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, 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 #include "gromacs/simd/general_x86_sse4_1.h"
  41
  42 #undef gmx_restrict
  43 #define gmx_restrict
  44
  45 /* Normal sum of four xmm registers */
  46 #define gmx_mm_sum4_ps(t0, t1, t2, t3)  _mm_add_ps(_mm_add_ps(t0, t1), _mm_add_ps(t2, t3))
  47
  48 static gmx_inline __m128
  49 gmx_mm_calc_rsq_ps(__m128 dx, __m128 dy, __m128 dz)
  50 {
  51     return _mm_add_ps( _mm_add_ps( _mm_mul_ps(dx, dx), _mm_mul_ps(dy, dy) ), _mm_mul_ps(dz, dz) );
  52 }
  53
  54 static gmx_inline int
  55 gmx_mm_any_lt(__m128 a, __m128 b)
  56 {
  57     return _mm_movemask_ps(_mm_cmplt_ps(a, b));
  58 }
  59
  60 /* Load a single value from 1-4 places, merge into xmm register */
  61
  62 static gmx_inline __m128
  63 gmx_mm_load_4real_swizzle_ps(const float * gmx_restrict ptrA,
  64                              const float * gmx_restrict ptrB,
  65                              const float * gmx_restrict ptrC,
  66                              const float * gmx_restrict ptrD)
  67 {
  68     __m128 t1, t2;
  69
  70     t1 = _mm_unpacklo_ps(_mm_load_ss(ptrA), _mm_load_ss(ptrC));
  71     t2 = _mm_unpacklo_ps(_mm_load_ss(ptrB), _mm_load_ss(ptrD));
  72     return _mm_unpacklo_ps(t1, t2);
  73 }
  74
  75 static gmx_inline void
  76 gmx_mm_store_4real_swizzle_ps(float * gmx_restrict ptrA,
  77                               float * gmx_restrict ptrB,
  78                               float * gmx_restrict ptrC,
  79                               float * gmx_restrict ptrD,
  80                               __m128               xmm1)
  81 {
  82     __m128 t2, t3, t4;
  83
  84     t3       = _mm_movehl_ps(_mm_setzero_ps(), xmm1);
  85     t2       = _mm_shuffle_ps(xmm1, xmm1, _MM_SHUFFLE(1, 1, 1, 1));
  86     t4       = _mm_shuffle_ps(t3, t3, _MM_SHUFFLE(1, 1, 1, 1));
  87     _mm_store_ss(ptrA, xmm1);
  88     _mm_store_ss(ptrB, t2);
  89     _mm_store_ss(ptrC, t3);
  90     _mm_store_ss(ptrD, t4);
  91 }
  92
  93 /* Similar to store, but increments value in memory */
  94 static gmx_inline void
  95 gmx_mm_increment_4real_swizzle_ps(float * gmx_restrict ptrA,
  96                                   float * gmx_restrict ptrB,
  97                                   float * gmx_restrict ptrC,
  98                                   float * gmx_restrict ptrD, __m128 xmm1)
  99 {
 100     __m128 tmp;
 101
 102     tmp = gmx_mm_load_4real_swizzle_ps(ptrA, ptrB, ptrC, ptrD);
 103     tmp = _mm_add_ps(tmp, xmm1);
 104     gmx_mm_store_4real_swizzle_ps(ptrA, ptrB, ptrC, ptrD, tmp);
 105 }
 106
 107
 108 static gmx_inline void
 109 gmx_mm_load_4pair_swizzle_ps(const float * gmx_restrict p1,
 110                              const float * gmx_restrict p2,
 111                              const float * gmx_restrict p3,
 112                              const float * gmx_restrict p4,
 113                              __m128 * gmx_restrict      c6,
 114                              __m128 * gmx_restrict      c12)
 115 {
 116     __m128 t1, t2, t3, t4;
 117
 118     t1   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)p1);   /* - - c12a  c6a */
 119     t2   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)p2);   /* - - c12b  c6b */
 120     t3   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)p3);   /* - - c12c  c6c */
 121     t4   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)p4);   /* - - c12d  c6d */
 122     t1   = _mm_unpacklo_ps(t1, t2);
 123     t2   = _mm_unpacklo_ps(t3, t4);
