src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/kernelutil_x86_sse4_1_single.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_single_h_
  36 #define _kernelutil_x86_sse4_1_single_h_
  37
  38 #include <math.h>
  39
  40 #undef gmx_restrict
  41 #define gmx_restrict
  42
  43 #define gmx_mm_castsi128_ps   _mm_castsi128_ps
  44 #define gmx_mm_extract_epi32  _mm_extract_epi32
  45
  46 /* Normal sum of four xmm registers */
  47 #define gmx_mm_sum4_ps(t0, t1, t2, t3)  _mm_add_ps(_mm_add_ps(t0, t1), _mm_add_ps(t2, t3))
  48
  49 static gmx_inline __m128
  50 gmx_mm_calc_rsq_ps(__m128 dx, __m128 dy, __m128 dz)
  51 {
  52     return _mm_add_ps( _mm_add_ps( _mm_mul_ps(dx, dx), _mm_mul_ps(dy, dy) ), _mm_mul_ps(dz, dz) );
  53 }
  54
  55 static gmx_inline int
  56 gmx_mm_any_lt(__m128 a, __m128 b)
  57 {
  58     return _mm_movemask_ps(_mm_cmplt_ps(a, b));
  59 }
  60
  61 /* Load a single value from 1-4 places, merge into xmm register */
  62
  63 static gmx_inline __m128
  64 gmx_mm_load_4real_swizzle_ps(const float * gmx_restrict ptrA,
  65                              const float * gmx_restrict ptrB,
  66                              const float * gmx_restrict ptrC,
  67                              const float * gmx_restrict ptrD)
  68 {
  69     __m128 t1, t2;
  70
  71     t1 = _mm_unpacklo_ps(_mm_load_ss(ptrA), _mm_load_ss(ptrC));
  72     t2 = _mm_unpacklo_ps(_mm_load_ss(ptrB), _mm_load_ss(ptrD));
  73     return _mm_unpacklo_ps(t1, t2);
  74 }
  75
  76 static gmx_inline void
  77 gmx_mm_store_4real_swizzle_ps(float * gmx_restrict ptrA,
  78                               float * gmx_restrict ptrB,
  79                               float * gmx_restrict ptrC,
  80                               float * gmx_restrict ptrD,
  81                               __m128               xmm1)
  82 {
  83     __m128 t2, t3, t4;
  84
  85     t3       = _mm_movehl_ps(_mm_setzero_ps(), xmm1);
  86     t2       = _mm_shuffle_ps(xmm1, xmm1, _MM_SHUFFLE(1, 1, 1, 1));
  87     t4       = _mm_shuffle_ps(t3, t3, _MM_SHUFFLE(1, 1, 1, 1));
  88     _mm_store_ss(ptrA, xmm1);
  89     _mm_store_ss(ptrB, t2);
  90     _mm_store_ss(ptrC, t3);
  91     _mm_store_ss(ptrD, t4);
  92 }
  93
  94 /* Similar to store, but increments value in memory */
  95 static gmx_inline void
  96 gmx_mm_increment_4real_swizzle_ps(float * gmx_restrict ptrA,
  97                                   float * gmx_restrict ptrB,
  98                                   float * gmx_restrict ptrC,
  99                                   float * gmx_restrict ptrD, __m128 xmm1)
 100 {
 101     __m128 tmp;
 102
 103     tmp = gmx_mm_load_4real_swizzle_ps(ptrA, ptrB, ptrC, ptrD);
 104     tmp = _mm_add_ps(tmp, xmm1);
 105     gmx_mm_store_4real_swizzle_ps(ptrA, ptrB, ptrC, ptrD, tmp);
 106 }
 107
 108
 109 static gmx_inline void
 110 gmx_mm_load_4pair_swizzle_ps(const float * gmx_restrict p1,
 111                              const float * gmx_restrict p2,
 112                              const float * gmx_restrict p3,
 113                              const float * gmx_restrict p4,
 114                              __m128 * gmx_restrict      c6,
 115                              __m128 * gmx_restrict      c12)
 116 {
 117     __m128 t1, t2, t3, t4;
 118
 119     t1   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)p1);   /* - - c12a  c6a */
 120     t2   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)p2);   /* - - c12b  c6b */
 121     t3   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)p3);   /* - - c12c  c6c */
 122     t4   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)p4);   /* - - c12d  c6d */
 123     t1   = _mm_unpacklo_ps(t1, t2);
 124     t2   = _mm_unpacklo_ps(t3, t4);
 125     *c6  = _mm_movelh_ps(t1, t2);
 126     *c12 = _mm_movehl_ps(t2, t1);
 127 }
 128
 129
 130 static gmx_inline void
 131 gmx_mm_load_shift_and_1rvec_broadcast_ps(const float * gmx_restrict xyz_shift,
 132                                          const float * gmx_restrict xyz,
 133                                          __m128 * gmx_restrict      x1,
 134                                          __m128 * gmx_restrict      y1,
 135                                          __m128 * gmx_restrict      z1)
 136 {
 137     __m128 t1, t2, t3, t4;
 138
 139     t1   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)xyz_shift);
 140     t2   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)xyz);
 141     t3   = _mm_load_ss(xyz_shift+2);
 142     t4   = _mm_load_ss(xyz+2);
 143     t1   = _mm_add_ps(t1, t2);
 144     t3   = _mm_add_ss(t3, t4);
 145
 146     *x1  = _mm_shuffle_ps(t1, t1, _MM_SHUFFLE(0, 0, 0, 0));
 147     *y1  = _mm_shuffle_ps(t1, t1, _MM_SHUFFLE(1, 1, 1, 1));
 148     *z1  = _mm_shuffle_ps(t3, t3, _MM_SHUFFLE(0, 0, 0, 0));
 149 }
 150
 151
 152 static gmx_inline void
