src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_double/kernelutil_x86_sse2_double.h

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, 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_sse2_double_h_
  36 #define _kernelutil_x86_sse2_double_h_
  37
  38 #include <math.h>
  39
  40
  41 #include <stdio.h>
  42
  43
  44 /* Normal sum of four ymm registers */
  45 #define gmx_mm_sum4_pd(t0, t1, t2, t3)  _mm_add_pd(_mm_add_pd(t0, t1), _mm_add_pd(t2, t3))
  46 #define gmx_mm_extract_epi32(x, imm) _mm_cvtsi128_si32(_mm_srli_si128((x), 4 * (imm)))
  47 #define gmx_mm_castsi128_pd(a) _mm_castsi128_pd(a)
  48
  49 #define GMX_MM_TRANSPOSE2_PD(row0, row1) {           \
  50         __m128d __gmx_t1 = row0;                         \
  51         row0           = _mm_unpacklo_pd(row0, row1);     \
  52         row1           = _mm_unpackhi_pd(__gmx_t1, row1); \
  53 }
  54
  55 static int
  56 gmx_mm_any_lt(__m128d a, __m128d b)
  57 {
  58     return _mm_movemask_pd(_mm_cmplt_pd(a, b));
  59 }
  60
  61 static gmx_inline __m128d
  62 gmx_mm_calc_rsq_pd(__m128d dx, __m128d dy, __m128d dz)
  63 {
  64     return _mm_add_pd( _mm_add_pd( _mm_mul_pd(dx, dx), _mm_mul_pd(dy, dy) ), _mm_mul_pd(dz, dz) );
  65 }
  66
  67
  68 /* Load a double value from 1-2 places, merge into xmm register */
  69 static gmx_inline __m128d
  70 gmx_mm_load_2real_swizzle_pd(const double * gmx_restrict ptrA,
  71                              const double * gmx_restrict ptrB)
  72 {
  73     return _mm_unpacklo_pd(_mm_load_sd(ptrA), _mm_load_sd(ptrB));
  74 }
  75
  76 static gmx_inline __m128d
  77 gmx_mm_load_1real_pd(const double * gmx_restrict ptrA)
  78 {
  79     return _mm_load_sd(ptrA);
  80 }
  81
  82
  83 static gmx_inline void
  84 gmx_mm_store_2real_swizzle_pd(double * gmx_restrict ptrA,
  85                               double * gmx_restrict ptrB,
  86                               __m128d               xmm1)
  87 {
  88     __m128d t2;
  89
  90     t2       = _mm_unpackhi_pd(xmm1, xmm1);
  91     _mm_store_sd(ptrA, xmm1);
  92     _mm_store_sd(ptrB, t2);
  93 }
  94
  95 static gmx_inline void
  96 gmx_mm_store_1real_pd(double * gmx_restrict ptrA, __m128d xmm1)
  97 {
  98     _mm_store_sd(ptrA, xmm1);
  99 }
 100
 101
 102 /* Similar to store, but increments value in memory */
 103 static gmx_inline void
 104 gmx_mm_increment_2real_swizzle_pd(double * gmx_restrict ptrA,
 105                                   double * gmx_restrict ptrB, __m128d xmm1)
 106 {
 107     __m128d t1;
 108
 109     t1   = _mm_unpackhi_pd(xmm1, xmm1);
 110     xmm1 = _mm_add_sd(xmm1, _mm_load_sd(ptrA));
 111     t1   = _mm_add_sd(t1, _mm_load_sd(ptrB));
 112     _mm_store_sd(ptrA, xmm1);
 113     _mm_store_sd(ptrB, t1);
 114 }
 115
 116 static gmx_inline void
 117 gmx_mm_increment_1real_pd(double * gmx_restrict ptrA, __m128d xmm1)
 118 {
 119     __m128d tmp;
 120
 121     tmp = gmx_mm_load_1real_pd(ptrA);
 122     tmp = _mm_add_sd(tmp, xmm1);
 123     gmx_mm_store_1real_pd(ptrA, tmp);
 124 }
 125
 126
 127 static gmx_inline void
 128 gmx_mm_load_2pair_swizzle_pd(const double * gmx_restrict p1,
 129                              const double * gmx_restrict p2,
 130                              __m128d * gmx_restrict      c6,
 131                              __m128d * gmx_restrict      c12)
 132 {
 133     __m128d t1, t2, t3;
 134
 135     t1   = _mm_loadu_pd(p1);
 136     t2   = _mm_loadu_pd(p2);
 137     *c6  = _mm_unpacklo_pd(t1, t2);
 138     *c12 = _mm_unpackhi_pd(t1, t2);
 139 }
 140
 141 static gmx_inline void
 142 gmx_mm_load_1pair_swizzle_pd(const double * gmx_restrict p1,
 143                              __m128d * gmx_restrict      c6,
 144                              __m128d * gmx_restrict      c12)
 145 {
 146     *c6     = _mm_load_sd(p1);
 147     *c12    = _mm_load_sd(p1+1);
 148 }
 149
 150
 151
 152 static gmx_inline void
 153 gmx_mm_load_shift_and_1rvec_broadcast_pd(const double * gmx_restrict xyz_shift,
 154                                          const double * gmx_restrict xyz,
 155                                          __m128d * gmx_restrict      x1,
 156                                          __m128d * gmx_restrict      y1,
 157                                          __m128d * gmx_restrict      z1)
 158 {
 159     __m128d mem_xy, mem_z, mem_sxy, mem_sz;
 160
 161     mem_xy  = _mm_loadu_pd(xyz);
 162     mem_z   = _mm_load_sd(xyz+2);
 163     mem_sxy = _mm_loadu_pd(xyz_shift);
 164     mem_sz  = _mm_load_sd(xyz_shift+2);
 165
 166     mem_xy  = _mm_add_pd(mem_xy, mem_sxy);
 167     mem_z   = _mm_add_pd(mem_z, mem_sz);
 168
 169     *x1  = _mm_shuffle_pd(mem_xy, mem_xy, _MM_SHUFFLE2(0, 0));
 170     *y1  = _mm_shuffle_pd(mem_xy, mem_xy, _MM_SHUFFLE2(1, 1));
 171     *z1  = _mm_shuffle_pd(mem_z, mem_z, _MM_SHUFFLE2(0, 0));
 172 }
 173
 174
 175 static gmx_inline void
 176 gmx_mm_load_shift_and_3rvec_broadcast_pd(const double * gmx_restrict xyz_shift,
 177                                          const double * gmx_restrict xyz,