 124     *c6  = _mm_movelh_ps(t1, t2);
 125     *c12 = _mm_movehl_ps(t2, t1);
 126 }
 127
 128
 129 static gmx_inline void
 130 gmx_mm_load_shift_and_1rvec_broadcast_ps(const float * gmx_restrict xyz_shift,
 131                                          const float * gmx_restrict xyz,
 132                                          __m128 * gmx_restrict      x1,
 133                                          __m128 * gmx_restrict      y1,
 134                                          __m128 * gmx_restrict      z1)
 135 {
 136     __m128 t1, t2, t3, t4;
 137
 138     t1   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)xyz_shift);
 139     t2   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)xyz);
 140     t3   = _mm_load_ss(xyz_shift+2);
 141     t4   = _mm_load_ss(xyz+2);
 142     t1   = _mm_add_ps(t1, t2);
 143     t3   = _mm_add_ss(t3, t4);
 144
 145     *x1  = _mm_shuffle_ps(t1, t1, _MM_SHUFFLE(0, 0, 0, 0));
 146     *y1  = _mm_shuffle_ps(t1, t1, _MM_SHUFFLE(1, 1, 1, 1));
 147     *z1  = _mm_shuffle_ps(t3, t3, _MM_SHUFFLE(0, 0, 0, 0));
 148 }
 149
 150
 151 static gmx_inline void
 152 gmx_mm_load_shift_and_3rvec_broadcast_ps(const float * gmx_restrict xyz_shift,
 153                                          const float * gmx_restrict xyz,
 154                                          __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 155                                          __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 156                                          __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3)
 157 {
 158     __m128 tA, tB;
 159     __m128 t1, t2, t3, t4, t5, t6;
 160
 161     tA   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)xyz_shift);
 162     tB   = _mm_load_ss(xyz_shift+2);
 163
 164     t1   = _mm_loadu_ps(xyz);
 165     t2   = _mm_loadu_ps(xyz+4);
 166     t3   = _mm_load_ss(xyz+8);
 167
 168     tA   = _mm_movelh_ps(tA, tB);
 169     t4   = _mm_shuffle_ps(tA, tA, _MM_SHUFFLE(0, 2, 1, 0));
 170     t5   = _mm_shuffle_ps(tA, tA, _MM_SHUFFLE(1, 0, 2, 1));
 171     t6   = _mm_shuffle_ps(tA, tA, _MM_SHUFFLE(2, 1, 0, 2));
 172
 173     t1   = _mm_add_ps(t1, t4);
 174     t2   = _mm_add_ps(t2, t5);
 175     t3   = _mm_add_ss(t3, t6);
 176
 177     *x1  = _mm_shuffle_ps(t1, t1, _MM_SHUFFLE(0, 0, 0, 0));
 178     *y1  = _mm_shuffle_ps(t1, t1, _MM_SHUFFLE(1, 1, 1, 1));
 179     *z1  = _mm_shuffle_ps(t1, t1, _MM_SHUFFLE(2, 2, 2, 2));
 180     *x2  = _mm_shuffle_ps(t1, t1, _MM_SHUFFLE(3, 3, 3, 3));
 181     *y2  = _mm_shuffle_ps(t2, t2, _MM_SHUFFLE(0, 0, 0, 0));
 182     *z2  = _mm_shuffle_ps(t2, t2, _MM_SHUFFLE(1, 1, 1, 1));
 183     *x3  = _mm_shuffle_ps(t2, t2, _MM_SHUFFLE(2, 2, 2, 2));
 184     *y3  = _mm_shuffle_ps(t2, t2, _MM_SHUFFLE(3, 3, 3, 3));
 185     *z3  = _mm_shuffle_ps(t3, t3, _MM_SHUFFLE(0, 0, 0, 0));
 186 }
 187
 188
 189 static gmx_inline void
 190 gmx_mm_load_shift_and_4rvec_broadcast_ps(const float * gmx_restrict xyz_shift,
 191                                          const float * gmx_restrict xyz,
 192                                          __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 193                                          __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 194                                          __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3,
 195                                          __m128 * gmx_restrict x4, __m128 * gmx_restrict y4, __m128 * gmx_restrict z4)
 196 {
 197     __m128 tA, tB;
 198     __m128 t1, t2, t3, t4, t5, t6;
 199
