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

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 #ifndef _kernelutil_x86_avx_128_fma_double_h_
  36 #define _kernelutil_x86_avx_128_fma_double_h_
  37
  38 #include "gmx_x86_avx_128_fma.h"
  39
  40
  41 static int
  42 gmx_mm_any_lt(__m128d a, __m128d b)
  43 {
  44     return _mm_movemask_pd(_mm_cmplt_pd(a, b));
  45 }
  46
  47
  48 static gmx_inline __m128d
  49 gmx_mm_calc_rsq_pd(__m128d dx, __m128d dy, __m128d dz)
  50 {
  51     return _mm_macc_pd(dx, dx, _mm_macc_pd(dy, dy, _mm_mul_pd(dz, dz)));
  52 }
  53
  54 /* Normal sum of four ymm registers */
  55 #define gmx_mm_sum4_pd(t0, t1, t2, t3)  _mm_add_pd(_mm_add_pd(t0, t1), _mm_add_pd(t2, t3))
  56
  57
  58
  59 /* Load a double value from 1-2 places, merge into xmm register */
  60
  61
  62 static __m128d
  63 gmx_mm_load_2real_swizzle_pd(const double * gmx_restrict ptrA,
  64                              const double * gmx_restrict ptrB)
  65 {
  66     return _mm_unpacklo_pd(_mm_load_sd(ptrA), _mm_load_sd(ptrB));
  67 }
  68
  69 static __m128d
  70 gmx_mm_load_1real_pd(const double * gmx_restrict ptrA)
  71 {
  72     return _mm_load_sd(ptrA);
  73 }
  74
  75
  76 static void
  77 gmx_mm_store_2real_swizzle_pd(double * gmx_restrict ptrA,
  78                               double * gmx_restrict ptrB,
  79                               __m128d               xmm1)
  80 {
  81     __m128d t2;
  82
  83     t2       = _mm_unpackhi_pd(xmm1, xmm1);
  84     _mm_store_sd(ptrA, xmm1);
  85     _mm_store_sd(ptrB, t2);
  86 }
  87
  88 static void
  89 gmx_mm_store_1real_pd(double * gmx_restrict ptrA, __m128d xmm1)
  90 {
  91     _mm_store_sd(ptrA, xmm1);
  92 }
  93
  94
  95 /* Similar to store, but increments value in memory */
  96 static void
  97 gmx_mm_increment_2real_swizzle_pd(double * gmx_restrict ptrA,
  98                                   double * gmx_restrict ptrB, __m128d xmm1)
  99 {
 100     __m128d t1;
 101
 102     t1   = _mm_unpackhi_pd(xmm1, xmm1);
 103     xmm1 = _mm_add_sd(xmm1, _mm_load_sd(ptrA));
 104     t1   = _mm_add_sd(t1, _mm_load_sd(ptrB));
 105     _mm_store_sd(ptrA, xmm1);
 106     _mm_store_sd(ptrB, t1);
 107 }
 108
 109 static void
 110 gmx_mm_increment_1real_pd(double * gmx_restrict ptrA, __m128d xmm1)
 111 {
 112     __m128d tmp;
 113
 114     tmp = gmx_mm_load_1real_pd(ptrA);
 115     tmp = _mm_add_sd(tmp, xmm1);
 116     gmx_mm_store_1real_pd(ptrA, tmp);
 117 }
 118
 119
 120
 121 static gmx_inline void
 122 gmx_mm_load_2pair_swizzle_pd(const double * gmx_restrict p1,
 123                              const double * gmx_restrict p2,
 124                              __m128d * gmx_restrict      c6,
 125                              __m128d * gmx_restrict      c12)
 126 {
 127     __m128d t1, t2, t3;
 128
 129     /* The c6/c12 array should be aligned */
 130     t1   = _mm_loadu_pd(p1);
 131     t2   = _mm_loadu_pd(p2);
 132     *c6  = _mm_unpacklo_pd(t1, t2);
 133     *c12 = _mm_unpackhi_pd(t1, t2);
 134 }
 135
 136 static gmx_inline void
 137 gmx_mm_load_1pair_swizzle_pd(const double * gmx_restrict p1,
 138                              __m128d * gmx_restrict      c6,
 139                              __m128d * gmx_restrict      c12)
 140 {
 141     *c6     = _mm_load_sd(p1);
 142     *c12    = _mm_load_sd(p1+1);
 143 }
 144
 145
 146 static gmx_inline void
 147 gmx_mm_load_shift_and_1rvec_broadcast_pd(const double * gmx_restrict xyz_shift,
 148                                          const double * gmx_restrict xyz,
 149                                          __m128d * gmx_restrict      x1,
 150                                          __m128d * gmx_restrict      y1,
 151                                          __m128d * gmx_restrict      z1)
 152 {
 153     __m128d mem_xy, mem_z, mem_sxy, mem_sz;
 154
 155     mem_xy  = _mm_loadu_pd(xyz);
 156     mem_z   = _mm_load_sd(xyz+2);
 157     mem_sxy = _mm_loadu_pd(xyz_shift);
 158     mem_sz  = _mm_load_sd(xyz_shift+2);
 159
 160     mem_xy  = _mm_add_pd(mem_xy, mem_sxy);
 161     mem_z   = _mm_add_pd(mem_z, mem_sz);
 162
 163     *x1  = _mm_shuffle_pd(mem_xy, mem_xy, _MM_SHUFFLE2(0, 0));
 164     *y1  = _mm_shuffle_pd(mem_xy, mem_xy, _MM_SHUFFLE2(1, 1));
 165     *z1  = _mm_shuffle_pd(mem_z, mem_z, _MM_SHUFFLE2(0, 0));
 166 }
 167
 168
 169 static gmx_inline void
 170 gmx_mm_load_shift_and_3rvec_broadcast_pd(const double * gmx_restrict xyz_shift,
 171                                          const double * gmx_restrict xyz,
 172                                          __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 173                                          __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 174                                          __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3)
 175 {
 176     __m128d t1, t2, t3, t4, t5, sxy, sz, szx, syz;
 177