 178                                          __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 179                                          __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 180                                          __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3)
 181 {
 182     __m128d t1, t2, t3, t4, t5, sxy, sz, szx, syz;
 183
 184     t1  = _mm_loadu_pd(xyz);
 185     t2  = _mm_loadu_pd(xyz+2);
 186     t3  = _mm_loadu_pd(xyz+4);
 187     t4  = _mm_loadu_pd(xyz+6);
 188     t5  = _mm_load_sd(xyz+8);
 189
 190     sxy = _mm_loadu_pd(xyz_shift);
 191     sz  = _mm_load_sd(xyz_shift+2);
 192     szx = _mm_shuffle_pd(sz, sxy, _MM_SHUFFLE2(0, 0));
 193     syz = _mm_shuffle_pd(sxy, sz, _MM_SHUFFLE2(0, 1));
 194
 195     t1  = _mm_add_pd(t1, sxy);
 196     t2  = _mm_add_pd(t2, szx);
 197     t3  = _mm_add_pd(t3, syz);
 198     t4  = _mm_add_pd(t4, sxy);
 199     t5  = _mm_add_sd(t5, sz);
 200
 201     *x1  = _mm_shuffle_pd(t1, t1, _MM_SHUFFLE2(0, 0));
 202     *y1  = _mm_shuffle_pd(t1, t1, _MM_SHUFFLE2(1, 1));
 203     *z1  = _mm_shuffle_pd(t2, t2, _MM_SHUFFLE2(0, 0));
 204     *x2  = _mm_shuffle_pd(t2, t2, _MM_SHUFFLE2(1, 1));
 205     *y2  = _mm_shuffle_pd(t3, t3, _MM_SHUFFLE2(0, 0));
 206     *z2  = _mm_shuffle_pd(t3, t3, _MM_SHUFFLE2(1, 1));
 207     *x3  = _mm_shuffle_pd(t4, t4, _MM_SHUFFLE2(0, 0));
 208     *y3  = _mm_shuffle_pd(t4, t4, _MM_SHUFFLE2(1, 1));
 209     *z3  = _mm_shuffle_pd(t5, t5, _MM_SHUFFLE2(0, 0));
 210 }
 211
 212
 213 static gmx_inline void
 214 gmx_mm_load_shift_and_4rvec_broadcast_pd(const double * gmx_restrict xyz_shift,
 215                                          const double * gmx_restrict xyz,
 216                                          __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 217                                          __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 218                                          __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3,
 219                                          __m128d * gmx_restrict x4, __m128d * gmx_restrict y4, __m128d * gmx_restrict z4)
 220 {
 221     __m128d t1, t2, t3, t4, t5, t6, sxy, sz, szx, syz;
 222
 223     t1  = _mm_loadu_pd(xyz);
 224     t2  = _mm_loadu_pd(xyz+2);
 225     t3  = _mm_loadu_pd(xyz+4);
 226     t4  = _mm_loadu_pd(xyz+6);
 227     t5  = _mm_loadu_pd(xyz+8);
 228     t6  = _mm_loadu_pd(xyz+10);
 229
 230     sxy = _mm_loadu_pd(xyz_shift);
 231     sz  = _mm_load_sd(xyz_shift+2);
 232     szx = _mm_shuffle_pd(sz, sxy, _MM_SHUFFLE2(0, 0));
 233     syz = _mm_shuffle_pd(sxy, sz, _MM_SHUFFLE2(0, 1));
 234
 235     t1  = _mm_add_pd(t1, sxy);
 236     t2  = _mm_add_pd(t2, szx);
 237     t3  = _mm_add_pd(t3, syz);
 238     t4  = _mm_add_pd(t4, sxy);
 239     t5  = _mm_add_pd(t5, szx);
 240     t6  = _mm_add_pd(t6, syz);
 241
 242     *x1  = _mm_shuffle_pd(t1, t1, _MM_SHUFFLE2(0, 0));
 243     *y1  = _mm_shuffle_pd(t1, t1, _MM_SHUFFLE2(1, 1));
 244     *z1  = _mm_shuffle_pd(t2, t2, _MM_SHUFFLE2(0, 0));
 245     *x2  = _mm_shuffle_pd(t2, t2, _MM_SHUFFLE2(1, 1));
 246     *y2  = _mm_shuffle_pd(t3, t3, _MM_SHUFFLE2(0, 0));
 247     *z2  = _mm_shuffle_pd(t3, t3, _MM_SHUFFLE2(1, 1));
 248     *x3  = _mm_shuffle_pd(t4, t4, _MM_SHUFFLE2(0, 0));
 249     *y3  = _mm_shuffle_pd(t4, t4, _MM_SHUFFLE2(1, 1));
 250     *z3  = _mm_shuffle_pd(t5, t5, _MM_SHUFFLE2(0, 0));
 251     *x4  = _mm_shuffle_pd(t5, t5, _MM_SHUFFLE2(1, 1));
 252     *y4  = _mm_shuffle_pd(t6, t6, _MM_SHUFFLE2(0, 0));
 253     *z4  = _mm_shuffle_pd(t6, t6, _MM_SHUFFLE2(1, 1));
 254 }
 255
 256
 257
 258
 259 static gmx_inline void
 260 gmx_mm_load_1rvec_1ptr_swizzle_pd(const double * gmx_restrict p1,
 261                                   __m128d * gmx_restrict x, __m128d * gmx_restrict y, __m128d * gmx_restrict z)
 262 {
 263     *x            = _mm_load_sd(p1);
 264     *y            = _mm_load_sd(p1+1);
 265     *z            = _mm_load_sd(p1+2);
 266 }
 267
 268 static gmx_inline void
 269 gmx_mm_load_3rvec_1ptr_swizzle_pd(const double * gmx_restrict p1,
 270                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 271                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 272                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3)
 273 {
 274     *x1            = _mm_load_sd(p1);
 275     *y1            = _mm_load_sd(p1+1);
 276     *z1            = _mm_load_sd(p1+2);
 277     *x2            = _mm_load_sd(p1+3);
 278     *y2            = _mm_load_sd(p1+4);
 279     *z2            = _mm_load_sd(p1+5);
 280     *x3            = _mm_load_sd(p1+6);
 281     *y3            = _mm_load_sd(p1+7);
 282     *z3            = _mm_load_sd(p1+8);
 283 }
 284
 285 static gmx_inline void
 286 gmx_mm_load_4rvec_1ptr_swizzle_pd(const double * gmx_restrict p1,
 287                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 288                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 289                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3,
 290                                   __m128d * gmx_restrict x4, __m128d * gmx_restrict y4, __m128d * gmx_restrict z4)
 291 {
 292     *x1            = _mm_load_sd(p1);