 153 gmx_mm_load_shift_and_3rvec_broadcast_ps(const float * gmx_restrict xyz_shift,
 154                                          const float * gmx_restrict xyz,
 155                                          __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 156                                          __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 157                                          __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3)
 158 {
 159     __m128 tA, tB;
 160     __m128 t1, t2, t3, t4, t5, t6;
 161
 162     tA   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)xyz_shift);
 163     tB   = _mm_load_ss(xyz_shift+2);
 164
 165     t1   = _mm_loadu_ps(xyz);
 166     t2   = _mm_loadu_ps(xyz+4);
 167     t3   = _mm_load_ss(xyz+8);
 168
 169     tA   = _mm_movelh_ps(tA, tB);
 170     t4   = _mm_shuffle_ps(tA, tA, _MM_SHUFFLE(0, 2, 1, 0));
 171     t5   = _mm_shuffle_ps(tA, tA, _MM_SHUFFLE(1, 0, 2, 1));
 172     t6   = _mm_shuffle_ps(tA, tA, _MM_SHUFFLE(2, 1, 0, 2));
 173
 174     t1   = _mm_add_ps(t1, t4);
 175     t2   = _mm_add_ps(t2, t5);
 176     t3   = _mm_add_ss(t3, t6);
 177
 178     *x1  = _mm_shuffle_ps(t1, t1, _MM_SHUFFLE(0, 0, 0, 0));
 179     *y1  = _mm_shuffle_ps(t1, t1, _MM_SHUFFLE(1, 1, 1, 1));
 180     *z1  = _mm_shuffle_ps(t1, t1, _MM_SHUFFLE(2, 2, 2, 2));
 181     *x2  = _mm_shuffle_ps(t1, t1, _MM_SHUFFLE(3, 3, 3, 3));
 182     *y2  = _mm_shuffle_ps(t2, t2, _MM_SHUFFLE(0, 0, 0, 0));
 183     *z2  = _mm_shuffle_ps(t2, t2, _MM_SHUFFLE(1, 1, 1, 1));
 184     *x3  = _mm_shuffle_ps(t2, t2, _MM_SHUFFLE(2, 2, 2, 2));
 185     *y3  = _mm_shuffle_ps(t2, t2, _MM_SHUFFLE(3, 3, 3, 3));
 186     *z3  = _mm_shuffle_ps(t3, t3, _MM_SHUFFLE(0, 0, 0, 0));
 187 }
 188
 189
 190 static gmx_inline void
 191 gmx_mm_load_shift_and_4rvec_broadcast_ps(const float * gmx_restrict xyz_shift,
 192                                          const float * gmx_restrict xyz,
 193                                          __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 194                                          __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 195                                          __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3,
 196                                          __m128 * gmx_restrict x4, __m128 * gmx_restrict y4, __m128 * gmx_restrict z4)
 197 {
 198     __m128 tA, tB;
 199     __m128 t1, t2, t3, t4, t5, t6;
 200
 201     tA   = _mm_castpd_ps(_mm_load_sd((const double *)xyz_shift));
 202     tB   = _mm_load_ss(xyz_shift+2);
 203
 204     t1   = _mm_loadu_ps(xyz);
 205     t2   = _mm_loadu_ps(xyz+4);
 206     t3   = _mm_loadu_ps(xyz+8);
 207
 208     tA   = _mm_movelh_ps(tA, tB);
 209     t4   = _mm_shuffle_ps(tA, tA, _MM_SHUFFLE(0, 2, 1, 0));
 210     t5   = _mm_shuffle_ps(tA, tA, _MM_SHUFFLE(1, 0, 2, 1));
 211     t6   = _mm_shuffle_ps(tA, tA, _MM_SHUFFLE(2, 1, 0, 2));
 212
 213     t1   = _mm_add_ps(t1, t4);
 214     t2   = _mm_add_ps(t2, t5);
 215     t3   = _mm_add_ps(t3, t6);
 216
 217     *x1  = _mm_shuffle_ps(t1, t1, _MM_SHUFFLE(0, 0, 0, 0));
 218     *y1  = _mm_shuffle_ps(t1, t1, _MM_SHUFFLE(1, 1, 1, 1));
 219     *z1  = _mm_shuffle_ps(t1, t1, _MM_SHUFFLE(2, 2, 2, 2));
 220     *x2  = _mm_shuffle_ps(t1, t1, _MM_SHUFFLE(3, 3, 3, 3));
 221     *y2  = _mm_shuffle_ps(t2, t2, _MM_SHUFFLE(0, 0, 0, 0));
 222     *z2  = _mm_shuffle_ps(t2, t2, _MM_SHUFFLE(1, 1, 1, 1));
 223     *x3  = _mm_shuffle_ps(t2, t2, _MM_SHUFFLE(2, 2, 2, 2));
 224     *y3  = _mm_shuffle_ps(t2, t2, _MM_SHUFFLE(3, 3, 3, 3));
 225     *z3  = _mm_shuffle_ps(t3, t3, _MM_SHUFFLE(0, 0, 0, 0));
 226     *x4  = _mm_shuffle_ps(t3, t3, _MM_SHUFFLE(1, 1, 1, 1));
 227     *y4  = _mm_shuffle_ps(t3, t3, _MM_SHUFFLE(2, 2, 2, 2));
 228     *z4  = _mm_shuffle_ps(t3, t3, _MM_SHUFFLE(3, 3, 3, 3));
 229 }
 230
 231
 232 static gmx_inline void
 233 gmx_mm_load_1rvec_4ptr_swizzle_ps(const float * gmx_restrict ptrA,
 234                                   const float * gmx_restrict ptrB,
 235                                   const float * gmx_restrict ptrC,
 236                                   const float * gmx_restrict ptrD,
 237                                   __m128 *      gmx_restrict x1,
 238                                   __m128 *      gmx_restrict y1,
 239                                   __m128 *      gmx_restrict z1)
 240 {
 241     __m128 t1, t2, t3, t4, t5, t6, t7, t8;
 242     t1   = _mm_castpd_ps(_mm_load_sd((const double *)ptrA));
 243     t2   = _mm_castpd_ps(_mm_load_sd((const double *)ptrB));
 244     t3   = _mm_castpd_ps(_mm_load_sd((const double *)ptrC));
 245     t4   = _mm_castpd_ps(_mm_load_sd((const double *)ptrD));
 246     t5   = _mm_load_ss(ptrA+2);
 247     t6   = _mm_load_ss(ptrB+2);
 248     t7   = _mm_load_ss(ptrC+2);
 249     t8   = _mm_load_ss(ptrD+2);