 178     t1  = _mm_loadu_pd(xyz);
 179     t2  = _mm_loadu_pd(xyz+2);
 180     t3  = _mm_loadu_pd(xyz+4);
 181     t4  = _mm_loadu_pd(xyz+6);
 182     t5  = _mm_load_sd(xyz+8);
 183
 184     sxy = _mm_loadu_pd(xyz_shift);
 185     sz  = _mm_load_sd(xyz_shift+2);
 186     szx = _mm_shuffle_pd(sz, sxy, _MM_SHUFFLE2(0, 0));
 187     syz = _mm_shuffle_pd(sxy, sz, _MM_SHUFFLE2(0, 1));
 188
 189     t1  = _mm_add_pd(t1, sxy);
 190     t2  = _mm_add_pd(t2, szx);
 191     t3  = _mm_add_pd(t3, syz);
 192     t4  = _mm_add_pd(t4, sxy);
 193     t5  = _mm_add_sd(t5, sz);
 194
 195     *x1  = _mm_shuffle_pd(t1, t1, _MM_SHUFFLE2(0, 0));
 196     *y1  = _mm_shuffle_pd(t1, t1, _MM_SHUFFLE2(1, 1));
 197     *z1  = _mm_shuffle_pd(t2, t2, _MM_SHUFFLE2(0, 0));
 198     *x2  = _mm_shuffle_pd(t2, t2, _MM_SHUFFLE2(1, 1));
 199     *y2  = _mm_shuffle_pd(t3, t3, _MM_SHUFFLE2(0, 0));
 200     *z2  = _mm_shuffle_pd(t3, t3, _MM_SHUFFLE2(1, 1));
 201     *x3  = _mm_shuffle_pd(t4, t4, _MM_SHUFFLE2(0, 0));
 202     *y3  = _mm_shuffle_pd(t4, t4, _MM_SHUFFLE2(1, 1));
 203     *z3  = _mm_shuffle_pd(t5, t5, _MM_SHUFFLE2(0, 0));
 204 }
 205
 206
 207 static gmx_inline void
 208 gmx_mm_load_shift_and_4rvec_broadcast_pd(const double * gmx_restrict xyz_shift,
 209                                          const double * gmx_restrict xyz,
 210                                          __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 211                                          __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 212                                          __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3,
 213                                          __m128d * gmx_restrict x4, __m128d * gmx_restrict y4, __m128d * gmx_restrict z4)
 214 {
 215     __m128d t1, t2, t3, t4, t5, t6, sxy, sz, szx, syz;
 216
 217     t1  = _mm_loadu_pd(xyz);
 218     t2  = _mm_loadu_pd(xyz+2);
 219     t3  = _mm_loadu_pd(xyz+4);
 220     t4  = _mm_loadu_pd(xyz+6);
 221     t5  = _mm_loadu_pd(xyz+8);
 222     t6  = _mm_loadu_pd(xyz+10);
 223
 224     sxy = _mm_loadu_pd(xyz_shift);
 225     sz  = _mm_load_sd(xyz_shift+2);
 226     szx = _mm_shuffle_pd(sz, sxy, _MM_SHUFFLE2(0, 0));
 227     syz = _mm_shuffle_pd(sxy, sz, _MM_SHUFFLE2(0, 1));
 228
 229     t1  = _mm_add_pd(t1, sxy);
 230     t2  = _mm_add_pd(t2, szx);
 231     t3  = _mm_add_pd(t3, syz);
 232     t4  = _mm_add_pd(t4, sxy);
 233     t5  = _mm_add_pd(t5, szx);
 234     t6  = _mm_add_pd(t6, syz);
 235
 236     *x1  = _mm_shuffle_pd(t1, t1, _MM_SHUFFLE2(0, 0));
 237     *y1  = _mm_shuffle_pd(t1, t1, _MM_SHUFFLE2(1, 1));
 238     *z1  = _mm_shuffle_pd(t2, t2, _MM_SHUFFLE2(0, 0));
 239     *x2  = _mm_shuffle_pd(t2, t2, _MM_SHUFFLE2(1, 1));
 240     *y2  = _mm_shuffle_pd(t3, t3, _MM_SHUFFLE2(0, 0));
 241     *z2  = _mm_shuffle_pd(t3, t3, _MM_SHUFFLE2(1, 1));
 242     *x3  = _mm_shuffle_pd(t4, t4, _MM_SHUFFLE2(0, 0));
 243     *y3  = _mm_shuffle_pd(t4, t4, _MM_SHUFFLE2(1, 1));
 244     *z3  = _mm_shuffle_pd(t5, t5, _MM_SHUFFLE2(0, 0));
 245     *x4  = _mm_shuffle_pd(t5, t5, _MM_SHUFFLE2(1, 1));
 246     *y4  = _mm_shuffle_pd(t6, t6, _MM_SHUFFLE2(0, 0));
 247     *z4  = _mm_shuffle_pd(t6, t6, _MM_SHUFFLE2(1, 1));
 248 }
 249
 250
 251
 252 static gmx_inline void
 253 gmx_mm_load_1rvec_1ptr_swizzle_pd(const double * gmx_restrict p1,
 254                                   __m128d * gmx_restrict x, __m128d * gmx_restrict y, __m128d * gmx_restrict z)
 255 {
 256     *x            = _mm_load_sd(p1);
 257     *y            = _mm_load_sd(p1+1);
 258     *z            = _mm_load_sd(p1+2);
 259 }
 260
 261 static gmx_inline void
 262 gmx_mm_load_3rvec_1ptr_swizzle_pd(const double * gmx_restrict p1,
 263                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 264                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 265                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3)
 266 {
 267     *x1            = _mm_load_sd(p1);
 268     *y1            = _mm_load_sd(p1+1);
 269     *z1            = _mm_load_sd(p1+2);
 270     *x2            = _mm_load_sd(p1+3);
 271     *y2            = _mm_load_sd(p1+4);
 272     *z2            = _mm_load_sd(p1+5);
 273     *x3            = _mm_load_sd(p1+6);
 274     *y3            = _mm_load_sd(p1+7);
 275     *z3            = _mm_load_sd(p1+8);
 276 }
 277
 278 static gmx_inline void
 279 gmx_mm_load_4rvec_1ptr_swizzle_pd(const double * gmx_restrict p1,
 280                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 281                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 282                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3,
 283                                   __m128d * gmx_restrict x4, __m128d * gmx_restrict y4, __m128d * gmx_restrict z4)
 284 {
 285     *x1            = _mm_load_sd(p1);
 286     *y1            = _mm_load_sd(p1+1);
 287     *z1            = _mm_load_sd(p1+2);
