src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_double/kernelutil_x86_sse4_1_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_sse4_1_double_h_
  36 #define _kernelutil_x86_sse4_1_double_h_
  37
  38 #include <math.h>
  39
  40 #include <stdio.h>
  41
  42 #define gmx_mm_extract_epi32(x, imm) _mm_extract_epi32((x), (imm))
  43 #define gmx_mm_castsi128_pd(a) _mm_castsi128_pd(a)
  44
  45 #define GMX_MM_TRANSPOSE2_PD(row0, row1) {           \
  46         __m128d __gmx_t1 = row0;                         \
  47         row0           = _mm_unpacklo_pd(row0, row1);     \
  48         row1           = _mm_unpackhi_pd(__gmx_t1, row1); \
  49 }
  50
  51 /* Normal sum of four ymm registers */
  52 #define gmx_mm_sum4_pd(t0, t1, t2, t3)  _mm_add_pd(_mm_add_pd(t0, t1), _mm_add_pd(t2, t3))
  53
  54 static int
  55 gmx_mm_any_lt(__m128d a, __m128d b)
  56 {
  57     return _mm_movemask_pd(_mm_cmplt_pd(a, b));
  58 }
  59
  60 static gmx_inline __m128d
  61 gmx_mm_calc_rsq_pd(__m128d dx, __m128d dy, __m128d dz)
  62 {
  63     return _mm_add_pd( _mm_add_pd( _mm_mul_pd(dx, dx), _mm_mul_pd(dy, dy) ), _mm_mul_pd(dz, dz) );
  64 }
  65
  66
  67 /* Load a double value from 1-2 places, merge into xmm register */
  68 static gmx_inline __m128d
  69 gmx_mm_load_2real_swizzle_pd(const double * gmx_restrict ptrA,
  70                              const double * gmx_restrict ptrB)
  71 {
  72     return _mm_unpacklo_pd(_mm_load_sd(ptrA), _mm_load_sd(ptrB));
  73 }
  74
  75 static gmx_inline __m128d
  76 gmx_mm_load_1real_pd(const double * gmx_restrict ptrA)
  77 {
  78     return _mm_load_sd(ptrA);
  79 }
  80
  81
  82 static gmx_inline void
  83 gmx_mm_store_2real_swizzle_pd(double * gmx_restrict ptrA,
  84                               double * gmx_restrict ptrB,
  85                               __m128d               xmm1)
  86 {
  87     __m128d t2;
  88
  89     t2       = _mm_unpackhi_pd(xmm1, xmm1);
  90     _mm_store_sd(ptrA, xmm1);
  91     _mm_store_sd(ptrB, t2);
  92 }
  93
  94 static gmx_inline void
  95 gmx_mm_store_1real_pd(double * gmx_restrict ptrA, __m128d xmm1)
  96 {
  97     _mm_store_sd(ptrA, xmm1);
  98 }
  99
 100
 101 /* Similar to store, but increments value in memory */
 102 static gmx_inline void
 103 gmx_mm_increment_2real_swizzle_pd(double * gmx_restrict ptrA,
 104                                   double * gmx_restrict ptrB, __m128d xmm1)
 105 {
 106     __m128d t1;
 107
 108     t1   = _mm_unpackhi_pd(xmm1, xmm1);
 109     xmm1 = _mm_add_sd(xmm1, _mm_load_sd(ptrA));
 110     t1   = _mm_add_sd(t1, _mm_load_sd(ptrB));
 111     _mm_store_sd(ptrA, xmm1);
 112     _mm_store_sd(ptrB, t1);
 113 }
 114
 115 static gmx_inline void
 116 gmx_mm_increment_1real_pd(double * gmx_restrict ptrA, __m128d xmm1)
 117 {
 118     __m128d tmp;
 119
 120     tmp = gmx_mm_load_1real_pd(ptrA);
 121     tmp = _mm_add_sd(tmp, xmm1);
 122     gmx_mm_store_1real_pd(ptrA, tmp);
 123 }
 124
 125
 126 static gmx_inline void
 127 gmx_mm_load_2pair_swizzle_pd(const double * gmx_restrict p1,
 128                              const double * gmx_restrict p2,
 129                              __m128d * gmx_restrict      c6,
 130                              __m128d * gmx_restrict      c12)
 131 {
 132     __m128d t1, t2, t3;
 133
 134     t1   = _mm_loadu_pd(p1);
 135     t2   = _mm_loadu_pd(p2);
 136     *c6  = _mm_unpacklo_pd(t1, t2);
 137     *c12 = _mm_unpackhi_pd(t1, t2);
 138 }
 139
 140 static gmx_inline void
 141 gmx_mm_load_1pair_swizzle_pd(const double * gmx_restrict p1,
 142                              __m128d * gmx_restrict      c6,
 143                              __m128d * gmx_restrict      c12)
 144 {
 145     *c6     = _mm_load_sd(p1);
 146     *c12    = _mm_load_sd(p1+1);
 147 }
 148
 149
 150
 151 static gmx_inline void
 152 gmx_mm_load_shift_and_1rvec_broadcast_pd(const double * gmx_restrict xyz_shift,
 153                                          const double * gmx_restrict xyz,
 154                                          __m128d * gmx_restrict      x1,
 155                                          __m128d * gmx_restrict      y1,
 156                                          __m128d * gmx_restrict      z1)
 157 {
 158     __m128d mem_xy, mem_z, mem_sxy, mem_sz;
 159
 160     mem_xy  = _mm_loadu_pd(xyz);
 161     mem_z   = _mm_load_sd(xyz+2);
 162     mem_sxy = _mm_loadu_pd(xyz_shift);
 163     mem_sz  = _mm_load_sd(xyz_shift+2);
 164
 165     mem_xy  = _mm_add_pd(mem_xy, mem_sxy);
 166     mem_z   = _mm_add_pd(mem_z, mem_sz);
 167
 168     *x1  = _mm_shuffle_pd(mem_xy, mem_xy, _MM_SHUFFLE2(0, 0));
 169     *y1  = _mm_shuffle_pd(mem_xy, mem_xy, _MM_SHUFFLE2(1, 1));
 170     *z1  = _mm_shuffle_pd(mem_z, mem_z, _MM_SHUFFLE2(0, 0));
 171 }
 172
 173
 174 static gmx_inline void
 175 gmx_mm_load_shift_and_3rvec_broadcast_pd(const double * gmx_restrict xyz_shift,
 176                                          const double * gmx_restrict xyz,