 293     *y1            = _mm_load_sd(p1+1);
 294     *z1            = _mm_load_sd(p1+2);
 295     *x2            = _mm_load_sd(p1+3);
 296     *y2            = _mm_load_sd(p1+4);
 297     *z2            = _mm_load_sd(p1+5);
 298     *x3            = _mm_load_sd(p1+6);
 299     *y3            = _mm_load_sd(p1+7);
 300     *z3            = _mm_load_sd(p1+8);
 301     *x4            = _mm_load_sd(p1+9);
 302     *y4            = _mm_load_sd(p1+10);
 303     *z4            = _mm_load_sd(p1+11);
 304 }
 305
 306
 307 static gmx_inline void
 308 gmx_mm_load_1rvec_2ptr_swizzle_pd(const double * gmx_restrict ptrA,
 309                                   const double * gmx_restrict ptrB,
 310                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1)
 311 {
 312     __m128d t1, t2, t3, t4;
 313     t1           = _mm_loadu_pd(ptrA);
 314     t2           = _mm_loadu_pd(ptrB);
 315     t3           = _mm_load_sd(ptrA+2);
 316     t4           = _mm_load_sd(ptrB+2);
 317     GMX_MM_TRANSPOSE2_PD(t1, t2);
 318     *x1          = t1;
 319     *y1          = t2;
 320     *z1          = _mm_unpacklo_pd(t3, t4);
 321 }
 322
 323
 324 static gmx_inline void
 325 gmx_mm_load_3rvec_2ptr_swizzle_pd(const double * gmx_restrict ptrA, const double * gmx_restrict ptrB,
 326                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 327                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 328                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3)
 329 {
 330     __m128d t1, t2, t3, t4, t5, t6, t7, t8, t9, t10;
 331     t1           = _mm_loadu_pd(ptrA);
 332     t2           = _mm_loadu_pd(ptrB);
 333     t3           = _mm_loadu_pd(ptrA+2);
 334     t4           = _mm_loadu_pd(ptrB+2);
 335     t5           = _mm_loadu_pd(ptrA+4);
 336     t6           = _mm_loadu_pd(ptrB+4);
 337     t7           = _mm_loadu_pd(ptrA+6);
 338     t8           = _mm_loadu_pd(ptrB+6);
 339     t9           = _mm_load_sd(ptrA+8);
 340     t10          = _mm_load_sd(ptrB+8);
 341     GMX_MM_TRANSPOSE2_PD(t1, t2);
 342     GMX_MM_TRANSPOSE2_PD(t3, t4);
 343     GMX_MM_TRANSPOSE2_PD(t5, t6);
 344     GMX_MM_TRANSPOSE2_PD(t7, t8);
 345     *x1          = t1;
 346     *y1          = t2;
 347     *z1          = t3;
 348     *x2          = t4;
 349     *y2          = t5;
 350     *z2          = t6;
 351     *x3          = t7;
 352     *y3          = t8;
 353     *z3          = _mm_unpacklo_pd(t9, t10);
 354 }
 355
 356
 357 static gmx_inline void
 358 gmx_mm_load_4rvec_2ptr_swizzle_pd(const double * gmx_restrict ptrA, const double * gmx_restrict ptrB,
 359                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 360                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 361                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3,
 362                                   __m128d * gmx_restrict x4, __m128d * gmx_restrict y4, __m128d * gmx_restrict z4)
 363 {
 364     __m128d t1, t2, t3, t4, t5, t6;
 365     t1           = _mm_loadu_pd(ptrA);
 366     t2           = _mm_loadu_pd(ptrB);
 367     t3           = _mm_loadu_pd(ptrA+2);
 368     t4           = _mm_loadu_pd(ptrB+2);
 369     t5           = _mm_loadu_pd(ptrA+4);
 370     t6           = _mm_loadu_pd(ptrB+4);
 371     GMX_MM_TRANSPOSE2_PD(t1, t2);
 372     GMX_MM_TRANSPOSE2_PD(t3, t4);
 373     GMX_MM_TRANSPOSE2_PD(t5, t6);
 374     *x1          = t1;
 375     *y1          = t2;
 376     *z1          = t3;
 377     *x2          = t4;
 378     *y2          = t5;
 379     *z2          = t6;
 380     t1           = _mm_loadu_pd(ptrA+6);
 381     t2           = _mm_loadu_pd(ptrB+6);
 382     t3           = _mm_loadu_pd(ptrA+8);
 383     t4           = _mm_loadu_pd(ptrB+8);
 384     t5           = _mm_loadu_pd(ptrA+10);
 385     t6           = _mm_loadu_pd(ptrB+10);
 386     GMX_MM_TRANSPOSE2_PD(t1, t2);
 387     GMX_MM_TRANSPOSE2_PD(t3, t4);
 388     GMX_MM_TRANSPOSE2_PD(t5, t6);
 389     *x3          = t1;
 390     *y3          = t2;
 391     *z3          = t3;
 392     *x4          = t4;
 393     *y4          = t5;
 394     *z4          = t6;
 395 }
 396
 397
 398 /* Routines to decrement rvec in memory, typically use for j particle force updates */
 399 static gmx_inline void
 400 gmx_mm_decrement_1rvec_1ptr_noswizzle_pd(double * gmx_restrict ptrA,
 401                                          __m128d xy, __m128d z)
 402 {
 403     __m128d t1, t2;
 404
 405     t1 = _mm_loadu_pd(ptrA);
 406     t2 = _mm_load_sd(ptrA+2);
 407
 408     t1 = _mm_sub_pd(t1, xy);
 409     t2 = _mm_sub_sd(t2, z);
 410
 411     _mm_storeu_pd(ptrA, t1);
 412     _mm_store_sd(ptrA+2, t2);
 413 }
 414
 415
 416 static gmx_inline void
 417 gmx_mm_decrement_1rvec_1ptr_swizzle_pd(double * gmx_restrict ptrA,
 418                                        __m128d x1, __m128d y1, __m128d z1)
 419 {
 420     __m128d t1, t2, t3;
 421
 422     t1           = _mm_load_sd(ptrA);
 423     t2           = _mm_load_sd(ptrA+1);
 424     t3           = _mm_load_sd(ptrA+2);
 425
 426     t1           = _mm_sub_sd(t1, x1);
 427     t2           = _mm_sub_sd(t2, y1);
