzle_pd(ptrA, ptrB, _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3, _x4, _y4, _z4) \
 659     { \
 660         __m128d _t1, _t2, _t3, _t4, _t5, _t6, _t7, _t8, _t9, _t10, _t11, _t12; \
 661         __m128d _tA, _tB, _tC, _tD, _tE, _tF, _tG, _tH, _tI, _tJ, _tK, _tL; \
 662         _t1          = _mm_loadu_pd(ptrA); \
 663         _t2          = _mm_loadu_pd(ptrA+2); \
 664         _t3          = _mm_loadu_pd(ptrA+4); \
 665         _t4          = _mm_loadu_pd(ptrA+6); \
 666         _t5          = _mm_loadu_pd(ptrA+8); \
 667         _t6          = _mm_loadu_pd(ptrA+10); \
 668         _t7          = _mm_loadu_pd(ptrB); \
 669         _t8          = _mm_loadu_pd(ptrB+2); \
 670         _t9          = _mm_loadu_pd(ptrB+4); \
 671         _t10         = _mm_loadu_pd(ptrB+6); \
 672         _t11         = _mm_loadu_pd(ptrB+8); \
 673         _t12         = _mm_loadu_pd(ptrB+10); \
 674         _tA          = _mm_unpacklo_pd(_x1, _y1); \
 675         _tB          = _mm_unpackhi_pd(_x1, _y1); \
 676         _tC          = _mm_unpacklo_pd(_z1, _x2); \
 677         _tD          = _mm_unpackhi_pd(_z1, _x2); \
 678         _tE          = _mm_unpacklo_pd(_y2, _z2); \
 679         _tF          = _mm_unpackhi_pd(_y2, _z2); \
 680         _tG          = _mm_unpacklo_pd(_x3, _y3); \
 681         _tH          = _mm_unpackhi_pd(_x3, _y3); \
 682         _tI          = _mm_unpacklo_pd(_z3, _x4); \
 683         _tJ          = _mm_unpackhi_pd(_z3, _x4); \
 684         _tK          = _mm_unpacklo_pd(_y4, _z4); \
 685         _tL          = _mm_unpackhi_pd(_y4, _z4); \
 686         _t1          = _mm_sub_pd(_t1, _tA); \
 687         _t2          = _mm_sub_pd(_t2, _tC); \
 688         _t3          = _mm_sub_pd(_t3, _tE); \
 689         _t4          = _mm_sub_pd(_t4, _tG); \
 690         _t5          = _mm_sub_pd(_t5, _tI); \
 691         _t6          = _mm_sub_pd(_t6, _tK); \
 692         _t7          = _mm_sub_pd(_t7, _tB); \
 693         _t8          = _mm_sub_pd(_t8, _tD); \
 694         _t9          = _mm_sub_pd(_t9, _tF); \
 695         _t10         = _mm_sub_pd(_t10, _tH); \
 696         _t11         = _mm_sub_pd(_t11, _tJ); \
 697         _t12         = _mm_sub_pd(_t12, _tL); \
 698         _mm_storeu_pd(ptrA,  _t1); \
 699         _mm_storeu_pd(ptrA+2, _t2); \
 700         _mm_storeu_pd(ptrA+4, _t3); \
 701         _mm_storeu_pd(ptrA+6, _t4); \
 702         _mm_storeu_pd(ptrA+8, _t5); \
 703         _mm_storeu_pd(ptrA+10, _t6); \
 704         _mm_storeu_pd(ptrB,  _t7); \
 705         _mm_storeu_pd(ptrB+2, _t8); \
 706         _mm_storeu_pd(ptrB+4, _t9); \
 707         _mm_storeu_pd(ptrB+6, _t10); \
 708         _mm_storeu_pd(ptrB+8, _t11); \
 709         _mm_storeu_pd(ptrB+10, _t12); \
 710     }
 711 #else
 712 /* Real function for sane compilers */
 713 static void
 714 gmx_mm_decrement_4rvec_2ptr_swizzle_pd(double * gmx_restrict ptrA, double * gmx_restrict ptrB,
 715                                        __m128d x1, __m128d y1, __m128d z1,
 716                                        __m128d x2, __m128d y2, __m128d z2,
 717                                        __m128d x3, __m128d y3, __m128d z3,
 718                                        __m128d x4, __m128d y4, __m128d z4)
 719 {
 720     __m128d t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12;
 721     __m128d tA, tB, tC, tD, tE, tF, tG, tH, tI, tJ, tK, tL;
 722