 200     tA   = _mm_castpd_ps(_mm_load_sd((const double *)xyz_shift));
 201     tB   = _mm_load_ss(xyz_shift+2);
 202
 203     t1   = _mm_loadu_ps(xyz);
 204     t2   = _mm_loadu_ps(xyz+4);
 205     t3   = _mm_loadu_ps(xyz+8);
 206
 207     tA   = _mm_movelh_ps(tA, tB);
 208     t4   = _mm_shuffle_ps(tA, tA, _MM_SHUFFLE(0, 2, 1, 0));
 209     t5   = _mm_shuffle_ps(tA, tA, _MM_SHUFFLE(1, 0, 2, 1));
 210     t6   = _mm_shuffle_ps(tA, tA, _MM_SHUFFLE(2, 1, 0, 2));
 211
 212     t1   = _mm_add_ps(t1, t4);
 213     t2   = _mm_add_ps(t2, t5);
 214     t3   = _mm_add_ps(t3, t6);
 215
 216     *x1  = _mm_shuffle_ps(t1, t1, _MM_SHUFFLE(0, 0, 0, 0));
 217     *y1  = _mm_shuffle_ps(t1, t1, _MM_SHUFFLE(1, 1, 1, 1));
 218     *z1  = _mm_shuffle_ps(t1, t1, _MM_SHUFFLE(2, 2, 2, 2));
 219     *x2  = _mm_shuffle_ps(t1, t1, _MM_SHUFFLE(3, 3, 3, 3));
 220     *y2  = _mm_shuffle_ps(t2, t2, _MM_SHUFFLE(0, 0, 0, 0));
 221     *z2  = _mm_shuffle_ps(t2, t2, _MM_SHUFFLE(1, 1, 1, 1));
 222     *x3  = _mm_shuffle_ps(t2, t2, _MM_SHUFFLE(2, 2, 2, 2));
 223     *y3  = _mm_shuffle_ps(t2, t2, _MM_SHUFFLE(3, 3, 3, 3));
 224     *z3  = _mm_shuffle_ps(t3, t3, _MM_SHUFFLE(0, 0, 0, 0));
 225     *x4  = _mm_shuffle_ps(t3, t3, _MM_SHUFFLE(1, 1, 1, 1));
 226     *y4  = _mm_shuffle_ps(t3, t3, _MM_SHUFFLE(2, 2, 2, 2));
 227     *z4  = _mm_shuffle_ps(t3, t3, _MM_SHUFFLE(3, 3, 3, 3));
 228 }
 229
 230
 231 static gmx_inline void
 232 gmx_mm_load_1rvec_4ptr_swizzle_ps(const float * gmx_restrict ptrA,
 233                                   const float * gmx_restrict ptrB,
 234                                   const float * gmx_restrict ptrC,
 235                                   const float * gmx_restrict ptrD,
 236                                   __m128 *      gmx_restrict x1,
 237                                   __m128 *      gmx_restrict y1,
 238                                   __m128 *      gmx_restrict z1)
 239 {
 240     __m128 t1, t2, t3, t4, t5, t6, t7, t8;
 241     t1   = _mm_castpd_ps(_mm_load_sd((const double *)ptrA));
 242     t2   = _mm_castpd_ps(_mm_load_sd((const double *)ptrB));
 243     t3   = _mm_castpd_ps(_mm_load_sd((const double *)ptrC));
 244     t4   = _mm_castpd_ps(_mm_load_sd((const double *)ptrD));
 245     t5   = _mm_load_ss(ptrA+2);
 246     t6   = _mm_load_ss(ptrB+2);
 247     t7   = _mm_load_ss(ptrC+2);
 248     t8   = _mm_load_ss(ptrD+2);
 249     t1   = _mm_unpacklo_ps(t1, t2);
 250     t3   = _mm_unpacklo_ps(t3, t4);
 251     *x1  = _mm_movelh_ps(t1, t3);
 252     *y1  = _mm_movehl_ps(t3, t1);
 253     t5   = _mm_unpacklo_ps(t5, t6);
 254     t7   = _mm_unpacklo_ps(t7, t8);
 255     *z1  = _mm_movelh_ps(t5, t7);
 256 }
 257
 258
 259 static gmx_inline void
 260 gmx_mm_load_3rvec_4ptr_swizzle_ps(const float * gmx_restrict ptrA,
 261                                   const float * gmx_restrict ptrB,
 262                                   const float * gmx_restrict ptrC,
 263                                   const float * gmx_restrict ptrD,
 264                                   __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 265                                   __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 266                                   __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3)
 267 {
 268     __m128 t1, t2, t3, t4;
 269     t1            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)ptrA ) );
 270     t2            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)ptrB ) );
 271     t3            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)ptrC ) );