 428     t3           = _mm_sub_sd(t3, z1);
 429     _mm_store_sd(ptrA, t1);
 430     _mm_store_sd(ptrA+1, t2);
 431     _mm_store_sd(ptrA+2, t3);
 432 }
 433
 434
 435 #if defined (_MSC_VER) && defined(_M_IX86)
 436 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 437 #define gmx_mm_decrement_3rvec_1ptr_swizzle_pd(ptrA, _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3) \
 438     { \
 439         __m128d _t1, _t2, _t3, _t4, _t5; \
 440         _t1          = _mm_loadu_pd(ptrA); \
 441         _t2          = _mm_loadu_pd(ptrA+2); \
 442         _t3          = _mm_loadu_pd(ptrA+4); \
 443         _t4          = _mm_loadu_pd(ptrA+6); \
 444         _t5          = _mm_load_sd(ptrA+8); \
 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         _t1          = _mm_sub_pd(_t1, _x1); \
 450         _t2          = _mm_sub_pd(_t2, _z1); \
 451         _t3          = _mm_sub_pd(_t3, _y2); \
 452         _t4          = _mm_sub_pd(_t4, _x3); \
 453         _t5          = _mm_sub_sd(_t5, _z3); \
 454         _mm_storeu_pd(ptrA, _t1); \
 455         _mm_storeu_pd(ptrA+2, _t2); \
 456         _mm_storeu_pd(ptrA+4, _t3); \
 457         _mm_storeu_pd(ptrA+6, _t4); \
 458         _mm_store_sd(ptrA+8, _t5); \
 459     }
 460 #else
 461 /* Real function for sane compilers */
 462 static gmx_inline void
 463 gmx_mm_decrement_3rvec_1ptr_swizzle_pd(double * gmx_restrict ptrA,
 464                                        __m128d x1, __m128d y1, __m128d z1,
 465                                        __m128d x2, __m128d y2, __m128d z2,
 466                                        __m128d x3, __m128d y3, __m128d z3)
 467 {
 468     __m128d t1, t2, t3, t4, t5;
 469
 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_load_sd(ptrA+8);
 475
 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     /* nothing to be done for z3 */
 481
 482     t1          = _mm_sub_pd(t1, x1);
 483     t2          = _mm_sub_pd(t2, z1);
 484     t3          = _mm_sub_pd(t3, y2);
 485     t4          = _mm_sub_pd(t4, x3);
 486     t5          = _mm_sub_sd(t5, z3);
 487     _mm_storeu_pd(ptrA, t1);
 488     _mm_storeu_pd(ptrA+2, t2);
 489     _mm_storeu_pd(ptrA+4, t3);
 490     _mm_storeu_pd(ptrA+6, t4);
 491     _mm_store_sd(ptrA+8, t5);
 492 }
 493 #endif
 494
 495
 496 #if defined (_MSC_VER) && defined(_M_IX86)
 497 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 498 #define gmx_mm_decrement_4rvec_1ptr_swizzle_pd(ptrA, _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3, _x4, _y4, _z4) \
 499     { \
 500         __m128d _t1, _t2, _t3, _t4, _t5, _t6; \
 501         _t1          = _mm_loadu_pd(ptrA); \
 502         _t2          = _mm_loadu_pd(ptrA+2); \
 503         _t3          = _mm_loadu_pd(ptrA+4); \
 504         _t4          = _mm_loadu_pd(ptrA+6); \
 505         _t5          = _mm_loadu_pd(ptrA+8); \
 506         _t6          = _mm_loadu_pd(ptrA+10); \
 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         _mm_storeu_pd(ptrA,    _mm_sub_pd( _t1, _x1 )); \
 514         _mm_storeu_pd(ptrA+2,  _mm_sub_pd( _t2, _z1 )); \
 515         _mm_storeu_pd(ptrA+4,  _mm_sub_pd( _t3, _y2 )); \
 516         _mm_storeu_pd(ptrA+6,  _mm_sub_pd( _t4, _x3 )); \
 517         _mm_storeu_pd(ptrA+8,  _mm_sub_pd( _t5, _z3 )); \
 518         _mm_storeu_pd(ptrA+10, _mm_sub_pd( _t6, _y4 )); \
 519     }
 520 #else
 521 /* Real function for sane compilers */
 522 static gmx_inline void
 523 gmx_mm_decrement_4rvec_1ptr_swizzle_pd(double * gmx_restrict ptrA,
 524                                        __m128d x1, __m128d y1, __m128d z1,
 525                                        __m128d x2, __m128d y2, __m128d z2,
 526                                        __m128d x3, __m128d y3, __m128d z3,
 527                                        __m128d x4, __m128d y4, __m128d z4)
 528 {
 529     __m128d t1, t2, t3, t4, t5, t6;
 530
 531     t1          = _mm_loadu_pd(ptrA);
 532     t2          = _mm_loadu_pd(ptrA+2);
 533     t3          = _mm_loadu_pd(ptrA+4);
 534     t4          = _mm_loadu_pd(ptrA+6);
 535     t5          = _mm_loadu_pd(ptrA+8);
 536     t6          = _mm_loadu_pd(ptrA+10);
 537
 538     x1          = _mm_unpacklo_pd(x1, y1);
 539     z1          = _mm_unpacklo_pd(z1, x2);
 540     y2          = _mm_unpacklo_pd(y2, z2);
 541     x3          = _mm_unpacklo_pd(x3, y3);
 542     z3          = _mm_unpacklo_pd(z3, x4);
 543     y4          = _mm_unpacklo_pd(y4, z4);
 544
 545     _mm_storeu_pd(ptrA,    _mm_sub_pd( t1, x1 ));
 546     _mm_storeu_pd(ptrA+2,  _mm_sub_pd( t2, z1 ));
 547     _mm_storeu_pd(ptrA+4,  _mm_sub_pd( t3, y2 ));
 548     _mm_storeu_pd(ptrA+6,  _mm_sub_pd( t4, x3 ));
 549     _mm_storeu_pd(ptrA+8,  _mm_sub_pd( t5, z3 ));
 550     _mm_storeu_pd(ptrA+10, _mm_sub_pd( t6, y4 ));
 551 }
 552 #endif
 553
 554
 555 static gmx_inline void
 556 gmx_mm_decrement_1rvec_2ptr_swizzle_pd(double * gmx_restrict ptrA, double * gmx_restrict ptrB,
 557                                        __m128d x1, __m128d y1, __m128d z1)
 558 {
 559     __m128d t1, t2, t3, t4, t5, t6, t7;
 560
 561     t1          = _mm_loadu_pd(ptrA);
 562     t2          = _mm_load_sd(ptrA+2);