 250     t1   = _mm_unpacklo_ps(t1, t2);
 251     t3   = _mm_unpacklo_ps(t3, t4);
 252     *x1  = _mm_movelh_ps(t1, t3);
 253     *y1  = _mm_movehl_ps(t3, t1);
 254     t5   = _mm_unpacklo_ps(t5, t6);
 255     t7   = _mm_unpacklo_ps(t7, t8);
 256     *z1  = _mm_movelh_ps(t5, t7);
 257 }
 258
 259
 260 static gmx_inline void
 261 gmx_mm_load_3rvec_4ptr_swizzle_ps(const float * gmx_restrict ptrA,
 262                                   const float * gmx_restrict ptrB,
 263                                   const float * gmx_restrict ptrC,
 264                                   const float * gmx_restrict ptrD,
 265                                   __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 266                                   __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 267                                   __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3)
 268 {
 269     __m128 t1, t2, t3, t4;
 270     t1            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)ptrA ) );
 271     t2            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)ptrB ) );
 272     t3            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)ptrC ) );
 273     t4            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)ptrD ) );
 274     _MM_TRANSPOSE4_PS(t1, t2, t3, t4);
 275     *x1           = t1;
 276     *y1           = t2;
 277     *z1           = t3;
 278     *x2           = t4;
 279     t1            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrA+4) ) );
 280     t2            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrB+4) ) );
 281     t3            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrC+4) ) );
 282     t4            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrD+4) ) );
 283     _MM_TRANSPOSE4_PS(t1, t2, t3, t4);
 284     *y2           = t1;
 285     *z2           = t2;
 286     *x3           = t3;
 287     *y3           = t4;
 288     t1            = _mm_load_ss(ptrA+8);
 289     t2            = _mm_load_ss(ptrB+8);
 290     t3            = _mm_load_ss(ptrC+8);
 291     t4            = _mm_load_ss(ptrD+8);
 292     t1            = _mm_unpacklo_ps(t1, t3);
 293     t3            = _mm_unpacklo_ps(t2, t4);
 294     *z3           = _mm_unpacklo_ps(t1, t3);
 295 }
 296
 297
 298 static gmx_inline void
 299 gmx_mm_load_4rvec_4ptr_swizzle_ps(const float * gmx_restrict ptrA,
 300                                   const float * gmx_restrict ptrB,
 301                                   const float * gmx_restrict ptrC,
 302                                   const float * gmx_restrict ptrD,
 303                                   __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 304                                   __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 305                                   __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3,
 306                                   __m128 * gmx_restrict x4, __m128 * gmx_restrict y4, __m128 * gmx_restrict z4)
 307 {
 308     __m128 t1, t2, t3, t4;
 309     t1            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrA) ) );
 310     t2            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrB) ) );
 311     t3            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrC) ) );
 312     t4            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrD) ) );
 313     _MM_TRANSPOSE4_PS(t1, t2, t3, t4);
 314     *x1           = t1;
 315     *y1           = t2;
 316     *z1           = t3;
 317     *x2           = t4;
 318     t1            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrA+4) ) );
 319     t2            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrB+4) ) );
 320     t3            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrC+4) ) );
 321     t4            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrD+4) ) );
 322     _MM_TRANSPOSE4_PS(t1, t2, t3, t4);
 323     *y2           = t1;
 324     *z2           = t2;
 325     *x3           = t3;
 326     *y3           = t4;
 327     t1            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrA+8) ) );
 328     t2            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrB+8) ) );
 329     t3            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrC+8) ) );
 330     t4            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrD+8) ) );
 331     _MM_TRANSPOSE4_PS(t1, t2, t3, t4);
 332     *z3           = t1;
 333     *x4           = t2;
 334     *y4           = t3;
 335     *z4           = t4;
 336 }
 337
 338
 339
 340 static gmx_inline void
 341 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(float * ptrA,
 342                                        float * ptrB,
 343                                        float * ptrC,
 344                                        float * ptrD,