 272     t4            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)ptrD ) );
 273     _MM_TRANSPOSE4_PS(t1, t2, t3, t4);
 274     *x1           = t1;
 275     *y1           = t2;
 276     *z1           = t3;
 277     *x2           = t4;
 278     t1            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrA+4) ) );
 279     t2            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrB+4) ) );
 280     t3            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrC+4) ) );
 281     t4            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrD+4) ) );
 282     _MM_TRANSPOSE4_PS(t1, t2, t3, t4);
 283     *y2           = t1;
 284     *z2           = t2;
 285     *x3           = t3;
 286     *y3           = t4;
 287     t1            = _mm_load_ss(ptrA+8);
 288     t2            = _mm_load_ss(ptrB+8);
 289     t3            = _mm_load_ss(ptrC+8);
 290     t4            = _mm_load_ss(ptrD+8);
 291     t1            = _mm_unpacklo_ps(t1, t3);
 292     t3            = _mm_unpacklo_ps(t2, t4);
 293     *z3           = _mm_unpacklo_ps(t1, t3);
 294 }
 295
 296
 297 static gmx_inline void
 298 gmx_mm_load_4rvec_4ptr_swizzle_ps(const float * gmx_restrict ptrA,
 299                                   const float * gmx_restrict ptrB,
 300                                   const float * gmx_restrict ptrC,
 301                                   const float * gmx_restrict ptrD,
 302                                   __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 303                                   __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 304                                   __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3,
 305                                   __m128 * gmx_restrict x4, __m128 * gmx_restrict y4, __m128 * gmx_restrict z4)
 306 {
 307     __m128 t1, t2, t3, t4;
 308     t1            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrA) ) );
 309     t2            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrB) ) );
 310     t3            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrC) ) );
 311     t4            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrD) ) );
 312     _MM_TRANSPOSE4_PS(t1, t2, t3, t4);
 313     *x1           = t1;
 314     *y1           = t2;
 315     *z1           = t3;
 316     *x2           = t4;
 317     t1            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrA+4) ) );
 318     t2            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrB+4) ) );
 319     t3            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrC+4) ) );
 320     t4            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrD+4) ) );
 321     _MM_TRANSPOSE4_PS(t1, t2, t3, t4);
 322     *y2           = t1;
 323     *z2           = t2;
 324     *x3           = t3;
 325     *y3           = t4;
 326     t1            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrA+8) ) );
 327     t2            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrB+8) ) );
 328     t3            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrC+8) ) );
 329     t4            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrD+8) ) );
 330     _MM_TRANSPOSE4_PS(t1, t2, t3, t4);
 331     *z3           = t1;
 332     *x4           = t2;
 333     *y4           = t3;
 334     *z4           = t4;
 335 }
 336
 337
 338
 339 static gmx_inline void
 340 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(float * ptrA,
 341                                        float * ptrB,