 563     t3          = _mm_loadu_pd(ptrB);
 564     t4          = _mm_load_sd(ptrB+2);
 565
 566     t5          = _mm_unpacklo_pd(x1, y1);
 567     t6          = _mm_unpackhi_pd(x1, y1);
 568     t7          = _mm_unpackhi_pd(z1, z1);
 569
 570     t1          = _mm_sub_pd(t1, t5);
 571     t2          = _mm_sub_sd(t2, z1);
 572
 573     t3          = _mm_sub_pd(t3, t6);
 574     t4          = _mm_sub_sd(t4, t7);
 575
 576     _mm_storeu_pd(ptrA, t1);
 577     _mm_store_sd(ptrA+2, t2);
 578     _mm_storeu_pd(ptrB, t3);
 579     _mm_store_sd(ptrB+2, t4);
 580 }
 581
 582 #if defined (_MSC_VER) && defined(_M_IX86)
 583 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 584 #define gmx_mm_decrement_3rvec_2ptr_swizzle_pd(ptrA, ptrB, _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3) \
 585     { \
 586         __m128d _t1, _t2, _t3, _t4, _t5, _t6, _t7, _t8, _t9, _t10; \
 587         __m128d _tA, _tB, _tC, _tD, _tE, _tF, _tG, _tH, _tI; \
 588         _t1          = _mm_loadu_pd(ptrA); \
 589         _t2          = _mm_loadu_pd(ptrA+2); \
 590         _t3          = _mm_loadu_pd(ptrA+4); \
 591         _t4          = _mm_loadu_pd(ptrA+6); \
 592         _t5          = _mm_load_sd(ptrA+8); \
 593         _t6          = _mm_loadu_pd(ptrB); \
 594         _t7          = _mm_loadu_pd(ptrB+2); \
 595         _t8          = _mm_loadu_pd(ptrB+4); \
 596         _t9          = _mm_loadu_pd(ptrB+6); \
 597         _t10         = _mm_load_sd(ptrB+8); \
 598         _tA          = _mm_unpacklo_pd(_x1, _y1); \
 599         _tB          = _mm_unpackhi_pd(_x1, _y1); \
 600         _tC          = _mm_unpacklo_pd(_z1, _x2); \
 601         _tD          = _mm_unpackhi_pd(_z1, _x2); \
 602         _tE          = _mm_unpacklo_pd(_y2, _z2); \
 603         _tF          = _mm_unpackhi_pd(_y2, _z2); \
 604         _tG          = _mm_unpacklo_pd(_x3, _y3); \
 605         _tH          = _mm_unpackhi_pd(_x3, _y3); \
 606         _tI          = _mm_unpackhi_pd(_z3, _z3); \
 607         _t1          = _mm_sub_pd(_t1, _tA); \
 608         _t2          = _mm_sub_pd(_t2, _tC); \
 609         _t3          = _mm_sub_pd(_t3, _tE); \
 610         _t4          = _mm_sub_pd(_t4, _tG); \
 611         _t5          = _mm_sub_sd(_t5, _z3); \
 612         _t6          = _mm_sub_pd(_t6, _tB); \
 613         _t7          = _mm_sub_pd(_t7, _tD); \
 614         _t8          = _mm_sub_pd(_t8, _tF); \
 615         _t9          = _mm_sub_pd(_t9, _tH); \
 616         _t10         = _mm_sub_sd(_t10, _tI); \
 617         _mm_storeu_pd(ptrA, _t1); \
 618         _mm_storeu_pd(ptrA+2, _t2); \
 619         _mm_storeu_pd(ptrA+4, _t3); \
 620         _mm_storeu_pd(ptrA+6, _t4); \
 621         _mm_store_sd(ptrA+8, _t5); \
 622         _mm_storeu_pd(ptrB, _t6); \
 623         _mm_storeu_pd(ptrB+2, _t7); \
 624         _mm_storeu_pd(ptrB+4, _t8); \
 625         _mm_storeu_pd(ptrB+6, _t9); \
 626         _mm_store_sd(ptrB+8, _t10); \
 627     }
 628 #else
 629 /* Real function for sane compilers */
 630 static gmx_inline void
 631 gmx_mm_decrement_3rvec_2ptr_swizzle_pd(double * gmx_restrict ptrA, double * gmx_restrict ptrB,
 632                                        __m128d x1, __m128d y1, __m128d z1,
 633                                        __m128d x2, __m128d y2, __m128d z2,
 634                                        __m128d x3, __m128d y3, __m128d z3)
 635 {
 636     __m128d t1, t2, t3, t4, t5, t6, t7, t8, t9, t10;
 637     __m128d tA, tB, tC, tD, tE, tF, tG, tH, tI;
 638
 639     t1          = _mm_loadu_pd(ptrA);
 640     t2          = _mm_loadu_pd(ptrA+2);
 641     t3          = _mm_loadu_pd(ptrA+4);
 642     t4          = _mm_loadu_pd(ptrA+6);
 643     t5          = _mm_load_sd(ptrA+8);
 644     t6          = _mm_loadu_pd(ptrB);
 645     t7          = _mm_loadu_pd(ptrB+2);
 646     t8          = _mm_loadu_pd(ptrB+4);
 647     t9          = _mm_loadu_pd(ptrB+6);
 648     t10         = _mm_load_sd(ptrB+8);
 649
 650     tA          = _mm_unpacklo_pd(x1, y1);
 651     tB          = _mm_unpackhi_pd(x1, y1);
 652     tC          = _mm_unpacklo_pd(z1, x2);
 653     tD          = _mm_unpackhi_pd(z1, x2);
 654     tE          = _mm_unpacklo_pd(y2, z2);
 655     tF          = _mm_unpackhi_pd(y2, z2);
 656     tG          = _mm_unpacklo_pd(x3, y3);
 657     tH          = _mm_unpackhi_pd(x3, y3);
 658     tI          = _mm_unpackhi_pd(z3, z3);
 659
 660     t1          = _mm_sub_pd(t1, tA);
 661     t2          = _mm_sub_pd(t2, tC);
 662     t3          = _mm_sub_pd(t3, tE);
 663     t4          = _mm_sub_pd(t4, tG);
 664     t5          = _mm_sub_sd(t5, z3);
 665
 666     t6          = _mm_sub_pd(t6, tB);
 667     t7          = _mm_sub_pd(t7, tD);
 668     t8          = _mm_sub_pd(t8, tF);
 669     t9          = _mm_sub_pd(t9, tH);
 670     t10         = _mm_sub_sd(t10, tI);
 671
 672     _mm_storeu_pd(ptrA, t1);
 673     _mm_storeu_pd(ptrA+2, t2);
 674     _mm_storeu_pd(ptrA+4, t3);
 675     _mm_storeu_pd(ptrA+6, t4);
 676     _mm_store_sd(ptrA+8, t5);
 677     _mm_storeu_pd(ptrB, t6);
 678     _mm_storeu_pd(ptrB+2, t7);
 679     _mm_storeu_pd(ptrB+4, t8);