 288     *x2            = _mm_load_sd(p1+3);
 289     *y2            = _mm_load_sd(p1+4);
 290     *z2            = _mm_load_sd(p1+5);
 291     *x3            = _mm_load_sd(p1+6);
 292     *y3            = _mm_load_sd(p1+7);
 293     *z3            = _mm_load_sd(p1+8);
 294     *x4            = _mm_load_sd(p1+9);
 295     *y4            = _mm_load_sd(p1+10);
 296     *z4            = _mm_load_sd(p1+11);
 297 }
 298
 299
 300 static gmx_inline void
 301 gmx_mm_load_1rvec_2ptr_swizzle_pd(const double * gmx_restrict ptrA,
 302                                   const double * gmx_restrict ptrB,
 303                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1)
 304 {
 305     __m128d t1, t2, t3, t4;
 306     t1           = _mm_loadu_pd(ptrA);
 307     t2           = _mm_loadu_pd(ptrB);
 308     t3           = _mm_load_sd(ptrA+2);
 309     t4           = _mm_load_sd(ptrB+2);
 310     GMX_MM_TRANSPOSE2_PD(t1, t2);
 311     *x1          = t1;
 312     *y1          = t2;
 313     *z1          = _mm_unpacklo_pd(t3, t4);
 314 }
 315
 316 static gmx_inline void
 317 gmx_mm_load_3rvec_2ptr_swizzle_pd(const double * gmx_restrict ptrA, const double * gmx_restrict ptrB,
 318                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 319                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 320                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3)
 321 {
 322     __m128d t1, t2, t3, t4, t5, t6, t7, t8, t9, t10;
 323     t1           = _mm_loadu_pd(ptrA);
 324     t2           = _mm_loadu_pd(ptrB);
 325     t3           = _mm_loadu_pd(ptrA+2);
 326     t4           = _mm_loadu_pd(ptrB+2);
 327     t5           = _mm_loadu_pd(ptrA+4);
 328     t6           = _mm_loadu_pd(ptrB+4);
 329     t7           = _mm_loadu_pd(ptrA+6);
 330     t8           = _mm_loadu_pd(ptrB+6);
 331     t9           = _mm_load_sd(ptrA+8);
 332     t10          = _mm_load_sd(ptrB+8);
 333     GMX_MM_TRANSPOSE2_PD(t1, t2);
 334     GMX_MM_TRANSPOSE2_PD(t3, t4);
 335     GMX_MM_TRANSPOSE2_PD(t5, t6);
 336     GMX_MM_TRANSPOSE2_PD(t7, t8);
 337     *x1          = t1;
 338     *y1          = t2;
 339     *z1          = t3;
 340     *x2          = t4;
 341     *y2          = t5;
 342     *z2          = t6;
 343     *x3          = t7;
 344     *y3          = t8;
 345     *z3          = _mm_unpacklo_pd(t9, t10);
 346 }
 347
 348
 349 static gmx_inline void
 350 gmx_mm_load_4rvec_2ptr_swizzle_pd(const double * gmx_restrict ptrA, const double * gmx_restrict ptrB,
 351                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 352                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 353                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3,
 354                                   __m128d * gmx_restrict x4, __m128d * gmx_restrict y4, __m128d * gmx_restrict z4)
 355 {
 356     __m128d t1, t2, t3, t4, t5, t6;
 357     t1           = _mm_loadu_pd(ptrA);
 358     t2           = _mm_loadu_pd(ptrB);
 359     t3           = _mm_loadu_pd(ptrA+2);
 360     t4           = _mm_loadu_pd(ptrB+2);
 361     t5           = _mm_loadu_pd(ptrA+4);
 362     t6           = _mm_loadu_pd(ptrB+4);
 363     GMX_MM_TRANSPOSE2_PD(t1, t2);
 364     GMX_MM_TRANSPOSE2_PD(t3, t4);
 365     GMX_MM_TRANSPOSE2_PD(t5, t6);
 366     *x1          = t1;
 367     *y1          = t2;
 368     *z1          = t3;
 369     *x2          = t4;
 370     *y2          = t5;
 371     *z2          = t6;
 372     t1           = _mm_loadu_pd(ptrA+6);
 373     t2           = _mm_loadu_pd(ptrB+6);
 374     t3           = _mm_loadu_pd(ptrA+8);
 375     t4           = _mm_loadu_pd(ptrB+8);
 376     t5           = _mm_loadu_pd(ptrA+10);
 377     t6           = _mm_loadu_pd(ptrB+10);
 378     GMX_MM_TRANSPOSE2_PD(t1, t2);
 379     GMX_MM_TRANSPOSE2_PD(t3, t4);
 380     GMX_MM_TRANSPOSE2_PD(t5, t6);
 381     *x3          = t1;
 382     *y3          = t2;
 383     *z3          = t3;
 384     *x4          = t4;
 385     *y4          = t5;
 386     *z4          = t6;
 387 }
 388
 389
 390 /* Routines to decrement rvec in memory, typically use for j particle force updates */
 391 static void
 392 gmx_mm_decrement_1rvec_1ptr_swizzle_pd(double * gmx_restrict ptrA,
 393                                        __m128d x1, __m128d y1, __m128d z1)
 394 {
 395     __m128d t1, t2, t3;
 396
 397     t1           = _mm_load_sd(ptrA);
 398     t2           = _mm_load_sd(ptrA+1);
 399     t3           = _mm_load_sd(ptrA+2);
 400
 401     t1           = _mm_sub_sd(t1, x1);
 402     t2           = _mm_sub_sd(t2, y1);
 403     t3           = _mm_sub_sd(t3, z1);
 404     _mm_store_sd(ptrA, t1);
 405     _mm_store_sd(ptrA+1, t2);
 406     _mm_store_sd(ptrA+2, t3);
 407 }
 408
 409
 410 #if defined (_MSC_VER) && defined(_M_IX86)
 411 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 412 #define gmx_mm_decrement_3rvec_1ptr_swizzle_pd(ptrA, _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3) \