 342                                        float * ptrC,
 343                                        float * ptrD,
 344                                        __m128 x1, __m128 y1, __m128 z1)
 345 {
 346     __m128 t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12;
 347     t5          = _mm_unpacklo_ps(y1, z1);
 348     t6          = _mm_unpackhi_ps(y1, z1);
 349     t7          = _mm_shuffle_ps(x1, t5, _MM_SHUFFLE(1, 0, 0, 0));
 350     t8          = _mm_shuffle_ps(x1, t5, _MM_SHUFFLE(3, 2, 0, 1));
 351     t9          = _mm_shuffle_ps(x1, t6, _MM_SHUFFLE(1, 0, 0, 2));
 352     t10         = _mm_shuffle_ps(x1, t6, _MM_SHUFFLE(3, 2, 0, 3));
 353     t1          = _mm_load_ss(ptrA);
 354     t1          = _mm_loadh_pi(t1, (__m64 *)(ptrA+1));
 355     t1          = _mm_sub_ps(t1, t7);
 356     _mm_store_ss(ptrA, t1);
 357     _mm_storeh_pi((__m64 *)(ptrA+1), t1);
 358     t2          = _mm_load_ss(ptrB);
 359     t2          = _mm_loadh_pi(t2, (__m64 *)(ptrB+1));
 360     t2          = _mm_sub_ps(t2, t8);
 361     _mm_store_ss(ptrB, t2);
 362     _mm_storeh_pi((__m64 *)(ptrB+1), t2);
 363     t3          = _mm_load_ss(ptrC);
 364     t3          = _mm_loadh_pi(t3, (__m64 *)(ptrC+1));
 365     t3          = _mm_sub_ps(t3, t9);
 366     _mm_store_ss(ptrC, t3);
 367     _mm_storeh_pi((__m64 *)(ptrC+1), t3);
 368     t4          = _mm_load_ss(ptrD);
 369     t4          = _mm_loadh_pi(t4, (__m64 *)(ptrD+1));
 370     t4          = _mm_sub_ps(t4, t10);
 371     _mm_store_ss(ptrD, t4);
 372     _mm_storeh_pi((__m64 *)(ptrD+1), t4);
 373 }
 374
 375
 376
 377 #if defined (_MSC_VER) && defined(_M_IX86)
 378 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 379 #define gmx_mm_decrement_3rvec_4ptr_swizzle_ps(ptrA, ptrB, ptrC, ptrD, \
 380                                                _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3) \
 381     { \
 382         __m128 _t1, _t2, _t3, _t4, _t5, _t6, _t7, _t8, _t9, _t10; \
 383         __m128 _t11, _t12, _t13, _t14, _t15, _t16, _t17, _t18, _t19; \
 384         __m128 _t20, _t21, _t22, _t23, _t24, _t25; \
 385         _t13         = _mm_unpackhi_ps(_x1, _y1); \
 386         _x1          = _mm_unpacklo_ps(_x1, _y1); \
 387         _t14         = _mm_unpackhi_ps(_z1, _x2); \
 388         _z1          = _mm_unpacklo_ps(_z1, _x2); \
 389         _t15         = _mm_unpackhi_ps(_y2, _z2); \
 390         _y2          = _mm_unpacklo_ps(_y2, _z2); \
 391         _t16         = _mm_unpackhi_ps(_x3, _y3); \
 392         _x3          = _mm_unpacklo_ps(_x3, _y3); \
 393         _t17         = _mm_shuffle_ps(_z3, _z3, _MM_SHUFFLE(0, 0, 0, 1)); \
 394         _t18         = _mm_movehl_ps(_z3, _z3); \
 395         _t19         = _mm_shuffle_ps(_t18, _t18, _MM_SHUFFLE(0, 0, 0, 1)); \
 396         _t20         = _mm_movelh_ps(_x1, _z1); \
 397         _t21         = _mm_movehl_ps(_z1, _x1); \
 398         _t22         = _mm_movelh_ps(_t13, _t14); \
 399         _t14         = _mm_movehl_ps(_t14, _t13); \
 400         _t23         = _mm_movelh_ps(_y2, _x3); \
 401         _t24         = _mm_movehl_ps(_x3, _y2); \
 402         _t25         = _mm_movelh_ps(_t15, _t16); \
 403         _t16         = _mm_movehl_ps(_t16, _t15); \
 404         _t1          = _mm_loadu_ps(ptrA); \
 405         _t2          = _mm_loadu_ps(ptrA+4); \
 406         _t3          = _mm_load_ss(ptrA+8); \
 407         _t1          = _mm_sub_ps(_t1, _t20); \
 408         _t2          = _mm_sub_ps(_t2, _t23); \
 409         _t3          = _mm_sub_ss(_t3, _z3); \