 680     _mm_storeu_pd(ptrB+6, t9);
 681     _mm_store_sd(ptrB+8, t10);
 682 }
 683 #endif
 684
 685
 686 #if defined (_MSC_VER) && defined(_M_IX86)
 687 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 688 #define gmx_mm_decrement_4rvec_2ptr_swizzle_pd(ptrA, ptrB, _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3, _x4, _y4, _z4) \
 689     { \
 690         __m128d _t1, _t2, _t3, _t4, _t5, _t6, _t7, _t8, _t9, _t10, _t11, _t12; \
 691         __m128d _tA, _tB, _tC, _tD, _tE, _tF, _tG, _tH, _tI, _tJ, _tK, _tL; \
 692         _t1          = _mm_loadu_pd(ptrA); \
 693         _t2          = _mm_loadu_pd(ptrA+2); \
 694         _t3          = _mm_loadu_pd(ptrA+4); \
 695         _t4          = _mm_loadu_pd(ptrA+6); \
 696         _t5          = _mm_loadu_pd(ptrA+8); \
 697         _t6          = _mm_loadu_pd(ptrA+10); \
 698         _t7          = _mm_loadu_pd(ptrB); \
 699         _t8          = _mm_loadu_pd(ptrB+2); \
 700         _t9          = _mm_loadu_pd(ptrB+4); \
 701         _t10         = _mm_loadu_pd(ptrB+6); \
 702         _t11         = _mm_loadu_pd(ptrB+8); \
 703         _t12         = _mm_loadu_pd(ptrB+10); \
 704         _tA          = _mm_unpacklo_pd(_x1, _y1); \
 705         _tB          = _mm_unpackhi_pd(_x1, _y1); \
 706         _tC          = _mm_unpacklo_pd(_z1, _x2); \
 707         _tD          = _mm_unpackhi_pd(_z1, _x2); \
 708         _tE          = _mm_unpacklo_pd(_y2, _z2); \
 709         _tF          = _mm_unpackhi_pd(_y2, _z2); \
 710         _tG          = _mm_unpacklo_pd(_x3, _y3); \
 711         _tH          = _mm_unpackhi_pd(_x3, _y3); \
 712         _tI          = _mm_unpacklo_pd(_z3, _x4); \
 713         _tJ          = _mm_unpackhi_pd(_z3, _x4); \
 714         _tK          = _mm_unpacklo_pd(_y4, _z4); \
 715         _tL          = _mm_unpackhi_pd(_y4, _z4); \
 716         _t1          = _mm_sub_pd(_t1, _tA); \
 717         _t2          = _mm_sub_pd(_t2, _tC); \
 718         _t3          = _mm_sub_pd(_t3, _tE); \
 719         _t4          = _mm_sub_pd(_t4, _tG); \
 720         _t5          = _mm_sub_pd(_t5, _tI); \
 721         _t6          = _mm_sub_pd(_t6, _tK); \
 722         _t7          = _mm_sub_pd(_t7, _tB); \
 723         _t8          = _mm_sub_pd(_t8, _tD); \
 724         _t9          = _mm_sub_pd(_t9, _tF); \
 725         _t10         = _mm_sub_pd(_t10, _tH); \
 726         _t11         = _mm_sub_pd(_t11, _tJ); \
 727         _t12         = _mm_sub_pd(_t12, _tL); \
 728         _mm_storeu_pd(ptrA,  _t1); \
 729         _mm_storeu_pd(ptrA+2, _t2); \
 730         _mm_storeu_pd(ptrA+4, _t3); \
 731         _mm_storeu_pd(ptrA+6, _t4); \
 732         _mm_storeu_pd(ptrA+8, _t5); \
 733         _mm_storeu_pd(ptrA+10, _t6); \
 734         _mm_storeu_pd(ptrB,  _t7); \
 735         _mm_storeu_pd(ptrB+2, _t8); \
 736         _mm_storeu_pd(ptrB+4, _t9); \
 737         _mm_storeu_pd(ptrB+6, _t10); \
 738         _mm_storeu_pd(ptrB+8, _t11); \
 739         _mm_storeu_pd(ptrB+10, _t12); \
 740     }
 741 #else
 742 /* Real function for sane compilers */
 743 static gmx_inline void
 744 gmx_mm_decrement_4rvec_2ptr_swizzle_pd(double * gmx_restrict ptrA, double * gmx_restrict ptrB,
 745                                        __m128d x1, __m128d y1, __m128d z1,
 746                                        __m128d x2, __m128d y2, __m128d z2,
 747                                        __m128d x3, __m128d y3, __m128d z3,
 748                                        __m128d x4, __m128d y4, __m128d z4)
 749 {
 750     __m128d t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12;
 751     __m128d tA, tB, tC, tD, tE, tF, tG, tH, tI, tJ, tK, tL;
 752
 753     t1          = _mm_loadu_pd(ptrA);
 754     t2          = _mm_loadu_pd(ptrA+2);
 755     t3          = _mm_loadu_pd(ptrA+4);
 756     t4          = _mm_loadu_pd(ptrA+6);
 757     t5          = _mm_loadu_pd(ptrA+8);
 758     t6          = _mm_loadu_pd(ptrA+10);
 759     t7          = _mm_loadu_pd(ptrB);
 760     t8          = _mm_loadu_pd(ptrB+2);
 761     t9          = _mm_loadu_pd(ptrB+4);
 762     t10         = _mm_loadu_pd(ptrB+6);
 763     t11         = _mm_loadu_pd(ptrB+8);
 764     t12         = _mm_loadu_pd(ptrB+10);
 765
 766     tA          = _mm_unpacklo_pd(x1, y1);
 767     tB          = _mm_unpackhi_pd(x1, y1);
 768     tC          = _mm_unpacklo_pd(z1, x2);
 769     tD          = _mm_unpackhi_pd(z1, x2);
 770     tE          = _mm_unpacklo_pd(y2, z2);
 771     tF          = _mm_unpackhi_pd(y2, z2);
 772     tG          = _mm_unpacklo_pd(x3, y3);
 773     tH          = _mm_unpackhi_pd(x3, y3);
 774     tI          = _mm_unpacklo_pd(z3, x4);
 775     tJ          = _mm_unpackhi_pd(z3, x4);
 776     tK          = _mm_unpacklo_pd(y4, z4);
 777     tL          = _mm_unpackhi_pd(y4, z4);
 778
 779     t1          = _mm_sub_pd(t1, tA);
 780     t2          = _mm_sub_pd(t2, tC);
 781     t3          = _mm_sub_pd(t3, tE);
 782     t4          = _mm_sub_pd(t4, tG);
 783     t5          = _mm_sub_pd(t5, tI);
 784     t6          = _mm_sub_pd(t6, tK);
 785
 786     t7          = _mm_sub_pd(t7, tB);
 787     t8          = _mm_sub_pd(t8, tD);