 413     { \
 414         __m128d _t1, _t2, _t3, _t4, _t5; \
 415         _t1          = _mm_loadu_pd(ptrA); \
 416         _t2          = _mm_loadu_pd(ptrA+2); \
 417         _t3          = _mm_loadu_pd(ptrA+4); \
 418         _t4          = _mm_loadu_pd(ptrA+6); \
 419         _t5          = _mm_load_sd(ptrA+8); \
 420         _x1          = _mm_unpacklo_pd(_x1, _y1); \
 421         _z1          = _mm_unpacklo_pd(_z1, _x2); \
 422         _y2          = _mm_unpacklo_pd(_y2, _z2); \
 423         _x3          = _mm_unpacklo_pd(_x3, _y3); \
 424         _t1          = _mm_sub_pd(_t1, _x1); \
 425         _t2          = _mm_sub_pd(_t2, _z1); \
 426         _t3          = _mm_sub_pd(_t3, _y2); \
 427         _t4          = _mm_sub_pd(_t4, _x3); \
 428         _t5          = _mm_sub_sd(_t5, _z3); \
 429         _mm_storeu_pd(ptrA, _t1); \
 430         _mm_storeu_pd(ptrA+2, _t2); \
 431         _mm_storeu_pd(ptrA+4, _t3); \
 432         _mm_storeu_pd(ptrA+6, _t4); \
 433         _mm_store_sd(ptrA+8, _t5); \
 434     }
 435 #else
 436 /* Real function for sane compilers */
 437 static void
 438 gmx_mm_decrement_3rvec_1ptr_swizzle_pd(double * gmx_restrict ptrA,
 439                                        __m128d x1, __m128d y1, __m128d z1,
 440                                        __m128d x2, __m128d y2, __m128d z2,
 441                                        __m128d x3, __m128d y3, __m128d z3)
 442 {
 443     __m128d t1, t2, t3, t4, t5;
 444
 445     t1          = _mm_loadu_pd(ptrA);
 446     t2          = _mm_loadu_pd(ptrA+2);
 447     t3          = _mm_loadu_pd(ptrA+4);
 448     t4          = _mm_loadu_pd(ptrA+6);
 449     t5          = _mm_load_sd(ptrA+8);
 450
 451     x1          = _mm_unpacklo_pd(x1, y1);
 452     z1          = _mm_unpacklo_pd(z1, x2);
 453     y2          = _mm_unpacklo_pd(y2, z2);
 454     x3          = _mm_unpacklo_pd(x3, y3);
 455     /* nothing to be done for z3 */
 456
 457     t1          = _mm_sub_pd(t1, x1);
 458     t2          = _mm_sub_pd(t2, z1);
 459     t3          = _mm_sub_pd(t3, y2);
 460     t4          = _mm_sub_pd(t4, x3);
 461     t5          = _mm_sub_sd(t5, z3);
 462     _mm_storeu_pd(ptrA, t1);
 463     _mm_storeu_pd(ptrA+2, t2);
 464     _mm_storeu_pd(ptrA+4, t3);
 465     _mm_storeu_pd(ptrA+6, t4);
 466     _mm_store_sd(ptrA+8, t5);
 467 }
 468 #endif
 469
 470
 471 #if defined (_MSC_VER) && defined(_M_IX86)
 472 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 473 #define gmx_mm_decrement_4rvec_1ptr_swizzle_pd(ptrA, _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3, _x4, _y4, _z4) \
 474     { \
 475         __m128d _t1, _t2, _t3, _t4, _t5, _t6; \
 476         _t1          = _mm_loadu_pd(ptrA); \
 477         _t2          = _mm_loadu_pd(ptrA+2); \
 478         _t3          = _mm_loadu_pd(ptrA+4); \
 479         _t4          = _mm_loadu_pd(ptrA+6); \
 480         _t5          = _mm_loadu_pd(ptrA+8); \
 481         _t6          = _mm_loadu_pd(ptrA+10); \
 482         _x1          = _mm_unpacklo_pd(_x1, _y1); \
 483         _z1          = _mm_unpacklo_pd(_z1, _x2); \
 484         _y2          = _mm_unpacklo_pd(_y2, _z2); \
 485         _x3          = _mm_unpacklo_pd(_x3, _y3); \
 486         _z3          = _mm_unpacklo_pd(_z3, _x4); \
 487         _y4          = _mm_unpacklo_pd(_y4, _z4); \
 488         _mm_storeu_pd(ptrA,    _mm_sub_pd( _t1, _x1 )); \
 489         _mm_storeu_pd(ptrA+2,  _mm_sub_pd( _t2, _z1 )); \
 490         _mm_storeu_pd(ptrA+4,  _mm_sub_pd( _t3, _y2 )); \
 491         _mm_storeu_pd(ptrA+6,  _mm_sub_pd( _t4, _x3 )); \
 492         _mm_storeu_pd(ptrA+8,  _mm_sub_pd( _t5, _z3 )); \
 493         _mm_storeu_pd(ptrA+10, _mm_sub_pd( _t6, _y4 )); \
 494     }
 495 #else
 496 /* Real function for sane compilers */
 497 static void
 498 gmx_mm_decrement_4rvec_1ptr_swizzle_pd(double * gmx_restrict ptrA,
 499                                        __m128d x1, __m128d y1, __m128d z1,
 500                                        __m128d x2, __m128d y2, __m128d z2,
 501                                        __m128d x3, __m128d y3, __m128d z3,
 502                                        __m128d x4, __m128d y4, __m128d z4)
 503 {
 504     __m128d t1, t2, t3, t4, t5, t6;
 505
 506     t1          = _mm_loadu_pd(ptrA);
 507     t2          = _mm_loadu_pd(ptrA+2);
 508     t3          = _mm_loadu_pd(ptrA+4);
 509     t4          = _mm_loadu_pd(ptrA+6);
 510     t5          = _mm_loadu_pd(ptrA+8);
 511     t6          = _mm_loadu_pd(ptrA+10);
 512
 513     x1          = _mm_unpacklo_pd(x1, y1);
 514     z1          = _mm_unpacklo_pd(z1, x2);
 515     y2          = _mm_unpacklo_pd(y2, z2);
 516     x3          = _mm_unpacklo_pd(x3, y3);
 517     z3          = _mm_unpacklo_pd(z3, x4);
 518     y4          = _mm_unpacklo_pd(y4, z4);
 519
 520     _mm_storeu_pd(ptrA,    _mm_sub_pd( t1, x1 ));