 177                                          __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 178                                          __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 179                                          __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3)
 180 {
 181     __m128d t1, t2, t3, t4, t5, sxy, sz, szx, syz;
 182
 183     t1  = _mm_loadu_pd(xyz);
 184     t2  = _mm_loadu_pd(xyz+2);
 185     t3  = _mm_loadu_pd(xyz+4);
 186     t4  = _mm_loadu_pd(xyz+6);
 187     t5  = _mm_load_sd(xyz+8);
 188
 189     sxy = _mm_loadu_pd(xyz_shift);
 190     sz  = _mm_load_sd(xyz_shift+2);
 191     szx = _mm_shuffle_pd(sz, sxy, _MM_SHUFFLE2(0, 0));
 192     syz = _mm_shuffle_pd(sxy, sz, _MM_SHUFFLE2(0, 1));
 193
 194     t1  = _mm_add_pd(t1, sxy);
 195     t2  = _mm_add_pd(t2, szx);
 196     t3  = _mm_add_pd(t3, syz);
 197     t4  = _mm_add_pd(t4, sxy);
 198     t5  = _mm_add_sd(t5, sz);
 199
 200     *x1  = _mm_shuffle_pd(t1, t1, _MM_SHUFFLE2(0, 0));
 201     *y1  = _mm_shuffle_pd(t1, t1, _MM_SHUFFLE2(1, 1));
 202     *z1  = _mm_shuffle_pd(t2, t2, _MM_SHUFFLE2(0, 0));
 203     *x2  = _mm_shuffle_pd(t2, t2, _MM_SHUFFLE2(1, 1));
 204     *y2  = _mm_shuffle_pd(t3, t3, _MM_SHUFFLE2(0, 0));
 205     *z2  = _mm_shuffle_pd(t3, t3, _MM_SHUFFLE2(1, 1));
 206     *x3  = _mm_shuffle_pd(t4, t4, _MM_SHUFFLE2(0, 0));
 207     *y3  = _mm_shuffle_pd(t4, t4, _MM_SHUFFLE2(1, 1));
 208     *z3  = _mm_shuffle_pd(t5, t5, _MM_SHUFFLE2(0, 0));
 209 }
 210
 211
 212 static gmx_inline void
 213 gmx_mm_load_shift_and_4rvec_broadcast_pd(const double * gmx_restrict xyz_shift,
 214                                          const double * gmx_restrict xyz,
 215                                          __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 216                                          __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 217                                          __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3,
 218                                          __m128d * gmx_restrict x4, __m128d * gmx_restrict y4, __m128d * gmx_restrict z4)
 219 {
 220     __m128d t1, t2, t3, t4, t5, t6, sxy, sz, szx, syz;
 221
 222     t1  = _mm_loadu_pd(xyz);
 223     t2  = _mm_loadu_pd(xyz+2);
 224     t3  = _mm_loadu_pd(xyz+4);
 225     t4  = _mm_loadu_pd(xyz+6);
 226     t5  = _mm_loadu_pd(xyz+8);
 227     t6  = _mm_loadu_pd(xyz+10);
 228
 229     sxy = _mm_loadu_pd(xyz_shift);
 230     sz  = _mm_load_sd(xyz_shift+2);
 231     szx = _mm_shuffle_pd(sz, sxy, _MM_SHUFFLE2(0, 0));
 232     syz = _mm_shuffle_pd(sxy, sz, _MM_SHUFFLE2(0, 1));
 233
 234     t1  = _mm_add_pd(t1, sxy);
 235     t2  = _mm_add_pd(t2, szx);
 236     t3  = _mm_add_pd(t3, syz);
 237     t4  = _mm_add_pd(t4, sxy);
 238     t5  = _mm_add_pd(t5, szx);
 239     t6  = _mm_add_pd(t6, syz);
 240
 241     *x1  = _mm_shuffle_pd(t1, t1, _MM_SHUFFLE2(0, 0));
 242     *y1  = _mm_shuffle_pd(t1, t1, _MM_SHUFFLE2(1, 1));
 243     *z1  = _mm_shuffle_pd(t2, t2, _MM_SHUFFLE2(0, 0));
 244     *x2  = _mm_shuffle_pd(t2, t2, _MM_SHUFFLE2(1, 1));
 245     *y2  = _mm_shuffle_pd(t3, t3, _MM_SHUFFLE2(0, 0));
 246     *z2  = _mm_shuffle_pd(t3, t3, _MM_SHUFFLE2(1, 1));
 247     *x3  = _mm_shuffle_pd(t4, t4, _MM_SHUFFLE2(0, 0));
 248     *y3  = _mm_shuffle_pd(t4, t4, _MM_SHUFFLE2(1, 1));
 249     *z3  = _mm_shuffle_pd(t5, t5, _MM_SHUFFLE2(0, 0));
 250     *x4  = _mm_shuffle_pd(t5, t5, _MM_SHUFFLE2(1, 1));
 251     *y4  = _mm_shuffle_pd(t6, t6, _MM_SHUFFLE2(0, 0));
 252     *z4  = _mm_shuffle_pd(t6, t6, _MM_SHUFFLE2(1, 1));
 253 }
 254
 255
 256
 257
 258 static gmx_inline void
 259 gmx_mm_load_1rvec_1ptr_swizzle_pd(const double * gmx_restrict p1,
 260                                   __m128d * gmx_restrict x, __m128d * gmx_restrict y, __m128d * gmx_restrict z)
 261 {
 262     *x            = _mm_load_sd(p1);
 263     *y            = _mm_load_sd(p1+1);
 264     *z            = _mm_load_sd(p1+2);
 265 }
 266
 267 static gmx_inline void
 268 gmx_mm_load_3rvec_1ptr_swizzle_pd(const double * gmx_restrict p1,
 269                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 270                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 271                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3)
 272 {
 273     *x1            = _mm_load_sd(p1);
 274     *y1            = _mm_load_sd(p1+1);
 275     *z1            = _mm_load_sd(p1+2);
 276     *x2            = _mm_load_sd(p1+3);
 277     *y2            = _mm_load_sd(p1+4);
 278     *z2            = _mm_load_sd(p1+5);
 279     *x3            = _mm_load_sd(p1+6);
 280     *y3            = _mm_load_sd(p1+7);
 281     *z3            = _mm_load_sd(p1+8);
 282 }
 283
 284 static gmx_inline void
 285 gmx_mm_load_4rvec_1ptr_swizzle_pd(const double * gmx_restrict p1,
 286                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 287                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 288                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3,