 410         _mm_storeu_ps(ptrA, _t1); \
 411         _mm_storeu_ps(ptrA+4, _t2); \
 412         _mm_store_ss(ptrA+8, _t3); \
 413         _t4          = _mm_loadu_ps(ptrB); \
 414         _t5          = _mm_loadu_ps(ptrB+4); \
 415         _t6          = _mm_load_ss(ptrB+8); \
 416         _t4          = _mm_sub_ps(_t4, _t21); \
 417         _t5          = _mm_sub_ps(_t5, _t24); \
 418         _t6          = _mm_sub_ss(_t6, _t17); \
 419         _mm_storeu_ps(ptrB, _t4); \
 420         _mm_storeu_ps(ptrB+4, _t5); \
 421         _mm_store_ss(ptrB+8, _t6); \
 422         _t7          = _mm_loadu_ps(ptrC); \
 423         _t8          = _mm_loadu_ps(ptrC+4); \
 424         _t9          = _mm_load_ss(ptrC+8); \
 425         _t7          = _mm_sub_ps(_t7, _t22); \
 426         _t8          = _mm_sub_ps(_t8, _t25); \
 427         _t9          = _mm_sub_ss(_t9, _t18); \
 428         _mm_storeu_ps(ptrC, _t7); \
 429         _mm_storeu_ps(ptrC+4, _t8); \
 430         _mm_store_ss(ptrC+8, _t9); \
 431         _t10         = _mm_loadu_ps(ptrD); \
 432         _t11         = _mm_loadu_ps(ptrD+4); \
 433         _t12         = _mm_load_ss(ptrD+8); \
 434         _t10         = _mm_sub_ps(_t10, _t14); \
 435         _t11         = _mm_sub_ps(_t11, _t16); \
 436         _t12         = _mm_sub_ss(_t12, _t19); \
 437         _mm_storeu_ps(ptrD, _t10); \
 438         _mm_storeu_ps(ptrD+4, _t11); \
 439         _mm_store_ss(ptrD+8, _t12); \
 440     }
 441 #else
 442 /* Real function for sane compilers */
 443 static gmx_inline void
 444 gmx_mm_decrement_3rvec_4ptr_swizzle_ps(float * gmx_restrict ptrA, float * gmx_restrict ptrB,
 445                                        float * gmx_restrict ptrC, float * gmx_restrict ptrD,
 446                                        __m128 x1, __m128 y1, __m128 z1,
 447                                        __m128 x2, __m128 y2, __m128 z2,
 448                                        __m128 x3, __m128 y3, __m128 z3)
 449 {
 450     __m128 t1, t2, t3, t4, t5, t6, t7, t8, t9, t10;
 451     __m128 t11, t12, t13, t14, t15, t16, t17, t18, t19;
 452     __m128 t20, t21, t22, t23, t24, t25;
 453
 454     t13         = _mm_unpackhi_ps(x1, y1);
 455     x1          = _mm_unpacklo_ps(x1, y1);
 456     t14         = _mm_unpackhi_ps(z1, x2);
 457     z1          = _mm_unpacklo_ps(z1, x2);
 458     t15         = _mm_unpackhi_ps(y2, z2);
 459     y2          = _mm_unpacklo_ps(y2, z2);
 460     t16         = _mm_unpackhi_ps(x3, y3);
 461     x3          = _mm_unpacklo_ps(x3, y3);
 462     t17         = _mm_shuffle_ps(z3, z3, _MM_SHUFFLE(0, 0, 0, 1));
 463     t18         = _mm_movehl_ps(z3, z3);
 464     t19         = _mm_shuffle_ps(t18, t18, _MM_SHUFFLE(0, 0, 0, 1));
 465     t20         = _mm_movelh_ps(x1, z1);
 466     t21         = _mm_movehl_ps(z1, x1);
 467     t22         = _mm_movelh_ps(t13, t14);
 468     t14         = _mm_movehl_ps(t14, t13);
 469     t23         = _mm_movelh_ps(y2, x3);
 470     t24         = _mm_movehl_ps(x3, y2);
 471     t25         = _mm_movelh_ps(t15, t16);
 472     t16         = _mm_movehl_ps(t16, t15);
 473     t1          = _mm_loadu_ps(ptrA);
 474     t2          = _mm_loadu_ps(ptrA+4);
 475     t3          = _mm_load_ss(ptrA+8);
 476     t4          = _mm_loadu_ps(ptrB);
 477     t5          = _mm_loadu_ps(ptrB+4);
 478     t6          = _mm_load_ss(ptrB+8);
 479     t7          = _mm_loadu_ps(ptrC);
 480     t8          = _mm_loadu_ps(ptrC+4);
 481     t9          = _mm_load_ss(ptrC+8);
 482     t10         = _mm_loadu_ps(ptrD);
 483     t11         = _mm_loadu_ps(ptrD+4);