 345                                        __m128 x1, __m128 y1, __m128 z1)
 346 {
 347     __m128 t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12;
 348     t5          = _mm_unpacklo_ps(y1, z1);
 349     t6          = _mm_unpackhi_ps(y1, z1);
 350     t7          = _mm_shuffle_ps(x1, t5, _MM_SHUFFLE(1, 0, 0, 0));
 351     t8          = _mm_shuffle_ps(x1, t5, _MM_SHUFFLE(3, 2, 0, 1));
 352     t9          = _mm_shuffle_ps(x1, t6, _MM_SHUFFLE(1, 0, 0, 2));
 353     t10         = _mm_shuffle_ps(x1, t6, _MM_SHUFFLE(3, 2, 0, 3));
 354     t1          = _mm_load_ss(ptrA);
 355     t1          = _mm_loadh_pi(t1, (__m64 *)(ptrA+1));
 356     t1          = _mm_sub_ps(t1, t7);
 357     _mm_store_ss(ptrA, t1);
 358     _mm_storeh_pi((__m64 *)(ptrA+1), t1);
 359     t2          = _mm_load_ss(ptrB);
 360     t2          = _mm_loadh_pi(t2, (__m64 *)(ptrB+1));
 361     t2          = _mm_sub_ps(t2, t8);
 362     _mm_store_ss(ptrB, t2);
 363     _mm_storeh_pi((__m64 *)(ptrB+1), t2);
 364     t3          = _mm_load_ss(ptrC);
 365     t3          = _mm_loadh_pi(t3, (__m64 *)(ptrC+1));
 366     t3          = _mm_sub_ps(t3, t9);
 367     _mm_store_ss(ptrC, t3);
 368     _mm_storeh_pi((__m64 *)(ptrC+1), t3);
 369     t4          = _mm_load_ss(ptrD);
 370     t4          = _mm_loadh_pi(t4, (__m64 *)(ptrD+1));
 371     t4          = _mm_sub_ps(t4, t10);
 372     _mm_store_ss(ptrD, t4);
 373     _mm_storeh_pi((__m64 *)(ptrD+1), t4);
 374 }
 375
 376
 377
 378 #if defined (_MSC_VER) && defined(_M_IX86)
 379 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 380 #define gmx_mm_decrement_3rvec_4ptr_swizzle_ps(ptrA, ptrB, ptrC, ptrD, \
 381                                                _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3) \
 382     { \
 383         __m128 _t1, _t2, _t3, _t4, _t5, _t6, _t7, _t8, _t9, _t10; \
 384         __m128 _t11, _t12, _t13, _t14, _t15, _t16, _t17, _t18, _t19; \
 385         __m128 _t20, _t21, _t22, _t23, _t24, _t25; \
 386         _t13         = _mm_unpackhi_ps(_x1, _y1); \
 387         _x1          = _mm_unpacklo_ps(_x1, _y1); \
 388         _t14         = _mm_unpackhi_ps(_z1, _x2); \
 389         _z1          = _mm_unpacklo_ps(_z1, _x2); \
 390         _t15         = _mm_unpackhi_ps(_y2, _z2); \
 391         _y2          = _mm_unpacklo_ps(_y2, _z2); \
 392         _t16         = _mm_unpackhi_ps(_x3, _y3); \
 393         _x3          = _mm_unpacklo_ps(_x3, _y3); \
 394         _t17         = _mm_shuffle_ps(_z3, _z3, _MM_SHUFFLE(0, 0, 0, 1)); \
 395         _t18         = _mm_movehl_ps(_z3, _z3); \
 396         _t19         = _mm_shuffle_ps(_t18, _t18, _MM_SHUFFLE(0, 0, 0, 1)); \
 397         _t20         = _mm_movelh_ps(_x1, _z1); \
 398         _t21         = _mm_movehl_ps(_z1, _x1); \
 399         _t22         = _mm_movelh_ps(_t13, _t14); \
 400         _t14         = _mm_movehl_ps(_t14, _t13); \
 401         _t23         = _mm_movelh_ps(_y2, _x3); \
 402         _t24         = _mm_movehl_ps(_x3, _y2); \
 403         _t25         = _mm_movelh_ps(_t15, _t16); \
 404         _t16         = _mm_movehl_ps(_t16, _t15); \
 405         _t1          = _mm_loadu_ps(ptrA); \
 406         _t2          = _mm_loadu_ps(ptrA+4); \
 407         _t3          = _mm_load_ss(ptrA+8); \
 408         _t1          = _mm_sub_ps(_t1, _t20); \
 409         _t2          = _mm_sub_ps(_t2, _t23); \
 410         _t3          = _mm_sub_ss(_t3, _z3); \
 411         _mm_storeu_ps(ptrA, _t1); \
 412         _mm_storeu_ps(ptrA+4, _t2); \
 413         _mm_store_ss(ptrA+8, _t3); \
 414         _t4          = _mm_loadu_ps(ptrB); \
 415         _t5          = _mm_loadu_ps(ptrB+4); \
 416         _t6          = _mm_load_ss(ptrB+8); \
 417         _t4          = _mm_sub_ps(_t4, _t21); \
 418         _t5          = _mm_sub_ps(_t5, _t24); \
 419         _t6          = _mm_sub_ss(_t6, _t17); \
 420         _mm_storeu_ps(ptrB, _t4); \
 421         _mm_storeu_ps(ptrB+4, _t5); \
 422         _mm_store_ss(ptrB+8, _t6); \
 423         _t7          = _mm_loadu_ps(ptrC); \
 424         _t8          = _mm_loadu_ps(ptrC+4); \
 425         _t9          = _mm_load_ss(ptrC+8); \
 426         _t7          = _mm_sub_ps(_t7, _t22); \
 427         _t8          = _mm_sub_ps(_t8, _t25); \
 428         _t9          = _mm_sub_ss(_t9, _t18); \
 429         _mm_storeu_ps(ptrC, _t7); \
 430         _mm_storeu_ps(ptrC+4, _t8); \
 431         _mm_store_ss(ptrC+8, _t9); \
 432         _t10         = _mm_loadu_ps(ptrD); \
 433         _t11         = _mm_loadu_ps(ptrD+4); \
 434         _t12         = _mm_load_ss(ptrD+8); \
 435         _t10         = _mm_sub_ps(_t10, _t14); \
 436         _t11         = _mm_sub_ps(_t11, _t16); \
 437         _t12         = _mm_sub_ss(_t12, _t19); \
 438         _mm_storeu_ps(ptrD, _t10); \
 439         _mm_storeu_ps(ptrD+4, _t11); \