 788     t9          = _mm_sub_pd(t9, tF);
 789     t10         = _mm_sub_pd(t10, tH);
 790     t11         = _mm_sub_pd(t11, tJ);
 791     t12         = _mm_sub_pd(t12, tL);
 792
 793     _mm_storeu_pd(ptrA,  t1);
 794     _mm_storeu_pd(ptrA+2, t2);
 795     _mm_storeu_pd(ptrA+4, t3);
 796     _mm_storeu_pd(ptrA+6, t4);
 797     _mm_storeu_pd(ptrA+8, t5);
 798     _mm_storeu_pd(ptrA+10, t6);
 799     _mm_storeu_pd(ptrB,  t7);
 800     _mm_storeu_pd(ptrB+2, t8);
 801     _mm_storeu_pd(ptrB+4, t9);
 802     _mm_storeu_pd(ptrB+6, t10);
 803     _mm_storeu_pd(ptrB+8, t11);
 804     _mm_storeu_pd(ptrB+10, t12);
 805 }
 806 #endif
 807
 808
 809
 810
 811
 812 static gmx_inline void
 813 gmx_mm_update_iforce_1atom_swizzle_pd(__m128d fix1, __m128d fiy1, __m128d fiz1,
 814                                       double * gmx_restrict fptr,
 815                                       double * gmx_restrict fshiftptr)
 816 {
 817     __m128d t1, t2, t3;
 818
 819     /* transpose data */
 820     t1   = fix1;
 821     fix1 = _mm_unpacklo_pd(fix1, fiy1); /* y0 x0 */
 822     fiy1 = _mm_unpackhi_pd(t1, fiy1);   /* y1 x1 */
 823
 824     fix1 = _mm_add_pd(fix1, fiy1);
 825     fiz1 = _mm_add_sd( fiz1, _mm_unpackhi_pd(fiz1, fiz1 ));
 826
 827     _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr), fix1 ));
 828     _mm_store_sd( fptr+2, _mm_add_sd( _mm_load_sd(fptr+2), fiz1 ));
 829
 830     _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 ));
 831     _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 ));
 832 }
 833
 834 #if defined (_MSC_VER) && defined(_M_IX86)
 835 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 836 #define gmx_mm_update_iforce_3atom_swizzle_pd(fix1, fiy1, fiz1, fix2, fiy2, fiz2, fix3, fiy3, fiz3, \
 837                                               fptr, fshiftptr) \
 838     { \
 839         __m128d _t1, _t2; \
 840         GMX_MM_TRANSPOSE2_PD(fix1, fiy1); \
 841         GMX_MM_TRANSPOSE2_PD(fiz1, fix2); \
 842         GMX_MM_TRANSPOSE2_PD(fiy2, fiz2); \
 843         _t1  = fix3; \
 844         fix3 = _mm_unpacklo_pd(fix3, fiy3); \
 845         fiy3 = _mm_unpackhi_pd(_t1, fiy3); \
 846         fix1 = _mm_add_pd(fix1, fiy1); \
 847         fiz1 = _mm_add_pd(fiz1, fix2); \
 848         fiy2 = _mm_add_pd(fiy2, fiz2); \
 849         fix3 = _mm_add_pd(fix3, fiy3); \
 850         fiz3 = _mm_add_sd( fiz3, _mm_unpackhi_pd(fiz3, fiz3)); \
 851         _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr), fix1 )); \
 852         _mm_storeu_pd( fptr+2, _mm_add_pd( _mm_loadu_pd(fptr+2), fiz1 )); \
 853         _mm_storeu_pd( fptr+4, _mm_add_pd( _mm_loadu_pd(fptr+4), fiy2 )); \
 854         _mm_storeu_pd( fptr+6, _mm_add_pd( _mm_loadu_pd(fptr+6), fix3 )); \
 855         _mm_store_sd( fptr+8, _mm_add_sd( _mm_load_sd(fptr+8), fiz3 )); \
 856         fix1  = _mm_add_pd(fix1, fix3); \
 857         _t1   = _mm_shuffle_pd(fiz1, fiy2, _MM_SHUFFLE2(0, 1)); \
 858         fix1  = _mm_add_pd(fix1, _t1); \
 859         _t2   = _mm_shuffle_pd(fiy2, fiy2, _MM_SHUFFLE2(1, 1)); \
 860         fiz1  = _mm_add_sd(fiz1, fiz3); \
 861         fiz1  = _mm_add_sd(fiz1, _t2); \
 862         _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 )); \
 863         _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 )); \
 864     }
 865 #else
 866 /* Real function for sane compilers */
 867 static gmx_inline void
 868 gmx_mm_update_iforce_3atom_swizzle_pd(__m128d fix1, __m128d fiy1, __m128d fiz1,
 869                                       __m128d fix2, __m128d fiy2, __m128d fiz2,
 870                                       __m128d fix3, __m128d fiy3, __m128d fiz3,
 871                                       double * gmx_restrict fptr,
 872                                       double * gmx_restrict fshiftptr)
 873 {
 874     __m128d t1, t2;
 875
 876     /* transpose data */
 877     GMX_MM_TRANSPOSE2_PD(fix1, fiy1);
 878     GMX_MM_TRANSPOSE2_PD(fiz1, fix2);
 879     GMX_MM_TRANSPOSE2_PD(fiy2, fiz2);
 880     t1   = fix3;
 881     fix3 = _mm_unpacklo_pd(fix3, fiy3); /* y0 x0 */
 882     fiy3 = _mm_unpackhi_pd(t1, fiy3);   /* y1 x1 */
 883
 884     fix1 = _mm_add_pd(fix1, fiy1);
 885     fiz1 = _mm_add_pd(fiz1, fix2);
 886     fiy2 = _mm_add_pd(fiy2, fiz2);
 887
 888     fix3 = _mm_add_pd(fix3, fiy3);
 889     fiz3 = _mm_add_sd( fiz3, _mm_unpackhi_pd(fiz3, fiz3));
 890
 891     _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr), fix1 ));
 892     _mm_storeu_pd( fptr+2, _mm_add_pd( _mm_loadu_pd(fptr+2), fiz1 ));
 893     _mm_storeu_pd( fptr+4, _mm_add_pd( _mm_loadu_pd(fptr+4), fiy2 ));
 894     _mm_storeu_pd( fptr+6, _mm_add_pd( _mm_loadu_pd(fptr+6), fix3 ));
 895     _mm_store_sd( fptr+8, _mm_add_sd( _mm_load_sd(fptr+8), fiz3 ));
 896
 897     fix1 = _mm_add_pd(fix1, fix3);
 898     t1   = _mm_shuffle_pd(fiz1, fiy2, _MM_SHUFFLE2(0, 1));
 899     fix1 = _mm_add_pd(fix1, t1); /* x and y sums */
 900
 901     t2   = _mm_shuffle_pd(fiy2, fiy2, _MM_SHUFFLE2(1, 1));
 902     fiz1 = _mm_add_sd(fiz1, fiz3);
 903     fiz1 = _mm_add_sd(fiz1, t2); /* z sum */
 904
 905     _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 ));