 484     t12         = _mm_load_ss(ptrD+8);
 485
 486     t1          = _mm_sub_ps(t1, t20);
 487     t2          = _mm_sub_ps(t2, t23);
 488     t3          = _mm_sub_ss(t3, z3);
 489     _mm_storeu_ps(ptrA, t1);
 490     _mm_storeu_ps(ptrA+4, t2);
 491     _mm_store_ss(ptrA+8, t3);
 492     t4          = _mm_sub_ps(t4, t21);
 493     t5          = _mm_sub_ps(t5, t24);
 494     t6          = _mm_sub_ss(t6, t17);
 495     _mm_storeu_ps(ptrB, t4);
 496     _mm_storeu_ps(ptrB+4, t5);
 497     _mm_store_ss(ptrB+8, t6);
 498     t7          = _mm_sub_ps(t7, t22);
 499     t8          = _mm_sub_ps(t8, t25);
 500     t9          = _mm_sub_ss(t9, t18);
 501     _mm_storeu_ps(ptrC, t7);
 502     _mm_storeu_ps(ptrC+4, t8);
 503     _mm_store_ss(ptrC+8, t9);
 504     t10         = _mm_sub_ps(t10, t14);
 505     t11         = _mm_sub_ps(t11, t16);
 506     t12         = _mm_sub_ss(t12, t19);
 507     _mm_storeu_ps(ptrD, t10);
 508     _mm_storeu_ps(ptrD+4, t11);
 509     _mm_store_ss(ptrD+8, t12);
 510 }
 511 #endif
 512
 513 #if defined (_MSC_VER) && defined(_M_IX86)
 514 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 515 #define gmx_mm_decrement_4rvec_4ptr_swizzle_ps(ptrA, ptrB, ptrC, ptrD, \
 516                                                _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3, _x4, _y4, _z4) \
 517     { \
 518         __m128 _t1, _t2, _t3, _t4, _t5, _t6, _t7, _t8, _t9, _t10, _t11; \
 519         __m128 _t12, _t13, _t14, _t15, _t16, _t17, _t18, _t19, _t20, _t21, _t22; \
 520         __m128 _t23, _t24; \
 521         _t13         = _mm_unpackhi_ps(_x1, _y1); \
 522         _x1          = _mm_unpacklo_ps(_x1, _y1); \
 523         _t14         = _mm_unpackhi_ps(_z1, _x2); \
 524         _z1          = _mm_unpacklo_ps(_z1, _x2); \
 525         _t15         = _mm_unpackhi_ps(_y2, _z2); \
 526         _y2          = _mm_unpacklo_ps(_y2, _z2); \
 527         _t16         = _mm_unpackhi_ps(_x3, _y3); \
 528         _x3          = _mm_unpacklo_ps(_x3, _y3); \
 529         _t17         = _mm_unpackhi_ps(_z3, _x4); \
 530         _z3          = _mm_unpacklo_ps(_z3, _x4); \
 531         _t18         = _mm_unpackhi_ps(_y4, _z4); \
 532         _y4          = _mm_unpacklo_ps(_y4, _z4); \
 533         _t19         = _mm_movelh_ps(_x1, _z1); \
 534         _z1          = _mm_movehl_ps(_z1, _x1); \
 535         _t20         = _mm_movelh_ps(_t13, _t14); \
 536         _t14         = _mm_movehl_ps(_t14, _t13); \
 537         _t21         = _mm_movelh_ps(_y2, _x3); \
 538         _x3          = _mm_movehl_ps(_x3, _y2); \
 539         _t22         = _mm_movelh_ps(_t15, _t16); \
 540         _t16         = _mm_movehl_ps(_t16, _t15); \
 541         _t23         = _mm_movelh_ps(_z3, _y4); \
 542         _y4          = _mm_movehl_ps(_y4, _z3); \
 543         _t24         = _mm_movelh_ps(_t17, _t18); \
 544         _t18         = _mm_movehl_ps(_t18, _t17); \
 545         _t1          = _mm_loadu_ps(ptrA); \
 546         _t2          = _mm_loadu_ps(ptrA+4); \
 547         _t3          = _mm_loadu_ps(ptrA+8); \
 548         _t1          = _mm_sub_ps(_t1, _t19); \
 549         _t2          = _mm_sub_ps(_t2, _t21); \
 550         _t3          = _mm_sub_ps(_t3, _t23); \
 551         _mm_storeu_ps(ptrA, _t1); \
 552         _mm_storeu_ps(ptrA+4, _t2); \
 553         _mm_storeu_ps(ptrA+8, _t3); \
 554         _t4          = _mm_loadu_ps(ptrB); \
 555         _t5          = _mm_loadu_ps(ptrB+4); \
 556         _t6          = _mm_loadu_ps(ptrB+8); \
 557         _t4          = _mm_sub_ps(_t4, _z1); \