 289                                   __m128d * gmx_restrict x4, __m128d * gmx_restrict y4, __m128d * gmx_restrict z4)
 290 {
 291     *x1            = _mm_load_sd(p1);
 292     *y1            = _mm_load_sd(p1+1);
 293     *z1            = _mm_load_sd(p1+2);
 294     *x2            = _mm_load_sd(p1+3);
 295     *y2            = _mm_load_sd(p1+4);
 296     *z2            = _mm_load_sd(p1+5);
 297     *x3            = _mm_load_sd(p1+6);
 298     *y3            = _mm_load_sd(p1+7);
 299     *z3            = _mm_load_sd(p1+8);
 300     *x4            = _mm_load_sd(p1+9);
 301     *y4            = _mm_load_sd(p1+10);
 302     *z4            = _mm_load_sd(p1+11);
 303 }
 304
 305
 306 static gmx_inline void
 307 gmx_mm_load_1rvec_2ptr_swizzle_pd(const double * gmx_restrict ptrA,
 308                                   const double * gmx_restrict ptrB,
 309                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1)
 310 {
 311     __m128d t1, t2, t3, t4;
 312     t1           = _mm_loadu_pd(ptrA);
 313     t2           = _mm_loadu_pd(ptrB);
 314     t3           = _mm_load_sd(ptrA+2);
 315     t4           = _mm_load_sd(ptrB+2);
 316     GMX_MM_TRANSPOSE2_PD(t1, t2);
 317     *x1          = t1;
 318     *y1          = t2;
 319     *z1          = _mm_unpacklo_pd(t3, t4);
 320 }
 321
 322
 323 static gmx_inline void
 324 gmx_mm_load_3rvec_2ptr_swizzle_pd(const double * gmx_restrict ptrA, const double * gmx_restrict ptrB,
 325                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 326                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 327                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3)
 328 {
 329     __m128d t1, t2, t3, t4, t5, t6, t7, t8, t9, t10;
 330     t1           = _mm_loadu_pd(ptrA);
 331     t2           = _mm_loadu_pd(ptrB);
 332     t3           = _mm_loadu_pd(ptrA+2);
 333     t4           = _mm_loadu_pd(ptrB+2);
 334     t5           = _mm_loadu_pd(ptrA+4);
 335     t6           = _mm_loadu_pd(ptrB+4);
 336     t7           = _mm_loadu_pd(ptrA+6);
 337     t8           = _mm_loadu_pd(ptrB+6);
 338     t9           = _mm_load_sd(ptrA+8);
 339     t10          = _mm_load_sd(ptrB+8);
 340     GMX_MM_TRANSPOSE2_PD(t1, t2);
 341     GMX_MM_TRANSPOSE2_PD(t3, t4);
 342     GMX_MM_TRANSPOSE2_PD(t5, t6);
 343     GMX_MM_TRANSPOSE2_PD(t7, t8);
 344     *x1          = t1;
 345     *y1          = t2;
 346     *z1          = t3;
 347     *x2          = t4;
 348     *y2          = t5;
 349     *z2          = t6;
 350     *x3          = t7;
 351     *y3          = t8;
 352     *z3          = _mm_unpacklo_pd(t9, t10);
 353 }
 354
 355
 356 static gmx_inline void
 357 gmx_mm_load_4rvec_2ptr_swizzle_pd(const double * gmx_restrict ptrA, const double * gmx_restrict ptrB,
 358                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 359                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 360                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3,
 361                                   __m128d * gmx_restrict x4, __m128d * gmx_restrict y4, __m128d * gmx_restrict z4)
 362 {
 363     __m128d t1, t2, t3, t4, t5, t6;
 364     t1           = _mm_loadu_pd(ptrA);
 365     t2           = _mm_loadu_pd(ptrB);
 366     t3           = _mm_loadu_pd(ptrA+2);
 367     t4           = _mm_loadu_pd(ptrB+2);
 368     t5           = _mm_loadu_pd(ptrA+4);
 369     t6           = _mm_loadu_pd(ptrB+4);
 370     GMX_MM_TRANSPOSE2_PD(t1, t2);
 371     GMX_MM_TRANSPOSE2_PD(t3, t4);
 372     GMX_MM_TRANSPOSE2_PD(t5, t6);
 373     *x1          = t1;
 374     *y1          = t2;
 375     *z1          = t3;
 376     *x2          = t4;
 377     *y2          = t5;
 378     *z2          = t6;
 379     t1           = _mm_loadu_pd(ptrA+6);
 380     t2           = _mm_loadu_pd(ptrB+6);
 381     t3           = _mm_loadu_pd(ptrA+8);
 382     t4           = _mm_loadu_pd(ptrB+8);
 383     t5           = _mm_loadu_pd(ptrA+10);
 384     t6           = _mm_loadu_pd(ptrB+10);
 385     GMX_MM_TRANSPOSE2_PD(t1, t2);
 386     GMX_MM_TRANSPOSE2_PD(t3, t4);
 387     GMX_MM_TRANSPOSE2_PD(t5, t6);
 388     *x3          = t1;
 389     *y3          = t2;
 390     *z3          = t3;
 391     *x4          = t4;
 392     *y4          = t5;
 393     *z4          = t6;
 394 }
 395
 396
 397 /* Routines to decrement rvec in memory, typically use for j particle force updates */
 398 static gmx_inline void
 399 gmx_mm_decrement_1rvec_1ptr_noswizzle_pd(double * gmx_restrict ptrA,
 400                                          __m128d xy, __m128d z)
 401 {
 402     __m128d t1, t2;
 403
 404     t1 = _mm_loadu_pd(ptrA);
 405     t2 = _mm_load_sd(ptrA+2);
 406
 407     t1 = _mm_sub_pd(t1, xy);
 408     t2 = _mm_sub_sd(t2, z);
 409
 410     _mm_storeu_pd(ptrA, t1);
 411     _mm_store_sd(ptrA+2, t2);
 412 }
 413
 414
 415 static gmx_inline void