 906     _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 ));
 907 }
 908 #endif
 909
 910
 911 #if defined (_MSC_VER) && defined(_M_IX86)
 912 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 913 #define gmx_mm_update_iforce_4atom_swizzle_pd(fix1, fiy1, fiz1, fix2, fiy2, fiz2, fix3, fiy3, fiz3, fix4, fiy4, fiz4, \
 914                                               fptr, fshiftptr) \
 915     { \
 916         __m128d _t1, _t2; \
 917         GMX_MM_TRANSPOSE2_PD(fix1, fiy1); \
 918         GMX_MM_TRANSPOSE2_PD(fiz1, fix2); \
 919         GMX_MM_TRANSPOSE2_PD(fiy2, fiz2); \
 920         GMX_MM_TRANSPOSE2_PD(fix3, fiy3); \
 921         GMX_MM_TRANSPOSE2_PD(fiz3, fix4); \
 922         GMX_MM_TRANSPOSE2_PD(fiy4, fiz4); \
 923         fix1 = _mm_add_pd(fix1, fiy1); \
 924         fiz1 = _mm_add_pd(fiz1, fix2); \
 925         fiy2 = _mm_add_pd(fiy2, fiz2); \
 926         fix3 = _mm_add_pd(fix3, fiy3); \
 927         fiz3 = _mm_add_pd(fiz3, fix4); \
 928         fiy4 = _mm_add_pd(fiy4, fiz4); \
 929         _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr),       fix1 )); \
 930         _mm_storeu_pd( fptr+2, _mm_add_pd( _mm_loadu_pd(fptr+2),   fiz1 )); \
 931         _mm_storeu_pd( fptr+4, _mm_add_pd( _mm_loadu_pd(fptr+4),   fiy2 )); \
 932         _mm_storeu_pd( fptr+6, _mm_add_pd( _mm_loadu_pd(fptr+6),   fix3 )); \
 933         _mm_storeu_pd( fptr+8, _mm_add_pd( _mm_loadu_pd(fptr+8),   fiz3 )); \
 934         _mm_storeu_pd( fptr+10, _mm_add_pd( _mm_loadu_pd(fptr+10), fiy4 )); \
 935         _t1  = _mm_shuffle_pd(fiz1, fiy2, _MM_SHUFFLE2(0, 1)); \
 936         fix1 = _mm_add_pd(fix1, _t1); \
 937         _t2  = _mm_shuffle_pd(fiz3, fiy4, _MM_SHUFFLE2(0, 1)); \
 938         fix3 = _mm_add_pd(fix3, _t2); \
 939         fix1 = _mm_add_pd(fix1, fix3); \
 940         fiz1 = _mm_add_sd(fiz1, _mm_unpackhi_pd(fiy2, fiy2)); \
 941         fiz3 = _mm_add_sd(fiz3, _mm_unpackhi_pd(fiy4, fiy4)); \
 942         fiz1 = _mm_add_sd(fiz1, fiz3); \
 943         _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 )); \
 944         _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 )); \
 945     }
 946 #else
 947 /* Real function for sane compilers */
 948 static gmx_inline void
 949 gmx_mm_update_iforce_4atom_swizzle_pd(__m128d fix1, __m128d fiy1, __m128d fiz1,
 950                                       __m128d fix2, __m128d fiy2, __m128d fiz2,
 951                                       __m128d fix3, __m128d fiy3, __m128d fiz3,
 952                                       __m128d fix4, __m128d fiy4, __m128d fiz4,
 953                                       double * gmx_restrict fptr,
 954                                       double * gmx_restrict fshiftptr)
 955 {
 956     __m128d t1, t2;
 957
 958     /* transpose data */
 959     GMX_MM_TRANSPOSE2_PD(fix1, fiy1);
 960     GMX_MM_TRANSPOSE2_PD(fiz1, fix2);
 961     GMX_MM_TRANSPOSE2_PD(fiy2, fiz2);
 962     GMX_MM_TRANSPOSE2_PD(fix3, fiy3);
 963     GMX_MM_TRANSPOSE2_PD(fiz3, fix4);
 964     GMX_MM_TRANSPOSE2_PD(fiy4, fiz4);
 965
 966     fix1 = _mm_add_pd(fix1, fiy1);
 967     fiz1 = _mm_add_pd(fiz1, fix2);
 968     fiy2 = _mm_add_pd(fiy2, fiz2);
 969     fix3 = _mm_add_pd(fix3, fiy3);
 970     fiz3 = _mm_add_pd(fiz3, fix4);
 971     fiy4 = _mm_add_pd(fiy4, fiz4);
 972
 973     _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr),       fix1 ));
 974     _mm_storeu_pd( fptr+2, _mm_add_pd( _mm_loadu_pd(fptr+2),   fiz1 ));
 975     _mm_storeu_pd( fptr+4, _mm_add_pd( _mm_loadu_pd(fptr+4),   fiy2 ));
 976     _mm_storeu_pd( fptr+6, _mm_add_pd( _mm_loadu_pd(fptr+6),   fix3 ));
 977     _mm_storeu_pd( fptr+8, _mm_add_pd( _mm_loadu_pd(fptr+8),   fiz3 ));
 978     _mm_storeu_pd( fptr+10, _mm_add_pd( _mm_loadu_pd(fptr+10), fiy4 ));
 979
 980     t1   = _mm_shuffle_pd(fiz1, fiy2, _MM_SHUFFLE2(0, 1));
 981     fix1 = _mm_add_pd(fix1, t1);
 982     t2   = _mm_shuffle_pd(fiz3, fiy4, _MM_SHUFFLE2(0, 1));
 983     fix3 = _mm_add_pd(fix3, t2);
 984     fix1 = _mm_add_pd(fix1, fix3); /* x and y sums */
 985
 986     fiz1 = _mm_add_sd(fiz1, _mm_unpackhi_pd(fiy2, fiy2));
 987     fiz3 = _mm_add_sd(fiz3, _mm_unpackhi_pd(fiy4, fiy4));
 988     fiz1 = _mm_add_sd(fiz1, fiz3); /* z sum */
 989
 990     _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 ));
 991     _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 ));
 992 }
 993 #endif
 994
 995
 996 static gmx_inline void
 997 gmx_mm_update_1pot_pd(__m128d pot1, double * gmx_restrict ptrA)
 998 {
 999     pot1 = _mm_add_pd(pot1, _mm_unpackhi_pd(pot1, pot1));
1000     _mm_store_sd(ptrA, _mm_add_sd(pot1, _mm_load_sd(ptrA)));
1001 }
1002
1003 static gmx_inline void
1004 gmx_mm_update_2pot_pd(__m128d pot1, double * gmx_restrict ptrA,
1005                       __m128d pot2, double * gmx_restrict ptrB)
1006 {
1007     GMX_MM_TRANSPOSE2_PD(pot1, pot2);
1008     pot1 = _mm_add_pd(pot1, pot2);
1009     pot2 = _mm_unpackhi_pd(pot1, pot1);
1010
1011     _mm_store_sd(ptrA, _mm_add_sd(pot1, _mm_load_sd(ptrA)));
1012     _mm_store_sd(ptrB, _mm_add_sd(pot2, _mm_load_sd(ptrB)));
1013 }
1014
1015
1016 #endif /* _kernelutil_x86_sse2_double_h_ */