 723     t1          = _mm_loadu_pd(ptrA);
 724     t2          = _mm_loadu_pd(ptrA+2);
 725     t3          = _mm_loadu_pd(ptrA+4);
 726     t4          = _mm_loadu_pd(ptrA+6);
 727     t5          = _mm_loadu_pd(ptrA+8);
 728     t6          = _mm_loadu_pd(ptrA+10);
 729     t7          = _mm_loadu_pd(ptrB);
 730     t8          = _mm_loadu_pd(ptrB+2);
 731     t9          = _mm_loadu_pd(ptrB+4);
 732     t10         = _mm_loadu_pd(ptrB+6);
 733     t11         = _mm_loadu_pd(ptrB+8);
 734     t12         = _mm_loadu_pd(ptrB+10);
 735
 736     tA          = _mm_unpacklo_pd(x1, y1);
 737     tB          = _mm_unpackhi_pd(x1, y1);
 738     tC          = _mm_unpacklo_pd(z1, x2);
 739     tD          = _mm_unpackhi_pd(z1, x2);
 740     tE          = _mm_unpacklo_pd(y2, z2);
 741     tF          = _mm_unpackhi_pd(y2, z2);
 742     tG          = _mm_unpacklo_pd(x3, y3);
 743     tH          = _mm_unpackhi_pd(x3, y3);
 744     tI          = _mm_unpacklo_pd(z3, x4);
 745     tJ          = _mm_unpackhi_pd(z3, x4);
 746     tK          = _mm_unpacklo_pd(y4, z4);
 747     tL          = _mm_unpackhi_pd(y4, z4);
 748
 749     t1          = _mm_sub_pd(t1, tA);
 750     t2          = _mm_sub_pd(t2, tC);
 751     t3          = _mm_sub_pd(t3, tE);
 752     t4          = _mm_sub_pd(t4, tG);
 753     t5          = _mm_sub_pd(t5, tI);
 754     t6          = _mm_sub_pd(t6, tK);
 755
 756     t7          = _mm_sub_pd(t7, tB);
 757     t8          = _mm_sub_pd(t8, tD);
 758     t9          = _mm_sub_pd(t9, tF);
 759     t10         = _mm_sub_pd(t10, tH);
 760     t11         = _mm_sub_pd(t11, tJ);
 761     t12         = _mm_sub_pd(t12, tL);
 762
 763     _mm_storeu_pd(ptrA,  t1);
 764     _mm_storeu_pd(ptrA+2, t2);
 765     _mm_storeu_pd(ptrA+4, t3);
 766     _mm_storeu_pd(ptrA+6, t4);
 767     _mm_storeu_pd(ptrA+8, t5);
 768     _mm_storeu_pd(ptrA+10, t6);
 769     _mm_storeu_pd(ptrB,  t7);
 770     _mm_storeu_pd(ptrB+2, t8);
 771     _mm_storeu_pd(ptrB+4, t9);
 772     _mm_storeu_pd(ptrB+6, t10);
 773     _mm_storeu_pd(ptrB+8, t11);
 774     _mm_storeu_pd(ptrB+10, t12);
 775 }
 776 #endif
 777
 778
 779 static gmx_inline void
 780 gmx_mm_update_iforce_1atom_swizzle_pd(__m128d fix1, __m128d fiy1, __m128d fiz1,
 781                                       double * gmx_restrict fptr,
 782                                       double * gmx_restrict fshiftptr)
 783 {
 784     fix1 = _mm_hadd_pd(fix1, fiy1);
 785     fiz1 = _mm_hadd_pd(fiz1, fiz1);
 786
 787     _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr), fix1 ));
 788     _mm_store_sd( fptr+2, _mm_add_sd( _mm_load_sd(fptr+2), fiz1 ));
 789
 790     _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 ));
 791     _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 ));
 792 }
 793
 794 #if defined (_MSC_VER) && defined(_M_IX86)
 795 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 796 #define gmx_mm_update_iforce_3atom_swizzle_pd(fix1, fiy1, fiz1, fix2, fiy2, fiz2, fix3, fiy3, fiz3, \
 797                                               fptr, fshiftptr) \
 798     { \
 799         __m128d _t1, _t2; \
 800         fix1 = _mm_hadd_pd(fix1, fiy1); \
 801         fiz1 = _mm_hadd_pd(fiz1, fix2); \
 802         fiy2 = _mm_hadd_pd(fiy2, fiz2); \
 803         fix3 = _mm_hadd_pd(fix3, fiy3); \
 804         fiz3 = _mm_hadd_pd(fiz3, fiz3); \
 805         _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr), fix1 )); \