 416 gmx_mm_decrement_1rvec_1ptr_swizzle_pd(double * gmx_restrict ptrA,
 417                                        __m128d x1, __m128d y1, __m128d z1)
 418 {
 419     __m128d t1, t2, t3;
 420
 421     t1           = _mm_load_sd(ptrA);
 422     t2           = _mm_load_sd(ptrA+1);
 423     t3           = _mm_load_sd(ptrA+2);
 424
 425     t1           = _mm_sub_sd(t1, x1);
 426     t2           = _mm_sub_sd(t2, y1);
 427     t3           = _mm_sub_sd(t3, z1);
 428     _mm_store_sd(ptrA, t1);
 429     _mm_store_sd(ptrA+1, t2);
 430     _mm_store_sd(ptrA+2, t3);
 431 }
 432
 433
 434 #if defined (_MSC_VER) && defined(_M_IX86)
 435 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 436 #define gmx_mm_decrement_3rvec_1ptr_swizzle_pd(ptrA, _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3) \
 437     { \
 438         __m128d _t1, _t2, _t3, _t4, _t5; \
 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         _x1          = _mm_unpacklo_pd(_x1, _y1); \
 445         _z1          = _mm_unpacklo_pd(_z1, _x2); \
 446         _y2          = _mm_unpacklo_pd(_y2, _z2); \
 447         _x3          = _mm_unpacklo_pd(_x3, _y3); \
 448         _t1          = _mm_sub_pd(_t1, _x1); \
 449         _t2          = _mm_sub_pd(_t2, _z1); \
 450         _t3          = _mm_sub_pd(_t3, _y2); \
 451         _t4          = _mm_sub_pd(_t4, _x3); \
 452         _t5          = _mm_sub_sd(_t5, _z3); \
 453         _mm_storeu_pd(ptrA, _t1); \
 454         _mm_storeu_pd(ptrA+2, _t2); \
 455         _mm_storeu_pd(ptrA+4, _t3); \
 456         _mm_storeu_pd(ptrA+6, _t4); \
 457         _mm_store_sd(ptrA+8, _t5); \
 458     }
 459 #else
 460 /* Real function for sane compilers */
 461 static gmx_inline void
 462 gmx_mm_decrement_3rvec_1ptr_swizzle_pd(double * gmx_restrict ptrA,
 463                                        __m128d x1, __m128d y1, __m128d z1,
 464                                        __m128d x2, __m128d y2, __m128d z2,
 465                                        __m128d x3, __m128d y3, __m128d z3)
 466 {
 467     __m128d t1, t2, t3, t4, t5;
 468
 469     t1          = _mm_loadu_pd(ptrA);
 470     t2          = _mm_loadu_pd(ptrA+2);
 471     t3          = _mm_loadu_pd(ptrA+4);
 472     t4          = _mm_loadu_pd(ptrA+6);
 473     t5          = _mm_load_sd(ptrA+8);
 474
 475     x1          = _mm_unpacklo_pd(x1, y1);
 476     z1          = _mm_unpacklo_pd(z1, x2);
 477     y2          = _mm_unpacklo_pd(y2, z2);
 478     x3          = _mm_unpacklo_pd(x3, y3);
 479     /* nothing to be done for z3 */
 480
 481     t1          = _mm_sub_pd(t1, x1);
 482     t2          = _mm_sub_pd(t2, z1);
 483     t3          = _mm_sub_pd(t3, y2);
 484     t4          = _mm_sub_pd(t4, x3);
 485     t5          = _mm_sub_sd(t5, z3);
 486     _mm_storeu_pd(ptrA, t1);
 487     _mm_storeu_pd(ptrA+2, t2);
 488     _mm_storeu_pd(ptrA+4, t3);
 489     _mm_storeu_pd(ptrA+6, t4);
 490     _mm_store_sd(ptrA+8, t5);
 491 }
 492 #endif
 493
 494
 495 #if defined (_MSC_VER) && defined(_M_IX86)
 496 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 497 #define gmx_mm_decrement_4rvec_1ptr_swizzle_pd(ptrA, _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3, _x4, _y4, _z4) \
 498     { \
 499         __m128d _t1, _t2, _t3, _t4, _t5, _t6; \
 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         _x1          = _mm_unpacklo_pd(_x1, _y1); \
 507         _z1          = _mm_unpacklo_pd(_z1, _x2); \
 508         _y2          = _mm_unpacklo_pd(_y2, _z2); \
 509         _x3          = _mm_unpacklo_pd(_x3, _y3); \
 510         _z3          = _mm_unpacklo_pd(_z3, _x4); \
 511         _y4          = _mm_unpacklo_pd(_y4, _z4); \
 512         _mm_storeu_pd(ptrA,    _mm_sub_pd( _t1, _x1 )); \
 513         _mm_storeu_pd(ptrA+2,  _mm_sub_pd( _t2, _z1 )); \
 514         _mm_storeu_pd(ptrA+4,  _mm_sub_pd( _t3, _y2 )); \
 515         _mm_storeu_pd(ptrA+6,  _mm_sub_pd( _t4, _x3 )); \
 516         _mm_storeu_pd(ptrA+8,  _mm_sub_pd( _t5, _z3 )); \
 517         _mm_storeu_pd(ptrA+10, _mm_sub_pd( _t6, _y4 )); \
 518     }
 519 #else
 520 /* Real function for sane compilers */
 521 static gmx_inline void
 522 gmx_mm_decrement_4rvec_1ptr_swizzle_pd(double * gmx_restrict ptrA,
 523                                        __m128d x1, __m128d y1, __m128d z1,
 524                                        __m128d x2, __m128d y2, __m128d z2,
 525                                        __m128d x3, __m128d y3, __m128d z3,
 526                                        __m128d x4, __m128d y4, __m128d z4)
 527 {
 528     __m128d t1, t2, t3, t4, t5, t6;
 529
 530     t1          = _mm_loadu_pd(ptrA);
 531     t2          = _mm_loadu_pd(ptrA+2);
 532     t3          = _mm_loadu_pd(ptrA+4);
 533     t4          = _mm_loadu_pd(ptrA+6);
 534     t5          = _mm_loadu_pd(ptrA+8);
 535     t6          = _mm_loadu_pd(ptrA+10);
 536
 537     x1          = _mm_unpacklo_pd(x1, y1);
 538     z1          = _mm_unpacklo_pd(z1, x2);