 521     _mm_storeu_pd(ptrA+2,  _mm_sub_pd( t2, z1 ));
 522     _mm_storeu_pd(ptrA+4,  _mm_sub_pd( t3, y2 ));
 523     _mm_storeu_pd(ptrA+6,  _mm_sub_pd( t4, x3 ));
 524     _mm_storeu_pd(ptrA+8,  _mm_sub_pd( t5, z3 ));
 525     _mm_storeu_pd(ptrA+10, _mm_sub_pd( t6, y4 ));
 526 }
 527 #endif
 528
 529
 530 static void
 531 gmx_mm_decrement_1rvec_2ptr_swizzle_pd(double * gmx_restrict ptrA, double * gmx_restrict ptrB,
 532                                        __m128d x1, __m128d y1, __m128d z1)
 533 {
 534     __m128d t1, t2, t3, t4, t5, t6, t7;
 535
 536     t1          = _mm_loadu_pd(ptrA);
 537     t2          = _mm_load_sd(ptrA+2);
 538     t3          = _mm_loadu_pd(ptrB);
 539     t4          = _mm_load_sd(ptrB+2);
 540
 541     t5          = _mm_unpacklo_pd(x1, y1);
 542     t6          = _mm_unpackhi_pd(x1, y1);
 543     t7          = _mm_unpackhi_pd(z1, z1);
 544
 545     t1          = _mm_sub_pd(t1, t5);
 546     t2          = _mm_sub_sd(t2, z1);
 547
 548     t3          = _mm_sub_pd(t3, t6);
 549     t4          = _mm_sub_sd(t4, t7);
 550
 551     _mm_storeu_pd(ptrA, t1);
 552     _mm_store_sd(ptrA+2, t2);
 553     _mm_storeu_pd(ptrB, t3);
 554     _mm_store_sd(ptrB+2, t4);
 555 }
 556
 557
 558 #if defined (_MSC_VER) && defined(_M_IX86)
 559 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 560 #define gmx_mm_decrement_3rvec_2ptr_swizzle_pd(ptrA, ptrB, _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3) \
 561     { \
 562         __m128d _t1, _t2, _t3, _t4, _t5, _t6, _t7, _t8, _t9, _t10; \
 563         __m128d _tA, _tB, _tC, _tD, _tE, _tF, _tG, _tH, _tI; \
 564         _t1          = _mm_loadu_pd(ptrA); \
 565         _t2          = _mm_loadu_pd(ptrA+2); \
 566         _t3          = _mm_loadu_pd(ptrA+4); \
 567         _t4          = _mm_loadu_pd(ptrA+6); \
 568         _t5          = _mm_load_sd(ptrA+8); \
 569         _t6          = _mm_loadu_pd(ptrB); \
 570         _t7          = _mm_loadu_pd(ptrB+2); \
 571         _t8          = _mm_loadu_pd(ptrB+4); \
 572         _t9          = _mm_loadu_pd(ptrB+6); \
 573         _t10         = _mm_load_sd(ptrB+8); \
 574         _tA          = _mm_unpacklo_pd(_x1, _y1); \
 575         _tB          = _mm_unpackhi_pd(_x1, _y1); \
 576         _tC          = _mm_unpacklo_pd(_z1, _x2); \
 577         _tD          = _mm_unpackhi_pd(_z1, _x2); \
 578         _tE          = _mm_unpacklo_pd(_y2, _z2); \
 579         _tF          = _mm_unpackhi_pd(_y2, _z2); \
 580         _tG          = _mm_unpacklo_pd(_x3, _y3); \
 581         _tH          = _mm_unpackhi_pd(_x3, _y3); \
 582         _tI          = _mm_unpackhi_pd(_z3, _z3); \
 583         _t1          = _mm_sub_pd(_t1, _tA); \
 584         _t2          = _mm_sub_pd(_t2, _tC); \
 585         _t3          = _mm_sub_pd(_t3, _tE); \
 586         _t4          = _mm_sub_pd(_t4, _tG); \
 587         _t5          = _mm_sub_sd(_t5, _z3); \
 588         _t6          = _mm_sub_pd(_t6, _tB); \
 589         _t7          = _mm_sub_pd(_t7, _tD); \
 590         _t8          = _mm_sub_pd(_t8, _tF); \
 591         _t9          = _mm_sub_pd(_t9, _tH); \
 592         _t10         = _mm_sub_sd(_t10, _tI); \
 593         _mm_storeu_pd(ptrA, _t1); \
 594         _mm_storeu_pd(ptrA+2, _t2); \
 595         _mm_storeu_pd(ptrA+4, _t3); \
 596         _mm_storeu_pd(ptrA+6, _t4); \
 597         _mm_store_sd(ptrA+8, _t5); \
 598         _mm_storeu_pd(ptrB, _t6); \
 599         _mm_storeu_pd(ptrB+2, _t7); \
 600         _mm_storeu_pd(ptrB+4, _t8); \
 601         _mm_storeu_pd(ptrB+6, _t9); \
 602         _mm_store_sd(ptrB+8, _t10); \
 603     }
 604 #else
 605 /* Real function for sane compilers */
 606 static void
 607 gmx_mm_decrement_3rvec_2ptr_swizzle_pd(double * gmx_restrict ptrA, double * gmx_restrict ptrB,
 608                                        __m128d x1, __m128d y1, __m128d z1,
 609                                        __m128d x2, __m128d y2, __m128d z2,
 610                                        __m128d x3, __m128d y3, __m128d z3)
 611 {
 612     __m128d t1, t2, t3, t4, t5, t6, t7, t8, t9, t10;
 613     __m128d tA, tB, tC, tD, tE, tF, tG, tH, tI;
 614
 615     t1          = _mm_loadu_pd(ptrA);
 616     t2          = _mm_loadu_pd(ptrA+2);
 617     t3          = _mm_loadu_pd(ptrA+4);
 618     t4          = _mm_loadu_pd(ptrA+6);
 619     t5          = _mm_load_sd(ptrA+8);
 620     t6          = _mm_loadu_pd(ptrB);
 621     t7          = _mm_loadu_pd(ptrB+2);
 622     t8          = _mm_loadu_pd(ptrB+4);
 623     t9          = _mm_loadu_pd(ptrB+6);
 624     t10         = _mm_load_sd(ptrB+8);
 625
 626     tA          = _mm_unpacklo_pd(x1, y1);
 627     tB          = _mm_unpackhi_pd(x1, y1);
 628     tC          = _mm_unpacklo_pd(z1, x2);
 629     tD          = _mm_unpackhi_pd(z1, x2);
 630     tE          = _mm_unpacklo_pd(y2, z2);
 631     tF          = _mm_unpackhi_pd(y2, z2);
 632     tG          = _mm_unpacklo_pd(x3, y3);