 440         _mm_store_ss(ptrD+8, _t12); \
 441     }
 442 #else
 443 /* Real function for sane compilers */
 444 static gmx_inline void
 445 gmx_mm_decrement_3rvec_4ptr_swizzle_ps(float * gmx_restrict ptrA, float * gmx_restrict ptrB,
 446                                        float * gmx_restrict ptrC, float * gmx_restrict ptrD,
 447                                        __m128 x1, __m128 y1, __m128 z1,
 448                                        __m128 x2, __m128 y2, __m128 z2,
 449                                        __m128 x3, __m128 y3, __m128 z3)
 450 {
 451     __m128 t1, t2, t3, t4, t5, t6, t7, t8, t9, t10;
 452     __m128 t11, t12, t13, t14, t15, t16, t17, t18, t19;
 453     __m128 t20, t21, t22, t23, t24, t25;
 454
 455     t13         = _mm_unpackhi_ps(x1, y1);
 456     x1          = _mm_unpacklo_ps(x1, y1);
 457     t14         = _mm_unpackhi_ps(z1, x2);
 458     z1          = _mm_unpacklo_ps(z1, x2);
 459     t15         = _mm_unpackhi_ps(y2, z2);
 460     y2          = _mm_unpacklo_ps(y2, z2);
 461     t16         = _mm_unpackhi_ps(x3, y3);
 462     x3          = _mm_unpacklo_ps(x3, y3);
 463     t17         = _mm_shuffle_ps(z3, z3, _MM_SHUFFLE(0, 0, 0, 1));
 464     t18         = _mm_movehl_ps(z3, z3);
 465     t19         = _mm_shuffle_ps(t18, t18, _MM_SHUFFLE(0, 0, 0, 1));
 466     t20         = _mm_movelh_ps(x1, z1);
 467     t21         = _mm_movehl_ps(z1, x1);
 468     t22         = _mm_movelh_ps(t13, t14);
 469     t14         = _mm_movehl_ps(t14, t13);
 470     t23         = _mm_movelh_ps(y2, x3);
 471     t24         = _mm_movehl_ps(x3, y2);
 472     t25         = _mm_movelh_ps(t15, t16);
 473     t16         = _mm_movehl_ps(t16, t15);
 474     t1          = _mm_loadu_ps(ptrA);
 475     t2          = _mm_loadu_ps(ptrA+4);
 476     t3          = _mm_load_ss(ptrA+8);
 477     t4          = _mm_loadu_ps(ptrB);
 478     t5          = _mm_loadu_ps(ptrB+4);
 479     t6          = _mm_load_ss(ptrB+8);
 480     t7          = _mm_loadu_ps(ptrC);
 481     t8          = _mm_loadu_ps(ptrC+4);
 482     t9          = _mm_load_ss(ptrC+8);
 483     t10         = _mm_loadu_ps(ptrD);
 484     t11         = _mm_loadu_ps(ptrD+4);
 485     t12         = _mm_load_ss(ptrD+8);
 486
 487     t1          = _mm_sub_ps(t1, t20);
 488     t2          = _mm_sub_ps(t2, t23);
 489     t3          = _mm_sub_ss(t3, z3);
 490     _mm_storeu_ps(ptrA, t1);
 491     _mm_storeu_ps(ptrA+4, t2);
 492     _mm_store_ss(ptrA+8, t3);
 493     t4          = _mm_sub_ps(t4, t21);
 494     t5          = _mm_sub_ps(t5, t24);
 495     t6          = _mm_sub_ss(t6, t17);
 496     _mm_storeu_ps(ptrB, t4);
 497     _mm_storeu_ps(ptrB+4, t5);
 498     _mm_store_ss(ptrB+8, t6);
 499     t7          = _mm_sub_ps(t7, t22);
 500     t8          = _mm_sub_ps(t8, t25);
 501     t9          = _mm_sub_ss(t9, t18);
 502     _mm_storeu_ps(ptrC, t7);
 503     _mm_storeu_ps(ptrC+4, t8);
 504     _mm_store_ss(ptrC+8, t9);
 505     t10         = _mm_sub_ps(t10, t14);
 506     t11         = _mm_sub_ps(t11, t16);
 507     t12         = _mm_sub_ss(t12, t19);
 508     _mm_storeu_ps(ptrD, t10);
 509     _mm_storeu_ps(ptrD+4, t11);
 510     _mm_store_ss(ptrD+8, t12);
 511 }
 512 #endif
 513
 514 #if defined (_MSC_VER) && defined(_M_IX86)
 515 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 516 #define gmx_mm_decrement_4rvec_4ptr_swizzle_ps(ptrA, ptrB, ptrC, ptrD, \
 517                                                _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3, _x4, _y4, _z4) \
 518     { \
 519         __m128 _t1, _t2, _t3, _t4, _t5, _t6, _t7, _t8, _t9, _t10, _t11; \
 520         __m128 _t12, _t13, _t14, _t15, _t16, _t17, _t18, _t19, _t20, _t21, _t22; \
 521         __m128 _t23, _t24; \
 522         _t13         = _mm_unpackhi_ps(_x1, _y1); \
 523         _x1          = _mm_unpacklo_ps(_x1, _y1); \
 524         _t14         = _mm_unpackhi_ps(_z1, _x2); \
 525         _z1          = _mm_unpacklo_ps(_z1, _x2); \
 526         _t15         = _mm_unpackhi_ps(_y2, _z2); \
 527         _y2          = _mm_unpacklo_ps(_y2, _z2); \
 528         _t16         = _mm_unpackhi_ps(_x3, _y3); \
 529         _x3          = _mm_unpacklo_ps(_x3, _y3); \
 530         _t17         = _mm_unpackhi_ps(_z3, _x4); \
 531         _z3          = _mm_unpacklo_ps(_z3, _x4); \
 532         _t18         = _mm_unpackhi_ps(_y4, _z4); \
 533         _y4          = _mm_unpacklo_ps(_y4, _z4); \
 534         _t19         = _mm_movelh_ps(_x1, _z1); \
 535         _z1          = _mm_movehl_ps(_z1, _x1); \
 536         _t20         = _mm_movelh_ps(_t13, _t14); \
 537         _t14         = _mm_movehl_ps(_t14, _t13); \
 538         _t21         = _mm_movelh_ps(_y2, _x3); \
 539         _x3          = _mm_movehl_ps(_x3, _y2); \
 540         _t22         = _mm_movelh_ps(_t15, _t16); \