 539     y2          = _mm_unpacklo_pd(y2, z2);
 540     x3          = _mm_unpacklo_pd(x3, y3);
 541     z3          = _mm_unpacklo_pd(z3, x4);
 542     y4          = _mm_unpacklo_pd(y4, z4);
 543
 544     _mm_storeu_pd(ptrA,    _mm_sub_pd( t1, x1 ));
 545     _mm_storeu_pd(ptrA+2,  _mm_sub_pd( t2, z1 ));
 546     _mm_storeu_pd(ptrA+4,  _mm_sub_pd( t3, y2 ));
 547     _mm_storeu_pd(ptrA+6,  _mm_sub_pd( t4, x3 ));
 548     _mm_storeu_pd(ptrA+8,  _mm_sub_pd( t5, z3 ));
 549     _mm_storeu_pd(ptrA+10, _mm_sub_pd( t6, y4 ));
 550 }
 551 #endif
 552
 553
 554 static gmx_inline void
 555 gmx_mm_decrement_1rvec_2ptr_swizzle_pd(double * gmx_restrict ptrA, double * gmx_restrict ptrB,
 556                                        __m128d x1, __m128d y1, __m128d z1)
 557 {
 558     __m128d t1, t2, t3, t4, t5, t6, t7;
 559
 560     t1          = _mm_loadu_pd(ptrA);
 561     t2          = _mm_load_sd(ptrA+2);
 562     t3          = _mm_loadu_pd(ptrB);
 563     t4          = _mm_load_sd(ptrB+2);
 564
 565     t5          = _mm_unpacklo_pd(x1, y1);
 566     t6          = _mm_unpackhi_pd(x1, y1);
 567     t7          = _mm_unpackhi_pd(z1, z1);
 568
 569     t1          = _mm_sub_pd(t1, t5);
 570     t2          = _mm_sub_sd(t2, z1);
 571
 572     t3          = _mm_sub_pd(t3, t6);
 573     t4          = _mm_sub_sd(t4, t7);
 574
 575     _mm_storeu_pd(ptrA, t1);
 576     _mm_store_sd(ptrA+2, t2);
 577     _mm_storeu_pd(ptrB, t3);
 578     _mm_store_sd(ptrB+2, t4);
 579 }
 580
 581 #if defined (_MSC_VER) && defined(_M_IX86)
 582 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 583 #define gmx_mm_decrement_3rvec_2ptr_swizzle_pd(ptrA, ptrB, _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3) \
 584     { \
 585         __m128d _t1, _t2, _t3, _t4, _t5, _t6, _t7, _t8, _t9, _t10; \
 586         __m128d _tA, _tB, _tC, _tD, _tE, _tF, _tG, _tH, _tI; \
 587         _t1          = _mm_loadu_pd(ptrA); \
 588         _t2          = _mm_loadu_pd(ptrA+2); \
 589         _t3          = _mm_loadu_pd(ptrA+4); \
 590         _t4          = _mm_loadu_pd(ptrA+6); \
 591         _t5          = _mm_load_sd(ptrA+8); \
 592         _t6          = _mm_loadu_pd(ptrB); \
 593         _t7          = _mm_loadu_pd(ptrB+2); \
 594         _t8          = _mm_loadu_pd(ptrB+4); \
 595         _t9          = _mm_loadu_pd(ptrB+6); \
 596         _t10         = _mm_load_sd(ptrB+8); \
 597         _tA          = _mm_unpacklo_pd(_x1, _y1); \
 598         _tB          = _mm_unpackhi_pd(_x1, _y1); \
 599         _tC          = _mm_unpacklo_pd(_z1, _x2); \
 600         _tD          = _mm_unpackhi_pd(_z1, _x2); \
 601         _tE          = _mm_unpacklo_pd(_y2, _z2); \
 602         _tF          = _mm_unpackhi_pd(_y2, _z2); \
 603         _tG          = _mm_unpacklo_pd(_x3, _y3); \
 604         _tH          = _mm_unpackhi_pd(_x3, _y3); \
 605         _tI          = _mm_unpackhi_pd(_z3, _z3); \
 606         _t1          = _mm_sub_pd(_t1, _tA); \
 607         _t2          = _mm_sub_pd(_t2, _tC); \
 608         _t3          = _mm_sub_pd(_t3, _tE); \
 609         _t4          = _mm_sub_pd(_t4, _tG); \
 610         _t5          = _mm_sub_sd(_t5, _z3); \
 611         _t6          = _mm_sub_pd(_t6, _tB); \
 612         _t7          = _mm_sub_pd(_t7, _tD); \
 613         _t8          = _mm_sub_pd(_t8, _tF); \
 614         _t9          = _mm_sub_pd(_t9, _tH); \
 615         _t10         = _mm_sub_sd(_t10, _tI); \
 616         _mm_storeu_pd(ptrA, _t1); \
 617         _mm_storeu_pd(ptrA+2, _t2); \
 618         _mm_storeu_pd(ptrA+4, _t3); \
 619         _mm_storeu_pd(ptrA+6, _t4); \
 620         _mm_store_sd(ptrA+8, _t5); \
 621         _mm_storeu_pd(ptrB, _t6); \
 622         _mm_storeu_pd(ptrB+2, _t7); \
 623         _mm_storeu_pd(ptrB+4, _t8); \
 624         _mm_storeu_pd(ptrB+6, _t9); \
 625         _mm_store_sd(ptrB+8, _t10); \
 626     }
 627 #else
 628 /* Real function for sane compilers */
 629 static gmx_inline void
 630 gmx_mm_decrement_3rvec_2ptr_swizzle_pd(double * gmx_restrict ptrA, double * gmx_restrict ptrB,
 631                                        __m128d x1, __m128d y1, __m128d z1,
 632                                        __m128d x2, __m128d y2, __m128d z2,
 633                                        __m128d x3, __m128d y3, __m128d z3)
 634 {
 635     __m128d t1, t2, t3, t4, t5, t6, t7, t8, t9, t10;
 636     __m128d tA, tB, tC, tD, tE, tF, tG, tH, tI;
 637
 638     t1          = _mm_loadu_pd(ptrA);
 639     t2          = _mm_loadu_pd(ptrA+2);
 640     t3          = _mm_loadu_pd(ptrA+4);
 641     t4          = _mm_loadu_pd(ptrA+6);
 642     t5          = _mm_load_sd(ptrA+8);
 643     t6          = _mm_loadu_pd(ptrB);
 644     t7          = _mm_loadu_pd(ptrB+2);
 645     t8          = _mm_loadu_pd(ptrB+4);
 646     t9          = _mm_loadu_pd(ptrB+6);
 647     t10         = _mm_load_sd(ptrB+8);
 648
 649     tA          = _mm_unpacklo_pd(x1, y1);
 650     tB          = _mm_unpackhi_pd(x1, y1);