 806         _mm_storeu_pd( fptr+2, _mm_add_pd( _mm_loadu_pd(fptr+2), fiz1 )); \
 807         _mm_storeu_pd( fptr+4, _mm_add_pd( _mm_loadu_pd(fptr+4), fiy2 )); \
 808         _mm_storeu_pd( fptr+6, _mm_add_pd( _mm_loadu_pd(fptr+6), fix3 )); \
 809         _mm_store_sd( fptr+8, _mm_add_sd( _mm_load_sd(fptr+8), fiz3 )); \
 810         fix1  = _mm_add_pd(fix1, fix3); \
 811         _t1   = _mm_shuffle_pd(fiz1, fiy2, _MM_SHUFFLE2(0, 1)); \
 812         fix1  = _mm_add_pd(fix1, _t1); \
 813         _t2   = _mm_shuffle_pd(fiy2, fiy2, _MM_SHUFFLE2(1, 1)); \
 814         fiz1  = _mm_add_sd(fiz1, fiz3); \
 815         fiz1  = _mm_add_sd(fiz1, _t2); \
 816         _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 )); \
 817         _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 )); \
 818     }
 819 #else
 820 /* Real function for sane compilers */
 821 static gmx_inline void
 822 gmx_mm_update_iforce_3atom_swizzle_pd(__m128d fix1, __m128d fiy1, __m128d fiz1,
 823                                       __m128d fix2, __m128d fiy2, __m128d fiz2,
 824                                       __m128d fix3, __m128d fiy3, __m128d fiz3,
 825                                       double * gmx_restrict fptr,
 826                                       double * gmx_restrict fshiftptr)
 827 {
 828     __m128d t1, t2;
 829
 830     fix1 = _mm_hadd_pd(fix1, fiy1);
 831     fiz1 = _mm_hadd_pd(fiz1, fix2);
 832     fiy2 = _mm_hadd_pd(fiy2, fiz2);
 833     fix3 = _mm_hadd_pd(fix3, fiy3);
 834     fiz3 = _mm_hadd_pd(fiz3, fiz3);
 835
 836     _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr), fix1 ));
 837     _mm_storeu_pd( fptr+2, _mm_add_pd( _mm_loadu_pd(fptr+2), fiz1 ));
 838     _mm_storeu_pd( fptr+4, _mm_add_pd( _mm_loadu_pd(fptr+4), fiy2 ));
 839     _mm_storeu_pd( fptr+6, _mm_add_pd( _mm_loadu_pd(fptr+6), fix3 ));
 840     _mm_store_sd( fptr+8, _mm_add_sd( _mm_load_sd(fptr+8), fiz3 ));
 841
 842     fix1 = _mm_add_pd(fix1, fix3);
 843     t1   = _mm_shuffle_pd(fiz1, fiy2, _MM_SHUFFLE2(0, 1));
 844     fix1 = _mm_add_pd(fix1, t1); /* x and y sums */
 845
 846     t2   = _mm_shuffle_pd(fiy2, fiy2, _MM_SHUFFLE2(1, 1));
 847     fiz1 = _mm_add_sd(fiz1, fiz3);
 848     fiz1 = _mm_add_sd(fiz1, t2); /* z sum */
 849
 850     _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 ));
 851     _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 ));
 852 }
 853 #endif
 854
 855 #if defined (_MSC_VER) && defined(_M_IX86)
 856 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 857 #define gmx_mm_update_iforce_4atom_swizzle_pd(fix1, fiy1, fiz1, fix2, fiy2, fiz2, fix3, fiy3, fiz3, fix4, fiy4, fiz4, \
 858                                               fptr, fshiftptr) \
 859     { \
 860         __m128d _t1, _t2; \
 861         fix1 = _mm_hadd_pd(fix1, fiy1); \
 862         fiz1 = _mm_hadd_pd(fiz1, fix2); \
 863         fiy2 = _mm_hadd_pd(fiy2, fiz2); \
 864         fix3 = _mm_hadd_pd(fix3, fiy3); \
 865         fiz3 = _mm_hadd_pd(fiz3, fix4); \
 866         fiy4 = _mm_hadd_pd(fiy4, fiz4); \
 867         _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr),       fix1 )); \
 868         _mm_storeu_pd( fptr+2, _mm_add_pd( _mm_loadu_pd(fptr+2),   fiz1 )); \
 869         _mm_storeu_pd( fptr+4, _mm_add_pd( _mm_loadu_pd(fptr+4),   fiy2 )); \
 870         _mm_storeu_pd( fptr+6, _mm_add_pd( _mm_loadu_pd(fptr+6),   fix3 )); \