 541         _t16         = _mm_movehl_ps(_t16, _t15); \
 542         _t23         = _mm_movelh_ps(_z3, _y4); \
 543         _y4          = _mm_movehl_ps(_y4, _z3); \
 544         _t24         = _mm_movelh_ps(_t17, _t18); \
 545         _t18         = _mm_movehl_ps(_t18, _t17); \
 546         _t1          = _mm_loadu_ps(ptrA); \
 547         _t2          = _mm_loadu_ps(ptrA+4); \
 548         _t3          = _mm_loadu_ps(ptrA+8); \
 549         _t1          = _mm_sub_ps(_t1, _t19); \
 550         _t2          = _mm_sub_ps(_t2, _t21); \
 551         _t3          = _mm_sub_ps(_t3, _t23); \
 552         _mm_storeu_ps(ptrA, _t1); \
 553         _mm_storeu_ps(ptrA+4, _t2); \
 554         _mm_storeu_ps(ptrA+8, _t3); \
 555         _t4          = _mm_loadu_ps(ptrB); \
 556         _t5          = _mm_loadu_ps(ptrB+4); \
 557         _t6          = _mm_loadu_ps(ptrB+8); \
 558         _t4          = _mm_sub_ps(_t4, _z1); \
 559         _t5          = _mm_sub_ps(_t5, _x3); \
 560         _t6          = _mm_sub_ps(_t6, _y4); \
 561         _mm_storeu_ps(ptrB, _t4); \
 562         _mm_storeu_ps(ptrB+4, _t5); \
 563         _mm_storeu_ps(ptrB+8, _t6); \
 564         _t7          = _mm_loadu_ps(ptrC); \
 565         _t8          = _mm_loadu_ps(ptrC+4); \
 566         _t9          = _mm_loadu_ps(ptrC+8); \
 567         _t7          = _mm_sub_ps(_t7, _t20); \
 568         _t8          = _mm_sub_ps(_t8, _t22); \
 569         _t9          = _mm_sub_ps(_t9, _t24); \
 570         _mm_storeu_ps(ptrC, _t7); \
 571         _mm_storeu_ps(ptrC+4, _t8); \
 572         _mm_storeu_ps(ptrC+8, _t9); \
 573         _t10         = _mm_loadu_ps(ptrD); \
 574         _t11         = _mm_loadu_ps(ptrD+4); \
 575         _t12         = _mm_loadu_ps(ptrD+8); \
 576         _t10         = _mm_sub_ps(_t10, _t14); \
 577         _t11         = _mm_sub_ps(_t11, _t16); \
 578         _t12         = _mm_sub_ps(_t12, _t18); \
 579         _mm_storeu_ps(ptrD, _t10); \
 580         _mm_storeu_ps(ptrD+4, _t11); \
 581         _mm_storeu_ps(ptrD+8, _t12); \
 582     }
 583 #else
 584 /* Real function for sane compilers */
 585 static gmx_inline void
 586 gmx_mm_decrement_4rvec_4ptr_swizzle_ps(float * gmx_restrict ptrA, float * gmx_restrict ptrB,
 587                                        float * gmx_restrict ptrC, float * gmx_restrict ptrD,
 588                                        __m128 x1, __m128 y1, __m128 z1,
 589                                        __m128 x2, __m128 y2, __m128 z2,
 590                                        __m128 x3, __m128 y3, __m128 z3,
 591                                        __m128 x4, __m128 y4, __m128 z4)
 592 {
 593     __m128 t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11;
 594     __m128 t12, t13, t14, t15, t16, t17, t18, t19, t20, t21, t22;
 595     __m128 t23, t24;
 596     t13         = _mm_unpackhi_ps(x1, y1);
 597     x1          = _mm_unpacklo_ps(x1, y1);
 598     t14         = _mm_unpackhi_ps(z1, x2);
 599     z1          = _mm_unpacklo_ps(z1, x2);
 600     t15         = _mm_unpackhi_ps(y2, z2);
 601     y2          = _mm_unpacklo_ps(y2, z2);
 602     t16         = _mm_unpackhi_ps(x3, y3);
 603     x3          = _mm_unpacklo_ps(x3, y3);
 604     t17         = _mm_unpackhi_ps(z3, x4);
 605     z3          = _mm_unpacklo_ps(z3, x4);
 606     t18         = _mm_unpackhi_ps(y4, z4);
 607     y4          = _mm_unpacklo_ps(y4, z4);
 608     t19         = _mm_movelh_ps(x1, z1);
 609     z1          = _mm_movehl_ps(z1, x1);
 610     t20         = _mm_movelh_ps(t13, t14);
 611     t14         = _mm_movehl_ps(t14, t13);
 612     t21         = _mm_movelh_ps(y2, x3);
 613     x3          = _mm_movehl_ps(x3, y2);
 614     t22         = _mm_movelh_ps(t15, t16);
 615     t16         = _mm_movehl_ps(t16, t15);
 616     t23         = _mm_movelh_ps(z3, y4);
 617     y4          = _mm_movehl_ps(y4, z3);
 618     t24         = _mm_movelh_ps(t17, t18);
 619     t18         = _mm_movehl_ps(t18, t17);
 620     t1          = _mm_loadu_ps(ptrA);
 621     t2          = _mm_loadu_ps(ptrA+4);
 622     t3          = _mm_loadu_ps(ptrA+8);
 623     t1          = _mm_sub_ps(t1, t19);
 624     t2          = _mm_sub_ps(t2, t21);
 625     t3          = _mm_sub_ps(t3, t23);
 626     _mm_storeu_ps(ptrA, t1);
 627     _mm_storeu_ps(ptrA+4, t2);
 628     _mm_storeu_ps(ptrA+8, t3);
 629     t4          = _mm_loadu_ps(ptrB);
 630     t5          = _mm_loadu_ps(ptrB+4);
 631     t6          = _mm_loadu_ps(ptrB+8);
 632     t4          = _mm_sub_ps(t4, z1);
 633     t5          = _mm_sub_ps(t5, x3);
 634     t6          = _mm_sub_ps(t6, y4);
 635     _mm_storeu_ps(ptrB, t4);
 636     _mm_storeu_ps(ptrB+4, t5);
 637     _mm_storeu_ps(ptrB+8, t6);
 638     t7          = _mm_loadu_ps(ptrC);
 639     t8          = _mm_loadu_ps(ptrC+4);
 640     t9          = _mm_loadu_ps(ptrC+8);