 651     tC          = _mm_unpacklo_pd(z1, x2);
 652     tD          = _mm_unpackhi_pd(z1, x2);
 653     tE          = _mm_unpacklo_pd(y2, z2);
 654     tF          = _mm_unpackhi_pd(y2, z2);
 655     tG          = _mm_unpacklo_pd(x3, y3);
 656     tH          = _mm_unpackhi_pd(x3, y3);
 657     tI          = _mm_unpackhi_pd(z3, z3);
 658
 659     t1          = _mm_sub_pd(t1, tA);
 660     t2          = _mm_sub_pd(t2, tC);
 661     t3          = _mm_sub_pd(t3, tE);
 662     t4          = _mm_sub_pd(t4, tG);
 663     t5          = _mm_sub_sd(t5, z3);
 664
 665     t6          = _mm_sub_pd(t6, tB);
 666     t7          = _mm_sub_pd(t7, tD);
 667     t8          = _mm_sub_pd(t8, tF);
 668     t9          = _mm_sub_pd(t9, tH);
 669     t10         = _mm_sub_sd(t10, tI);
 670
 671     _mm_storeu_pd(ptrA, t1);
 672     _mm_storeu_pd(ptrA+2, t2);
 673     _mm_storeu_pd(ptrA+4, t3);
 674     _mm_storeu_pd(ptrA+6, t4);
 675     _mm_store_sd(ptrA+8, t5);
 676     _mm_storeu_pd(ptrB, t6);
 677     _mm_storeu_pd(ptrB+2, t7);
 678     _mm_storeu_pd(ptrB+4, t8);
 679     _mm_storeu_pd(ptrB+6, t9);
 680     _mm_store_sd(ptrB+8, t10);
 681 }
 682 #endif
 683
 684
 685 #if defined (_MSC_VER) && defined(_M_IX86)
 686 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 687 #define gmx_mm_decrement_4rvec_2ptr_swizzle_pd(ptrA, ptrB, _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3, _x4, _y4, _z4) \
 688     { \
 689         __m128d _t1, _t2, _t3, _t4, _t5, _t6, _t7, _t8, _t9, _t10, _t11, _t12; \
 690         __m128d _tA, _tB, _tC, _tD, _tE, _tF, _tG, _tH, _tI, _tJ, _tK, _tL; \
 691         _t1          = _mm_loadu_pd(ptrA); \
 692         _t2          = _mm_loadu_pd(ptrA+2); \
 693         _t3          = _mm_loadu_pd(ptrA+4); \
 694         _t4          = _mm_loadu_pd(ptrA+6); \
 695         _t5          = _mm_loadu_pd(ptrA+8); \
 696         _t6          = _mm_loadu_pd(ptrA+10); \
 697         _t7          = _mm_loadu_pd(ptrB); \
 698         _t8          = _mm_loadu_pd(ptrB+2); \
 699         _t9          = _mm_loadu_pd(ptrB+4); \
 700         _t10         = _mm_loadu_pd(ptrB+6); \
 701         _t11         = _mm_loadu_pd(ptrB+8); \
 702         _t12         = _mm_loadu_pd(ptrB+10); \
 703         _tA          = _mm_unpacklo_pd(_x1, _y1); \
 704         _tB          = _mm_unpackhi_pd(_x1, _y1); \
 705         _tC          = _mm_unpacklo_pd(_z1, _x2); \
 706         _tD          = _mm_unpackhi_pd(_z1, _x2); \
 707         _tE          = _mm_unpacklo_pd(_y2, _z2); \
 708         _tF          = _mm_unpackhi_pd(_y2, _z2); \
 709         _tG          = _mm_unpacklo_pd(_x3, _y3); \
 710         _tH          = _mm_unpackhi_pd(_x3, _y3); \
 711         _tI          = _mm_unpacklo_pd(_z3, _x4); \
 712         _tJ          = _mm_unpackhi_pd(_z3, _x4); \
 713         _tK          = _mm_unpacklo_pd(_y4, _z4); \
 714         _tL          = _mm_unpackhi_pd(_y4, _z4); \
 715         _t1          = _mm_sub_pd(_t1, _tA); \
 716         _t2          = _mm_sub_pd(_t2, _tC); \
 717         _t3          = _mm_sub_pd(_t3, _tE); \
 718         _t4          = _mm_sub_pd(_t4, _tG); \
 719         _t5          = _mm_sub_pd(_t5, _tI); \
 720         _t6          = _mm_sub_pd(_t6, _tK); \
 721         _t7          = _mm_sub_pd(_t7, _tB); \
 722         _t8          = _mm_sub_pd(_t8, _tD); \
 723         _t9          = _mm_sub_pd(_t9, _tF); \
 724         _t10         = _mm_sub_pd(_t10, _tH); \
 725         _t11         = _mm_sub_pd(_t11, _tJ); \
 726         _t12         = _mm_sub_pd(_t12, _tL); \
 727         _mm_storeu_pd(ptrA,  _t1); \
 728         _mm_storeu_pd(ptrA+2, _t2); \
 729         _mm_storeu_pd(ptrA+4, _t3); \
 730         _mm_storeu_pd(ptrA+6, _t4); \
 731         _mm_storeu_pd(ptrA+8, _t5); \
 732         _mm_storeu_pd(ptrA+10, _t6); \
 733         _mm_storeu_pd(ptrB,  _t7); \
 734         _mm_storeu_pd(ptrB+2, _t8); \
 735         _mm_storeu_pd(ptrB+4, _t9); \
 736         _mm_storeu_pd(ptrB+6, _t10); \
 737         _mm_storeu_pd(ptrB+8, _t11); \
 738         _mm_storeu_pd(ptrB+10, _t12); \
 739     }
 740 #else
 741 /* Real function for sane compilers */
 742 static gmx_inline void
 743 gmx_mm_decrement_4rvec_2ptr_swizzle_pd(double * gmx_restrict ptrA, double * gmx_restrict ptrB,
 744                                        __m128d x1, __m128d y1, __m128d z1,
 745                                        __m128d x2, __m128d y2, __m128d z2,
 746                                        __m128d x3, __m128d y3, __m128d z3,
 747                                        __m128d x4, __m128d y4, __m128d z4)
 748 {
 749     __m128d t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12;
 750     __m128d tA, tB, tC, tD, tE, tF, tG, tH, tI, tJ, tK, tL;
 751
 752     t1          = _mm_loadu_pd(ptrA);
 753     t2          = _mm_loadu_pd(ptrA+2);
 754     t3          = _mm_loadu_pd(ptrA+4);
 755     t4          = _mm_loadu_pd(ptrA+6);
 756     t5          = _mm_loadu_pd(ptrA+8);