 871         _mm_storeu_pd( fptr+8, _mm_add_pd( _mm_loadu_pd(fptr+8),   fiz3 )); \
 872         _mm_storeu_pd( fptr+10, _mm_add_pd( _mm_loadu_pd(fptr+10), fiy4 )); \
 873         _t1  = _mm_shuffle_pd(fiz1, fiy2, _MM_SHUFFLE2(0, 1)); \
 874         fix1 = _mm_add_pd(fix1, _t1); \
 875         _t2  = _mm_shuffle_pd(fiz3, fiy4, _MM_SHUFFLE2(0, 1)); \
 876         fix3 = _mm_add_pd(fix3, _t2); \
 877         fix1 = _mm_add_pd(fix1, fix3); \
 878         fiz1 = _mm_add_sd(fiz1, _mm_unpackhi_pd(fiy2, fiy2)); \
 879         fiz3 = _mm_add_sd(fiz3, _mm_unpackhi_pd(fiy4, fiy4)); \
 880         fiz1 = _mm_add_sd(fiz1, fiz3); \
 881         _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 )); \
 882         _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 )); \
 883     }
 884 #else
 885 /* Real function for sane compilers */
 886 static gmx_inline void
 887 gmx_mm_update_iforce_4atom_swizzle_pd(__m128d fix1, __m128d fiy1, __m128d fiz1,
 888                                       __m128d fix2, __m128d fiy2, __m128d fiz2,
 889                                       __m128d fix3, __m128d fiy3, __m128d fiz3,
 890                                       __m128d fix4, __m128d fiy4, __m128d fiz4,
 891                                       double * gmx_restrict fptr,
 892                                       double * gmx_restrict fshiftptr)
 893 {
 894     __m128d t1, t2;
 895
 896     fix1 = _mm_hadd_pd(fix1, fiy1);
 897     fiz1 = _mm_hadd_pd(fiz1, fix2);
 898     fiy2 = _mm_hadd_pd(fiy2, fiz2);
 899     fix3 = _mm_hadd_pd(fix3, fiy3);
 900     fiz3 = _mm_hadd_pd(fiz3, fix4);
 901     fiy4 = _mm_hadd_pd(fiy4, fiz4);
 902
 903     _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr),       fix1 ));
 904     _mm_storeu_pd( fptr+2, _mm_add_pd( _mm_loadu_pd(fptr+2),   fiz1 ));
 905     _mm_storeu_pd( fptr+4, _mm_add_pd( _mm_loadu_pd(fptr+4),   fiy2 ));
 906     _mm_storeu_pd( fptr+6, _mm_add_pd( _mm_loadu_pd(fptr+6),   fix3 ));
 907     _mm_storeu_pd( fptr+8, _mm_add_pd( _mm_loadu_pd(fptr+8),   fiz3 ));
 908     _mm_storeu_pd( fptr+10, _mm_add_pd( _mm_loadu_pd(fptr+10), fiy4 ));
 909
 910     t1   = _mm_shuffle_pd(fiz1, fiy2, _MM_SHUFFLE2(0, 1));
 911     fix1 = _mm_add_pd(fix1, t1);
 912     t2   = _mm_shuffle_pd(fiz3, fiy4, _MM_SHUFFLE2(0, 1));
 913     fix3 = _mm_add_pd(fix3, t2);
 914     fix1 = _mm_add_pd(fix1, fix3); /* x and y sums */
 915
 916     fiz1 = _mm_add_sd(fiz1, _mm_unpackhi_pd(fiy2, fiy2));
 917     fiz3 = _mm_add_sd(fiz3, _mm_unpackhi_pd(fiy4, fiy4));
 918     fiz1 = _mm_add_sd(fiz1, fiz3); /* z sum */
 919
 920     _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 ));
 921     _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 ));
 922 }
 923 #endif
 924
 925
 926 static gmx_inline void
 927 gmx_mm_update_1pot_pd(__m128d pot1, double * gmx_restrict ptrA)
 928 {
 929     pot1 = _mm_hadd_pd(pot1, pot1);
 930     _mm_store_sd(ptrA, _mm_add_sd(pot1, _mm_load_sd(ptrA)));
 931 }
 932
 933 static gmx_inline void
 934 gmx_mm_update_2pot_pd(__m128d pot1, double * gmx_restrict ptrA,
 935                       __m128d pot2, double * gmx_restrict ptrB)
 936 {
 937     pot1 = _mm_hadd_pd(pot1, pot2);
 938     pot2 = _mm_unpackhi_pd(pot1, pot1);
 939
 940     _mm_store_sd(ptrA, _mm_add_sd(pot1, _mm_load_sd(ptrA)));
 941     _mm_store_sd(ptrB, _mm_add_sd(pot2, _mm_load_sd(ptrB)));
 942 }
 943
 944
 945 #endif /* _kernelutil_x86_avx_128_fma_double_h_ */