 641     t7          = _mm_sub_ps(t7, t20);
 642     t8          = _mm_sub_ps(t8, t22);
 643     t9          = _mm_sub_ps(t9, t24);
 644     _mm_storeu_ps(ptrC, t7);
 645     _mm_storeu_ps(ptrC+4, t8);
 646     _mm_storeu_ps(ptrC+8, t9);
 647     t10         = _mm_loadu_ps(ptrD);
 648     t11         = _mm_loadu_ps(ptrD+4);
 649     t12         = _mm_loadu_ps(ptrD+8);
 650     t10         = _mm_sub_ps(t10, t14);
 651     t11         = _mm_sub_ps(t11, t16);
 652     t12         = _mm_sub_ps(t12, t18);
 653     _mm_storeu_ps(ptrD, t10);
 654     _mm_storeu_ps(ptrD+4, t11);
 655     _mm_storeu_ps(ptrD+8, t12);
 656 }
 657 #endif
 658
 659
 660 static gmx_inline void
 661 gmx_mm_update_iforce_1atom_swizzle_ps(__m128 fix1, __m128 fiy1, __m128 fiz1,
 662                                       float * gmx_restrict fptr,
 663                                       float * gmx_restrict fshiftptr)
 664 {
 665     __m128 t2, t3;
 666
 667     fix1 = _mm_hadd_ps(fix1, fix1);
 668     fiy1 = _mm_hadd_ps(fiy1, fiz1);
 669
 670     fix1 = _mm_hadd_ps(fix1, fiy1); /* fiz1 fiy1 fix1 fix1 */
 671
 672     t2 = _mm_load_ss(fptr);
 673     t2 = _mm_loadh_pi(t2, (__m64 *)(fptr+1));
 674     t3 = _mm_load_ss(fshiftptr);
 675     t3 = _mm_loadh_pi(t3, (__m64 *)(fshiftptr+1));
 676
 677     t2 = _mm_add_ps(t2, fix1);
 678     t3 = _mm_add_ps(t3, fix1);
 679
 680     _mm_store_ss(fptr, t2);
 681     _mm_storeh_pi((__m64 *)(fptr+1), t2);
 682     _mm_store_ss(fshiftptr, t3);
 683     _mm_storeh_pi((__m64 *)(fshiftptr+1), t3);
 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_update_iforce_3atom_swizzle_ps(fix1, fiy1, fiz1, fix2, fiy2, fiz2, fix3, fiy3, fiz3, \
 689                                               fptr, fshiftptr) \
 690     { \
 691         __m128 _t1, _t2, _t3, _t4; \
 692 \
 693         fix1 = _mm_hadd_ps(fix1, fiy1); \
 694         fiz1 = _mm_hadd_ps(fiz1, fix2); \
 695         fiy2 = _mm_hadd_ps(fiy2, fiz2); \
 696         fix3 = _mm_hadd_ps(fix3, fiy3); \
 697         fiz3 = _mm_hadd_ps(fiz3, fiz3); \
 698         fix1 = _mm_hadd_ps(fix1, fiz1); \
 699         fiy2 = _mm_hadd_ps(fiy2, fix3); \
 700         fiz3 = _mm_hadd_ps(fiz3, fiz3); \
 701         _mm_storeu_ps(fptr,  _mm_add_ps(fix1, _mm_loadu_ps(fptr)  )); \
 702         _mm_storeu_ps(fptr+4, _mm_add_ps(fiy2, _mm_loadu_ps(fptr+4))); \
 703         _mm_store_ss (fptr+8, _mm_add_ss(fiz3, _mm_load_ss(fptr+8) )); \
 704         _t4 = _mm_load_ss(fshiftptr+2); \
 705         _t4 = _mm_loadh_pi(_t4, (__m64 *)(fshiftptr)); \
 706         _t1 = _mm_shuffle_ps(fiz3, fix1, _MM_SHUFFLE(1, 0, 0, 0)); \
 707         _t2 = _mm_shuffle_ps(fix1, fiy2, _MM_SHUFFLE(3, 2, 2, 2)); \
 708         _t3 = _mm_shuffle_ps(fiy2, fix1, _MM_SHUFFLE(3, 3, 0, 1)); \
 709         _t3 = _mm_shuffle_ps(_t3, _t3, _MM_SHUFFLE(1, 2, 0, 0)); \
 710         _t1 = _mm_add_ps(_t1, _t2); \
 711         _t3 = _mm_add_ps(_t3, _t4); \
 712         _t1 = _mm_add_ps(_t1, _t3); \
 713         _mm_store_ss(fshiftptr+2, _t1); \
 714         _mm_storeh_pi((__m64 *)(fshiftptr), _t1); \
 715     }
 716 #else
 717 /* Real function for sane compilers */
 718 static gmx_inline void
 719 gmx_mm_update_iforce_3atom_swizzle_ps(__m128 fix1, __m128 fiy1, __m128 fiz1,
 720                                       __m128 fix2, __m128 fiy2, __m128 fiz2,
 721                                       __m128 fix3, __m128 fiy3, __m128 fiz3,
 722                                       float * gmx_restrict fptr,
 723                                       float * gmx_restrict fshiftptr)
 724 {
 725     __m128 t1, t2, t3, t4;
 726
 727     fix1 = _mm_hadd_ps(fix1, fiy1);
 728     fiz1 = _mm_hadd_ps(fiz1, fix2);
 729     fiy2 = _mm_hadd_ps(fiy2, fiz2);
 730     fix3 = _mm_hadd_ps(fix3, fiy3);
 731     fiz3 = _mm_hadd_ps(fiz3, fiz3);
 732
 733     fix1 = _mm_hadd_ps(fix1, fiz1); /* fix2 fiz1 fiy1 fix1 */
 734     fiy2 = _mm_hadd_ps(fiy2, fix3); /* fiy3 fix3 fiz2 fiy2 */
 735     fiz3 = _mm_hadd_ps(fiz3, fiz3); /*  -    -    -   fiz3 */
 736
 737     _mm_storeu_ps(fptr,  _mm_add_ps(fix1, _mm_loadu_ps(fptr)  ));
 738     _mm_storeu_ps(fptr+4, _mm_add_ps(fiy2, _mm_loadu_ps(fptr+4)));
 739     _mm_store_ss (fptr+8, _mm_add_ss(fiz3, _mm_load_ss(fptr+8) ));
 740
 741     t4 = _mm_load_ss(fshiftptr+2);
 742     t4 = _mm_loadh_pi(t4, (__m64 *)(fshiftptr));
 743
 744     t1 = _mm_shuffle_ps(fiz3, fix1, _MM_SHUFFLE(1, 0, 0, 0)); /* fiy1 fix1  -   fiz3 */
 745     t2 = _mm_shuffle_ps(fix1, fiy2, _MM_SHUFFLE(3, 2, 2, 2)); /* fiy3 fix3  -   fiz1 */
 746     t3 = _mm_shuffle_ps(fiy2, fix1, _MM_SHUFFLE(3, 3, 0, 1)); /* fix2 fix2 fiy2 fiz2 */
 747     t3 = _mm_shuffle_ps(t3, t3, _MM_SHUFFLE(1, 2, 0, 0));     /* fiy2 fix2  -   fiz2 */
 748
 749     t1 = _mm_add_ps(t1, t2);
 750     t3 = _mm_add_ps(t3, t4);
 751     t1 = _mm_add_ps(t1, t3); /* y x - z */
 752
 753     _mm_store_ss(fshiftptr+2, t1);