 757     t6          = _mm_loadu_pd(ptrA+10);
 758     t7          = _mm_loadu_pd(ptrB);
 759     t8          = _mm_loadu_pd(ptrB+2);
 760     t9          = _mm_loadu_pd(ptrB+4);
 761     t10         = _mm_loadu_pd(ptrB+6);
 762     t11         = _mm_loadu_pd(ptrB+8);
 763     t12         = _mm_loadu_pd(ptrB+10);
 764
 765     tA          = _mm_unpacklo_pd(x1, y1);
 766     tB          = _mm_unpackhi_pd(x1, y1);
 767     tC          = _mm_unpacklo_pd(z1, x2);
 768     tD          = _mm_unpackhi_pd(z1, x2);
 769     tE          = _mm_unpacklo_pd(y2, z2);
 770     tF          = _mm_unpackhi_pd(y2, z2);
 771     tG          = _mm_unpacklo_pd(x3, y3);
 772     tH          = _mm_unpackhi_pd(x3, y3);
 773     tI          = _mm_unpacklo_pd(z3, x4);
 774     tJ          = _mm_unpackhi_pd(z3, x4);
 775     tK          = _mm_unpacklo_pd(y4, z4);
 776     tL          = _mm_unpackhi_pd(y4, z4);
 777
 778     t1          = _mm_sub_pd(t1, tA);
 779     t2          = _mm_sub_pd(t2, tC);
 780     t3          = _mm_sub_pd(t3, tE);
 781     t4          = _mm_sub_pd(t4, tG);
 782     t5          = _mm_sub_pd(t5, tI);
 783     t6          = _mm_sub_pd(t6, tK);
 784
 785     t7          = _mm_sub_pd(t7, tB);
 786     t8          = _mm_sub_pd(t8, tD);
 787     t9          = _mm_sub_pd(t9, tF);
 788     t10         = _mm_sub_pd(t10, tH);
 789     t11         = _mm_sub_pd(t11, tJ);
 790     t12         = _mm_sub_pd(t12, tL);
 791
 792     _mm_storeu_pd(ptrA,  t1);
 793     _mm_storeu_pd(ptrA+2, t2);
 794     _mm_storeu_pd(ptrA+4, t3);
 795     _mm_storeu_pd(ptrA+6, t4);
 796     _mm_storeu_pd(ptrA+8, t5);
 797     _mm_storeu_pd(ptrA+10, t6);
 798     _mm_storeu_pd(ptrB,  t7);
 799     _mm_storeu_pd(ptrB+2, t8);
 800     _mm_storeu_pd(ptrB+4, t9);
 801     _mm_storeu_pd(ptrB+6, t10);
 802     _mm_storeu_pd(ptrB+8, t11);
 803     _mm_storeu_pd(ptrB+10, t12);
 804 }
 805 #endif
 806
 807
 808
 809
 810 static gmx_inline void
 811 gmx_mm_update_iforce_1atom_swizzle_pd(__m128d fix1, __m128d fiy1, __m128d fiz1,
 812                                       double * gmx_restrict fptr,
 813                                       double * gmx_restrict fshiftptr)
 814 {
 815     fix1 = _mm_hadd_pd(fix1, fiy1);
 816     fiz1 = _mm_hadd_pd(fiz1, fiz1);
 817
 818     _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr), fix1 ));
 819     _mm_store_sd( fptr+2, _mm_add_sd( _mm_load_sd(fptr+2), fiz1 ));
 820
 821     _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 ));
 822     _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 ));
 823 }
 824
 825
 826 #if defined (_MSC_VER) && defined(_M_IX86)
 827 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 828 #define gmx_mm_update_iforce_3atom_swizzle_pd(fix1, fiy1, fiz1, fix2, fiy2, fiz2, fix3, fiy3, fiz3, \
 829                                               fptr, fshiftptr) \
 830     { \
 831         __m128d _t1, _t2; \
 832         fix1 = _mm_hadd_pd(fix1, fiy1); \
 833         fiz1 = _mm_hadd_pd(fiz1, fix2); \
 834         fiy2 = _mm_hadd_pd(fiy2, fiz2); \
 835         fix3 = _mm_hadd_pd(fix3, fiy3); \
 836         fiz3 = _mm_hadd_pd(fiz3, fiz3); \
 837         _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr), fix1 )); \
 838         _mm_storeu_pd( fptr+2, _mm_add_pd( _mm_loadu_pd(fptr+2), fiz1 )); \
 839         _mm_storeu_pd( fptr+4, _mm_add_pd( _mm_loadu_pd(fptr+4), fiy2 )); \
 840         _mm_storeu_pd( fptr+6, _mm_add_pd( _mm_loadu_pd(fptr+6), fix3 )); \
 841         _mm_store_sd( fptr+8, _mm_add_sd( _mm_load_sd(fptr+8), fiz3 )); \
 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); \
 845         _t2   = _mm_shuffle_pd(fiy2, fiy2, _MM_SHUFFLE2(1, 1)); \
 846         fiz1  = _mm_add_sd(fiz1, fiz3); \
 847         fiz1  = _mm_add_sd(fiz1, _t2); \
 848         _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 )); \
 849         _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 )); \
 850     }
 851 #else
 852 /* Real function for sane compilers */
 853 static gmx_inline void
 854 gmx_mm_update_iforce_3atom_swizzle_pd(__m128d fix1, __m128d fiy1, __m128d fiz1,
 855                                       __m128d fix2, __m128d fiy2, __m128d fiz2,
 856                                       __m128d fix3, __m128d fiy3, __m128d fiz3,
 857                                       double * gmx_restrict fptr,
 858                                       double * gmx_restrict fshiftptr)
 859 {
 860     __m128d t1, t2;
 861
 862     fix1 = _mm_hadd_pd(fix1, fiy1);
 863     fiz1 = _mm_hadd_pd(fiz1, fix2);
 864     fiy2 = _mm_hadd_pd(fiy2, fiz2);
 865     fix3 = _mm_hadd_pd(fix3, fiy3);
 866     fiz3 = _mm_hadd_pd(fiz3, fiz3);
 867
 868     _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr), fix1 ));
 869     _mm_storeu_pd( fptr+2, _mm_add_pd( _mm_loadu_pd(fptr+2), fiz1 ));
 870     _mm_storeu_pd( fptr+4, _mm_add_pd( _mm_loadu_pd(fptr+4), fiy2 ));
 871     _mm_storeu_pd( fptr+6, _mm_add_pd( _mm_loadu_pd(fptr+6), fix3 ));