 633     tH          = _mm_unpackhi_pd(x3, y3);
 634     tI          = _mm_unpackhi_pd(z3, z3);
 635
 636     t1          = _mm_sub_pd(t1, tA);
 637     t2          = _mm_sub_pd(t2, tC);
 638     t3          = _mm_sub_pd(t3, tE);
 639     t4          = _mm_sub_pd(t4, tG);
 640     t5          = _mm_sub_sd(t5, z3);
 641
 642     t6          = _mm_sub_pd(t6, tB);
 643     t7          = _mm_sub_pd(t7, tD);
 644     t8          = _mm_sub_pd(t8, tF);
 645     t9          = _mm_sub_pd(t9, tH);
 646     t10         = _mm_sub_sd(t10, tI);
 647
 648     _mm_storeu_pd(ptrA, t1);
 649     _mm_storeu_pd(ptrA+2, t2);
 650     _mm_storeu_pd(ptrA+4, t3);
 651     _mm_storeu_pd(ptrA+6, t4);
 652     _mm_store_sd(ptrA+8, t5);
 653     _mm_storeu_pd(ptrB, t6);
 654     _mm_storeu_pd(ptrB+2, t7);
 655     _mm_storeu_pd(ptrB+4, t8);
 656     _mm_storeu_pd(ptrB+6, t9);
 657     _mm_store_sd(ptrB+8, t10);
 658 }
 659 #endif
 660
 661
 662 #if defined (_MSC_VER) && defined(_M_IX86)
 663 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 664 #define gmx_mm_decrement_4rvec_2ptr_swizzle_pd(ptrA, ptrB, _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3, _x4, _y4, _z4) \
 665     { \
 666         __m128d _t1, _t2, _t3, _t4, _t5, _t6, _t7, _t8, _t9, _t10, _t11, _t12; \
 667         __m128d _tA, _tB, _tC, _tD, _tE, _tF, _tG, _tH, _tI, _tJ, _tK, _tL; \
 668         _t1          = _mm_loadu_pd(ptrA); \
 669         _t2          = _mm_loadu_pd(ptrA+2); \
 670         _t3          = _mm_loadu_pd(ptrA+4); \
 671         _t4          = _mm_loadu_pd(ptrA+6); \
 672         _t5          = _mm_loadu_pd(ptrA+8); \
 673         _t6          = _mm_loadu_pd(ptrA+10); \
 674         _t7          = _mm_loadu_pd(ptrB); \
 675         _t8          = _mm_loadu_pd(ptrB+2); \
 676         _t9          = _mm_loadu_pd(ptrB+4); \
 677         _t10         = _mm_loadu_pd(ptrB+6); \
 678         _t11         = _mm_loadu_pd(ptrB+8); \
 679         _t12         = _mm_loadu_pd(ptrB+10); \
 680         _tA          = _mm_unpacklo_pd(_x1, _y1); \
 681         _tB          = _mm_unpackhi_pd(_x1, _y1); \
 682         _tC          = _mm_unpacklo_pd(_z1, _x2); \
 683         _tD          = _mm_unpackhi_pd(_z1, _x2); \
 684         _tE          = _mm_unpacklo_pd(_y2, _z2); \
 685         _tF          = _mm_unpackhi_pd(_y2, _z2); \
 686         _tG          = _mm_unpacklo_pd(_x3, _y3); \
 687         _tH          = _mm_unpackhi_pd(_x3, _y3); \
 688         _tI          = _mm_unpacklo_pd(_z3, _x4); \
 689         _tJ          = _mm_unpackhi_pd(_z3, _x4); \
 690         _tK          = _mm_unpacklo_pd(_y4, _z4); \
 691         _tL          = _mm_unpackhi_pd(_y4, _z4); \
 692         _t1          = _mm_sub_pd(_t1, _tA); \
 693         _t2          = _mm_sub_pd(_t2, _tC); \
 694         _t3          = _mm_sub_pd(_t3, _tE); \
 695         _t4          = _mm_sub_pd(_t4, _tG); \
 696         _t5          = _mm_sub_pd(_t5, _tI); \
 697         _t6          = _mm_sub_pd(_t6, _tK); \
 698         _t7          = _mm_sub_pd(_t7, _tB); \
 699         _t8          = _mm_sub_pd(_t8, _tD); \
 700         _t9          = _mm_sub_pd(_t9, _tF); \
 701         _t10         = _mm_sub_pd(_t10, _tH); \
 702         _t11         = _mm_sub_pd(_t11, _tJ); \
 703         _t12         = _mm_sub_pd(_t12, _tL); \
 704         _mm_storeu_pd(ptrA,  _t1); \
 705         _mm_storeu_pd(ptrA+2, _t2); \
 706         _mm_storeu_pd(ptrA+4, _t3); \
 707         _mm_storeu_pd(ptrA+6, _t4); \
 708         _mm_storeu_pd(ptrA+8, _t5); \
 709         _mm_storeu_pd(ptrA+10, _t6); \
 710         _mm_storeu_pd(ptrB,  _t7); \
 711         _mm_storeu_pd(ptrB+2, _t8); \
 712         _mm_storeu_pd(ptrB+4, _t9); \
 713         _mm_storeu_pd(ptrB+6, _t10); \
 714         _mm_storeu_pd(ptrB+8, _t11); \
 715         _mm_storeu_pd(ptrB+10, _t12); \
 716     }
 717 #else
 718 /* Real function for sane compilers */
 719 static void
 720 gmx_mm_decrement_4rvec_2ptr_swizzle_pd(double * gmx_restrict ptrA, double * gmx_restrict ptrB,
 721                                        __m128d x1, __m128d y1, __m128d z1,
 722                                        __m128d x2, __m128d y2, __m128d z2,
 723                                        __m128d x3, __m128d y3, __m128d z3,
 724                                        __m128d x4, __m128d y4, __m128d z4)
 725 {
 726     __m128d t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12;
 727     __m128d tA, tB, tC, tD, tE, tF, tG, tH, tI, tJ, tK, tL;
 728
 729     t1          = _mm_loadu_pd(ptrA);
 730     t2          = _mm_loadu_pd(ptrA+2);
 731     t3          = _mm_loadu_pd(ptrA+4);
 732     t4          = _mm_loadu_pd(ptrA+6);
 733     t5          = _mm_loadu_pd(ptrA+8);
 734     t6          = _mm_loadu_pd(ptrA+10);
 735     t7          = _mm_loadu_pd(ptrB);
 736     t8          = _mm_loadu_pd(ptrB+2);
 737     t9          = _mm_loadu_pd(ptrB+4);