 754     _mm_storeh_pi((__m64 *)(fshiftptr), t1);
 755 }
 756 #endif
 757
 758 #if defined (_MSC_VER) && defined(_M_IX86)
 759 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 760 #define gmx_mm_update_iforce_4atom_swizzle_ps(fix1, fiy1, fiz1, fix2, fiy2, fiz2, fix3, fiy3, fiz3, fix4, fiy4, fiz4, \
 761                                               fptr, fshiftptr) \
 762     { \
 763         __m128 _t1, _t2, _t3, _t4, _t5; \
 764 \
 765         fix1 = _mm_hadd_ps(fix1, fiy1); \
 766         fiz1 = _mm_hadd_ps(fiz1, fix2); \
 767         fiy2 = _mm_hadd_ps(fiy2, fiz2); \
 768         fix3 = _mm_hadd_ps(fix3, fiy3); \
 769         fiz3 = _mm_hadd_ps(fiz3, fix4); \
 770         fiy4 = _mm_hadd_ps(fiy4, fiz4); \
 771         fix1 = _mm_hadd_ps(fix1, fiz1); \
 772         fiy2 = _mm_hadd_ps(fiy2, fix3); \
 773         fiz3 = _mm_hadd_ps(fiz3, fiy4); \
 774         _mm_storeu_ps(fptr,  _mm_add_ps(fix1, _mm_loadu_ps(fptr)  )); \
 775         _mm_storeu_ps(fptr+4, _mm_add_ps(fiy2, _mm_loadu_ps(fptr+4))); \
 776         _mm_storeu_ps(fptr+8, _mm_add_ps(fiz3, _mm_loadu_ps(fptr+8))); \
 777         _t5 = _mm_load_ss(fshiftptr+2); \
 778         _t5 = _mm_loadh_pi(_t5, (__m64 *)(fshiftptr)); \
 779         _t1 = _mm_shuffle_ps(fix1, fix1, _MM_SHUFFLE(1, 0, 2, 2)); \
 780         _t2 = _mm_shuffle_ps(fiy2, fiy2, _MM_SHUFFLE(3, 2, 1, 1)); \
 781         _t3 = _mm_shuffle_ps(fiz3, fiz3, _MM_SHUFFLE(2, 1, 0, 0)); \
 782         _t4 = _mm_shuffle_ps(fix1, fiy2, _MM_SHUFFLE(0, 0, 3, 3)); \
 783         _t4 = _mm_shuffle_ps(fiz3, _t4, _MM_SHUFFLE(2, 0, 3, 3)); \
 784         _t1 = _mm_add_ps(_t1, _t2); \
 785         _t3 = _mm_add_ps(_t3, _t4); \
 786         _t1 = _mm_add_ps(_t1, _t3); \
 787         _t5 = _mm_add_ps(_t5, _t1); \
 788         _mm_store_ss(fshiftptr+2, _t5); \
 789         _mm_storeh_pi((__m64 *)(fshiftptr), _t5); \
 790     }
 791 #else
 792 /* Real function for sane compilers */
 793 static gmx_inline void
 794 gmx_mm_update_iforce_4atom_swizzle_ps(__m128 fix1, __m128 fiy1, __m128 fiz1,
 795                                       __m128 fix2, __m128 fiy2, __m128 fiz2,
 796                                       __m128 fix3, __m128 fiy3, __m128 fiz3,
 797                                       __m128 fix4, __m128 fiy4, __m128 fiz4,
 798                                       float * gmx_restrict fptr,
 799                                       float * gmx_restrict fshiftptr)
 800 {
 801     __m128 t1, t2, t3, t4, t5;
 802
 803     fix1 = _mm_hadd_ps(fix1, fiy1);
 804     fiz1 = _mm_hadd_ps(fiz1, fix2);
 805     fiy2 = _mm_hadd_ps(fiy2, fiz2);
 806     fix3 = _mm_hadd_ps(fix3, fiy3);
 807     fiz3 = _mm_hadd_ps(fiz3, fix4);
 808     fiy4 = _mm_hadd_ps(fiy4, fiz4);
 809
 810     fix1 = _mm_hadd_ps(fix1, fiz1); /* fix2 fiz1 fiy1 fix1 */
 811     fiy2 = _mm_hadd_ps(fiy2, fix3); /* fiy3 fix3 fiz2 fiy2 */
 812     fiz3 = _mm_hadd_ps(fiz3, fiy4); /* fiz4 fiy4 fix4 fiz3 */
 813
 814     _mm_storeu_ps(fptr,  _mm_add_ps(fix1, _mm_loadu_ps(fptr)  ));
 815     _mm_storeu_ps(fptr+4, _mm_add_ps(fiy2, _mm_loadu_ps(fptr+4)));
 816     _mm_storeu_ps(fptr+8, _mm_add_ps(fiz3, _mm_loadu_ps(fptr+8)));
 817
 818     t5 = _mm_load_ss(fshiftptr+2);
 819     t5 = _mm_loadh_pi(t5, (__m64 *)(fshiftptr));
 820
 821     t1 = _mm_shuffle_ps(fix1, fix1, _MM_SHUFFLE(1, 0, 2, 2));
 822     t2 = _mm_shuffle_ps(fiy2, fiy2, _MM_SHUFFLE(3, 2, 1, 1));
 823     t3 = _mm_shuffle_ps(fiz3, fiz3, _MM_SHUFFLE(2, 1, 0, 0));
 824     t4 = _mm_shuffle_ps(fix1, fiy2, _MM_SHUFFLE(0, 0, 3, 3));
 825     t4 = _mm_shuffle_ps(fiz3, t4, _MM_SHUFFLE(2, 0, 3, 3));
 826
 827     t1 = _mm_add_ps(t1, t2);
 828     t3 = _mm_add_ps(t3, t4);
 829     t1 = _mm_add_ps(t1, t3);
 830     t5 = _mm_add_ps(t5, t1);
 831
 832     _mm_store_ss(fshiftptr+2, t5);
 833     _mm_storeh_pi((__m64 *)(fshiftptr), t5);
 834 }
 835 #endif
 836
 837
 838 static gmx_inline void
 839 gmx_mm_update_1pot_ps(__m128 pot1, float * gmx_restrict ptrA)
 840 {
 841     pot1 = _mm_add_ps(pot1, _mm_movehl_ps(_mm_setzero_ps(), pot1));
 842     pot1 = _mm_add_ps(pot1, _mm_shuffle_ps(pot1, pot1, _MM_SHUFFLE(0, 0, 0, 1)));
 843     _mm_store_ss(ptrA, _mm_add_ss(pot1, _mm_load_ss(ptrA)));
 844 }
 845
 846 static gmx_inline void
 847 gmx_mm_update_2pot_ps(__m128 pot1, float * gmx_restrict ptrA,
 848                       __m128 pot2, float * gmx_restrict ptrB)
 849 {
 850     __m128 t1, t2;
 851     t1   = _mm_movehl_ps(pot2, pot1);
 852     t2   = _mm_movelh_ps(pot1, pot2);
 853     t1   = _mm_add_ps(t1, t2);
 854     t2   = _mm_shuffle_ps(t1, t1, _MM_SHUFFLE(3, 3, 1, 1));
 855     pot1 = _mm_add_ps(t1, t2);
 856     pot2 = _mm_movehl_ps(t2, pot1);
 857     _mm_store_ss(ptrA, _mm_add_ss(pot1, _mm_load_ss(ptrA)));
 858     _mm_store_ss(ptrB, _mm_add_ss(pot2, _mm_load_ss(ptrB)));
 859 }
 860
 861
 862 #endif /* _kernelutil_x86_sse4_1_single_h_ */