 872     _mm_store_sd( fptr+8, _mm_add_sd( _mm_load_sd(fptr+8), fiz3 ));
 873
 874     fix1 = _mm_add_pd(fix1, fix3);
 875     t1   = _mm_shuffle_pd(fiz1, fiy2, _MM_SHUFFLE2(0, 1));
 876     fix1 = _mm_add_pd(fix1, t1); /* x and y sums */
 877
 878     t2   = _mm_shuffle_pd(fiy2, fiy2, _MM_SHUFFLE2(1, 1));
 879     fiz1 = _mm_add_sd(fiz1, fiz3);
 880     fiz1 = _mm_add_sd(fiz1, t2); /* z sum */
 881
 882     _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 ));
 883     _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 ));
 884 }
 885 #endif
 886
 887 #if defined (_MSC_VER) && defined(_M_IX86)
 888 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 889 #define gmx_mm_update_iforce_4atom_swizzle_pd(fix1, fiy1, fiz1, fix2, fiy2, fiz2, fix3, fiy3, fiz3, fix4, fiy4, fiz4, \
 890                                               fptr, fshiftptr) \
 891     { \
 892         __m128d _t1, _t2; \
 893         fix1 = _mm_hadd_pd(fix1, fiy1); \
 894         fiz1 = _mm_hadd_pd(fiz1, fix2); \
 895         fiy2 = _mm_hadd_pd(fiy2, fiz2); \
 896         fix3 = _mm_hadd_pd(fix3, fiy3); \
 897         fiz3 = _mm_hadd_pd(fiz3, fix4); \
 898         fiy4 = _mm_hadd_pd(fiy4, fiz4); \
 899         _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr),       fix1 )); \
 900         _mm_storeu_pd( fptr+2, _mm_add_pd( _mm_loadu_pd(fptr+2),   fiz1 )); \
 901         _mm_storeu_pd( fptr+4, _mm_add_pd( _mm_loadu_pd(fptr+4),   fiy2 )); \
 902         _mm_storeu_pd( fptr+6, _mm_add_pd( _mm_loadu_pd(fptr+6),   fix3 )); \
 903         _mm_storeu_pd( fptr+8, _mm_add_pd( _mm_loadu_pd(fptr+8),   fiz3 )); \
 904         _mm_storeu_pd( fptr+10, _mm_add_pd( _mm_loadu_pd(fptr+10), fiy4 )); \
 905         _t1  = _mm_shuffle_pd(fiz1, fiy2, _MM_SHUFFLE2(0, 1)); \
 906         fix1 = _mm_add_pd(fix1, _t1); \
 907         _t2  = _mm_shuffle_pd(fiz3, fiy4, _MM_SHUFFLE2(0, 1)); \
 908         fix3 = _mm_add_pd(fix3, _t2); \
 909         fix1 = _mm_add_pd(fix1, fix3); \
 910         fiz1 = _mm_add_sd(fiz1, _mm_unpackhi_pd(fiy2, fiy2)); \
 911         fiz3 = _mm_add_sd(fiz3, _mm_unpackhi_pd(fiy4, fiy4)); \
 912         fiz1 = _mm_add_sd(fiz1, fiz3); \
 913         _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 )); \
 914         _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 )); \
 915     }
 916 #else
 917 /* Real function for sane compilers */
 918 static gmx_inline void
 919 gmx_mm_update_iforce_4atom_swizzle_pd(__m128d fix1, __m128d fiy1, __m128d fiz1,
 920                                       __m128d fix2, __m128d fiy2, __m128d fiz2,
 921                                       __m128d fix3, __m128d fiy3, __m128d fiz3,
 922                                       __m128d fix4, __m128d fiy4, __m128d fiz4,
 923                                       double * gmx_restrict fptr,
 924                                       double * gmx_restrict fshiftptr)
 925 {
 926     __m128d t1, t2;
 927
 928     fix1 = _mm_hadd_pd(fix1, fiy1);
 929     fiz1 = _mm_hadd_pd(fiz1, fix2);
 930     fiy2 = _mm_hadd_pd(fiy2, fiz2);
 931     fix3 = _mm_hadd_pd(fix3, fiy3);
 932     fiz3 = _mm_hadd_pd(fiz3, fix4);
 933     fiy4 = _mm_hadd_pd(fiy4, fiz4);
 934
 935     _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr),       fix1 ));
 936     _mm_storeu_pd( fptr+2, _mm_add_pd( _mm_loadu_pd(fptr+2),   fiz1 ));
 937     _mm_storeu_pd( fptr+4, _mm_add_pd( _mm_loadu_pd(fptr+4),   fiy2 ));
 938     _mm_storeu_pd( fptr+6, _mm_add_pd( _mm_loadu_pd(fptr+6),   fix3 ));
 939     _mm_storeu_pd( fptr+8, _mm_add_pd( _mm_loadu_pd(fptr+8),   fiz3 ));
 940     _mm_storeu_pd( fptr+10, _mm_add_pd( _mm_loadu_pd(fptr+10), fiy4 ));
 941
 942     t1   = _mm_shuffle_pd(fiz1, fiy2, _MM_SHUFFLE2(0, 1));
 943     fix1 = _mm_add_pd(fix1, t1);
 944     t2   = _mm_shuffle_pd(fiz3, fiy4, _MM_SHUFFLE2(0, 1));
 945     fix3 = _mm_add_pd(fix3, t2);
 946     fix1 = _mm_add_pd(fix1, fix3); /* x and y sums */
 947
 948     fiz1 = _mm_add_sd(fiz1, _mm_unpackhi_pd(fiy2, fiy2));
 949     fiz3 = _mm_add_sd(fiz3, _mm_unpackhi_pd(fiy4, fiy4));
 950     fiz1 = _mm_add_sd(fiz1, fiz3); /* z sum */
 951
 952     _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 ));
 953     _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 ));
 954 }
 955 #endif
 956
 957 static gmx_inline void
 958 gmx_mm_update_1pot_pd(__m128d pot1, double * gmx_restrict ptrA)
 959 {
 960     pot1 = _mm_hadd_pd(pot1, pot1);
 961     _mm_store_sd(ptrA, _mm_add_sd(pot1, _mm_load_sd(ptrA)));
 962 }
 963
 964 static gmx_inline void
 965 gmx_mm_update_2pot_pd(__m128d pot1, double * gmx_restrict ptrA,
 966                       __m128d pot2, double * gmx_restrict ptrB)
 967 {
 968     pot1 = _mm_hadd_pd(pot1, pot2);
 969     pot2 = _mm_unpackhi_pd(pot1, pot1);
 970
 971     _mm_store_sd(ptrA, _mm_add_sd(pot1, _mm_load_sd(ptrA)));
 972     _mm_store_sd(ptrB, _mm_add_sd(pot2, _mm_load_sd(ptrB)));
 973 }
 974
 975
 976 #endif /* _kernelutil_x86_sse4_1_double_h_ */