 738     t10         = _mm_loadu_pd(ptrB+6);
 739     t11         = _mm_loadu_pd(ptrB+8);
 740     t12         = _mm_loadu_pd(ptrB+10);
 741
 742     tA          = _mm_unpacklo_pd(x1, y1);
 743     tB          = _mm_unpackhi_pd(x1, y1);
 744     tC          = _mm_unpacklo_pd(z1, x2);
 745     tD          = _mm_unpackhi_pd(z1, x2);
 746     tE          = _mm_unpacklo_pd(y2, z2);
 747     tF          = _mm_unpackhi_pd(y2, z2);
 748     tG          = _mm_unpacklo_pd(x3, y3);
 749     tH          = _mm_unpackhi_pd(x3, y3);
 750     tI          = _mm_unpacklo_pd(z3, x4);
 751     tJ          = _mm_unpackhi_pd(z3, x4);
 752     tK          = _mm_unpacklo_pd(y4, z4);
 753     tL          = _mm_unpackhi_pd(y4, z4);
 754
 755     t1          = _mm_sub_pd(t1, tA);
 756     t2          = _mm_sub_pd(t2, tC);
 757     t3          = _mm_sub_pd(t3, tE);
 758     t4          = _mm_sub_pd(t4, tG);
 759     t5          = _mm_sub_pd(t5, tI);
 760     t6          = _mm_sub_pd(t6, tK);
 761
 762     t7          = _mm_sub_pd(t7, tB);
 763     t8          = _mm_sub_pd(t8, tD);
 764     t9          = _mm_sub_pd(t9, tF);
 765     t10         = _mm_sub_pd(t10, tH);
 766     t11         = _mm_sub_pd(t11, tJ);
 767     t12         = _mm_sub_pd(t12, tL);
 768
 769     _mm_storeu_pd(ptrA,  t1);
 770     _mm_storeu_pd(ptrA+2, t2);
 771     _mm_storeu_pd(ptrA+4, t3);
 772     _mm_storeu_pd(ptrA+6, t4);
 773     _mm_storeu_pd(ptrA+8, t5);
 774     _mm_storeu_pd(ptrA+10, t6);
 775     _mm_storeu_pd(ptrB,  t7);
 776     _mm_storeu_pd(ptrB+2, t8);
 777     _mm_storeu_pd(ptrB+4, t9);
 778     _mm_storeu_pd(ptrB+6, t10);
 779     _mm_storeu_pd(ptrB+8, t11);
 780     _mm_storeu_pd(ptrB+10, t12);
 781 }
 782 #endif
 783
 784
 785 static gmx_inline void
 786 gmx_mm_update_iforce_1atom_swizzle_pd(__m128d fix1, __m128d fiy1, __m128d fiz1,
 787                                       double * gmx_restrict fptr,
 788                                       double * gmx_restrict fshiftptr)
 789 {
 790     fix1 = _mm_hadd_pd(fix1, fiy1);
 791     fiz1 = _mm_hadd_pd(fiz1, fiz1);
 792
 793     _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr), fix1 ));
 794     _mm_store_sd( fptr+2, _mm_add_sd( _mm_load_sd(fptr+2), fiz1 ));
 795
 796     _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 ));
 797     _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 ));
 798 }
 799
 800 #if defined (_MSC_VER) && defined(_M_IX86)
 801 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 802 #define gmx_mm_update_iforce_3atom_swizzle_pd(fix1, fiy1, fiz1, fix2, fiy2, fiz2, fix3, fiy3, fiz3, \
 803                                               fptr, fshiftptr) \
 804     { \
 805         __m128d _t1, _t2; \
 806         fix1 = _mm_hadd_pd(fix1, fiy1); \
 807         fiz1 = _mm_hadd_pd(fiz1, fix2); \
 808         fiy2 = _mm_hadd_pd(fiy2, fiz2); \
 809         fix3 = _mm_hadd_pd(fix3, fiy3); \
 810         fiz3 = _mm_hadd_pd(fiz3, fiz3); \
 811         _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr), fix1 )); \
 812         _mm_storeu_pd( fptr+2, _mm_add_pd( _mm_loadu_pd(fptr+2), fiz1 )); \
 813         _mm_storeu_pd( fptr+4, _mm_add_pd( _mm_loadu_pd(fptr+4), fiy2 )); \
 814         _mm_storeu_pd( fptr+6, _mm_add_pd( _mm_loadu_pd(fptr+6), fix3 )); \
 815         _mm_store_sd( fptr+8, _mm_add_sd( _mm_load_sd(fptr+8), fiz3 )); \
 816         fix1  = _mm_add_pd(fix1, fix3); \
 817         _t1   = _mm_shuffle_pd(fiz1, fiy2, _MM_SHUFFLE2(0, 1)); \
 818         fix1  = _mm_add_pd(fix1, _t1); \
 819         _t2   = _mm_shuffle_pd(fiy2, fiy2, _MM_SHUFFLE2(1, 1)); \
 820         fiz1  = _mm_add_sd(fiz1, fiz3); \
 821         fiz1  = _mm_add_sd(fiz1, _t2); \
 822         _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 )); \
 823         _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 )); \
 824     }
 825 #else
 826 /* Real function for sane compilers */
 827 static gmx_inline void
 828 gmx_mm_update_iforce_3atom_swizzle_pd(__m128d fix1, __m128d fiy1, __m128d fiz1,
 829                                       __m128d fix2, __m128d fiy2, __m128d fiz2,
 830                                       __m128d fix3, __m128d fiy3, __m128d fiz3,
 831                                       double * gmx_restrict fptr,
 832                                       double * gmx_restrict fshiftptr)
 833 {
 834     __m128d t1, t2;
 835
 836     fix1 = _mm_hadd_pd(fix1, fiy1);
 837     fiz1 = _mm_hadd_pd(fiz1, fix2);
 838     fiy2 = _mm_hadd_pd(fiy2, fiz2);
 839     fix3 = _mm_hadd_pd(fix3, fiy3);
 840     fiz3 = _mm_hadd_pd(fiz3, fiz3);
 841
 842     _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr), fix1 ));
 843     _mm_storeu_pd( fptr+2, _mm_add_pd( _mm_loadu_pd(fptr+2), fiz1 ));
 844     _mm_storeu_pd( fptr+4, _mm_add_pd( _mm_loadu_pd(fptr+4), fiy2 ));
 845     _mm_storeu_pd( fptr+6, _mm_add_pd( _mm_loadu_pd(fptr+6), fix3 ));
 846     _mm_store_sd( fptr+8, _mm_add_sd( _mm_load_sd(fptr+8), fiz3 ));
 847
 848     fix1 = _mm_add_pd(fix1, fix3);
 849     t1   = _mm_shuffle_pd(fiz1, fiy2, _MM_SHUFFLE2(0, 1));
 850     fix1 = _mm_add_pd(fix1, t1); /* x and y sums */
 851
 852     t2   = _mm_shuffle_pd(fiy2, fiy2, _MM_SHUFFLE2(1, 1));
 853     fiz1 = _mm_add_sd(fiz1, fiz3);
 854     fiz1 = _mm_add_sd(fiz1, t2); /* z sum */
 855
 856     _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 ));
 857     _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 ));
 858 }
 859 #endif
 860
 861 #if defined (_MSC_VER) && defined(_M_IX86)
 862 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 863 #define gmx_mm_update_iforce_4atom_swizzle_pd(fix1, fiy1, fiz1, fix2, fiy2, fiz2, fix3, fiy3, fiz3, fix4, fiy4, fiz4, \
 864                                               fptr, fshiftptr) \
 865     { \
 866         __m128d _t1, _t2; \
 867         fix1 = _mm_hadd_pd(fix1, fiy1); \
 868         fiz1 = _mm_hadd_pd(fiz1, fix2); \
 869         fiy2 = _mm_hadd_pd(fiy2, fiz2); \
 870         fix3 = _mm_hadd_pd(fix3, fiy3); \
 871         fiz3 = _mm_hadd_pd(fiz3, fix4); \
 872         fiy4 = _mm_hadd_pd(fiy4, fiz4); \
 873         _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr),       fix1 )); \
 874         _mm_storeu_pd( fptr+2, _mm_add_pd( _mm_loadu_pd(fptr+2),   fiz1 )); \
 875         _mm_storeu_pd( fptr+4, _mm_add_pd( _mm_loadu_pd(fptr+4),   fiy2 )); \
 876         _mm_storeu_pd( fptr+6, _mm_add_pd( _mm_loadu_pd(fptr+6),   fix3 )); \
 877         _mm_storeu_pd( fptr+8, _mm_add_pd( _mm_loadu_pd(fptr+8),   fiz3 )); \
 878         _mm_storeu_pd( fptr+10, _mm_add_pd( _mm_loadu_pd(fptr+10), fiy4 )); \
 879         _t1  = _mm_shuffle_pd(fiz1, fiy2, _MM_SHUFFLE2(0, 1)); \
 880         fix1 = _mm_add_pd(fix1, _t1); \
 881         _t2  = _mm_shuffle_pd(fiz3, fiy4, _MM_SHUFFLE2(0, 1)); \
 882         fix3 = _mm_add_pd(fix3, _t2); \
 883         fix1 = _mm_add_pd(fix1, fix3); \
 884         fiz1 = _mm_add_sd(fiz1, _mm_unpackhi_pd(fiy2, fiy2)); \
 885         fiz3 = _mm_add_sd(fiz3, _mm_unpackhi_pd(fiy4, fiy4)); \
 886         fiz1 = _mm_add_sd(fiz1, fiz3); \
 887         _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 )); \
 888         _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 )); \
 889     }
 890 #else
 891 /* Real function for sane compilers */
 892 static gmx_inline void
 893 gmx_mm_update_iforce_4atom_swizzle_pd(__m128d fix1, __m128d fiy1, __m128d fiz1,
 894                                       __m128d fix2, __m128d fiy2, __m128d fiz2,
 895                                       __m128d fix3, __m128d fiy3, __m128d fiz3,
 896                                       __m128d fix4, __m128d fiy4, __m128d fiz4,
 897                                       double * gmx_restrict fptr,
 898                                       double * gmx_restrict fshiftptr)
 899 {
 900     __m128d t1, t2;
 901
 902     fix1 = _mm_hadd_pd(fix1, fiy1);
 903     fiz1 = _mm_hadd_pd(fiz1, fix2);
 904     fiy2 = _mm_hadd_pd(fiy2, fiz2);
 905     fix3 = _mm_hadd_pd(fix3, fiy3);
 906     fiz3 = _mm_hadd_pd(fiz3, fix4);
 907     fiy4 = _mm_hadd_pd(fiy4, fiz4);
 908
 909     _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr),       fix1 ));
 910     _mm_storeu_pd( fptr+2, _mm_add_pd( _mm_loadu_pd(fptr+2),   fiz1 ));
 911     _mm_storeu_pd( fptr+4, _mm_add_pd( _mm_loadu_pd(fptr+4),   fiy2 ));
 912     _mm_storeu_pd( fptr+6, _mm_add_pd( _mm_loadu_pd(fptr+6),   fix3 ));
 913     _mm_storeu_pd( fptr+8, _mm_add_pd( _mm_loadu_pd(fptr+8),   fiz3 ));
 914     _mm_storeu_pd( fptr+10, _mm_add_pd( _mm_loadu_pd(fptr+10), fiy4 ));
 915
 916     t1   = _mm_shuffle_pd(fiz1, fiy2, _MM_SHUFFLE2(0, 1));
 917     fix1 = _mm_add_pd(fix1, t1);
 918     t2   = _mm_shuffle_pd(fiz3, fiy4, _MM_SHUFFLE2(0, 1));
 919     fix3 = _mm_add_pd(fix3, t2);
 920     fix1 = _mm_add_pd(fix1, fix3); /* x and y sums */
 921
 922     fiz1 = _mm_add_sd(fiz1, _mm_unpackhi_pd(fiy2, fiy2));
 923     fiz3 = _mm_add_sd(fiz3, _mm_unpackhi_pd(fiy4, fiy4));
 924     fiz1 = _mm_add_sd(fiz1, fiz3); /* z sum */
 925
 926     _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 ));
 927     _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 ));
 928 }
 929 #endif
 930
 931
 932 static gmx_inline void
 933 gmx_mm_update_1pot_pd(__m128d pot1, double * gmx_restrict ptrA)
 934 {
 935     pot1 = _mm_hadd_pd(pot1, pot1);
 936     _mm_store_sd(ptrA, _mm_add_sd(pot1, _mm_load_sd(ptrA)));
 937 }
 938
 939 static gmx_inline void
 940 gmx_mm_update_2pot_pd(__m128d pot1, double * gmx_restrict ptrA,
 941                       __m128d pot2, double * gmx_restrict ptrB)
 942 {
 943     pot1 = _mm_hadd_pd(pot1, pot2);
 944     pot2 = _mm_unpackhi_pd(pot1, pot1);
 945
 946     _mm_store_sd(ptrA, _mm_add_sd(pot1, _mm_load_sd(ptrA)));
 947     _mm_store_sd(ptrB, _mm_add_sd(pot2, _mm_load_sd(ptrB)));
 948 }
 949
 950
 951 #endif /* _kernelutil_x86_avx_128_fma_double_h_ */