 558         _t5          = _mm_sub_ps(_t5, _x3); \
 559         _t6          = _mm_sub_ps(_t6, _y4); \
 560         _mm_storeu_ps(ptrB, _t4); \
 561         _mm_storeu_ps(ptrB+4, _t5); \
 562         _mm_storeu_ps(ptrB+8, _t6); \
 563         _t7          = _mm_loadu_ps(ptrC); \
 564         _t8          = _mm_loadu_ps(ptrC+4); \
 565         _t9          = _mm_loadu_ps(ptrC+8); \
 566         _t7          = _mm_sub_ps(_t7, _t20); \
 567         _t8          = _mm_sub_ps(_t8, _t22); \
 568         _t9          = _mm_sub_ps(_t9, _t24); \
 569         _mm_storeu_ps(ptrC, _t7); \
 570         _mm_storeu_ps(ptrC+4, _t8); \
 571         _mm_storeu_ps(ptrC+8, _t9); \
 572         _t10         = _mm_loadu_ps(ptrD); \
 573         _t11         = _mm_loadu_ps(ptrD+4); \
 574         _t12         = _mm_loadu_ps(ptrD+8); \
 575         _t10         = _mm_sub_ps(_t10, _t14); \
 576         _t11         = _mm_sub_ps(_t11, _t16); \
 577         _t12         = _mm_sub_ps(_t12, _t18); \
 578         _mm_storeu_ps(ptrD, _t10); \
 579         _mm_storeu_ps(ptrD+4, _t11); \
 580         _mm_storeu_ps(ptrD+8, _t12); \
 581     }
 582 #else
 583 /* Real function for sane compilers */
 584 static gmx_inline void
 585 gmx_mm_decrement_4rvec_4ptr_swizzle_ps(float * gmx_restrict ptrA, float * gmx_restrict ptrB,
 586                                        float * gmx_restrict ptrC, float * gmx_restrict ptrD,
 587                                        __m128 x1, __m128 y1, __m128 z1,
 588                                        __m128 x2, __m128 y2, __m128 z2,
 589                                        __m128 x3, __m128 y3, __m128 z3,
 590                                        __m128 x4, __m128 y4, __m128 z4)
 591 {
 592     __m128 t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11;
 593     __m128 t12, t13, t14, t15, t16, t17, t18, t19, t20, t21, t22;
 594     __m128 t23, t24;
 595     t13         = _mm_unpackhi_ps(x1, y1);
 596     x1          = _mm_unpacklo_ps(x1, y1);
 597     t14         = _mm_unpackhi_ps(z1, x2);
 598     z1          = _mm_unpacklo_ps(z1, x2);
 599     t15         = _mm_unpackhi_ps(y2, z2);
 600     y2          = _mm_unpacklo_ps(y2, z2);
 601     t16         = _mm_unpackhi_ps(x3, y3);
 602     x3          = _mm_unpacklo_ps(x3, y3);
 603     t17         = _mm_unpackhi_ps(z3, x4);
 604     z3          = _mm_unpacklo_ps(z3, x4);
 605     t18         = _mm_unpackhi_ps(y4, z4);
 606     y4          = _mm_unpacklo_ps(y4, z4);
 607     t19         = _mm_movelh_ps(x1, z1);
 608     z1          = _mm_movehl_ps(z1, x1);
 609     t20         = _mm_movelh_ps(t13, t14);
 610     t14         = _mm_movehl_ps(t14, t13);
 611     t21         = _mm_movelh_ps(y2, x3);
 612     x3          = _mm_movehl_ps(x3, y2);
 613     t22         = _mm_movelh_ps(t15, t16);
 614     t16         = _mm_movehl_ps(t16, t15);
 615     t23         = _mm_movelh_ps(z3, y4);
 616     y4          = _mm_movehl_ps(y4, z3);
 617     t24         = _mm_movelh_ps(t17, t18);
 618     t18         = _mm_movehl_ps(t18, t17);
 619     t1          = _mm_loadu_ps(ptrA);
 620     t2          = _mm_loadu_ps(ptrA+4);
 621     t3          = _mm_loadu_ps(ptrA+8);
 622     t1          = _mm_sub_ps(t1, t19);
 623     t2          = _mm_sub_ps(t2, t21);
 624     t3          = _mm_sub_ps(t3, t23);
 625     _mm_storeu_ps(ptrA, t1);
 626     _mm_storeu_ps(ptrA+4, t2);
 627     _mm_storeu_ps(ptrA+8, t3);
 628     t4          = _mm_loadu_ps(ptrB);
 629     t5          = _mm_loadu_ps(ptrB+4);
 630     t6          = _mm_loadu_ps(ptrB+8);
 631     t4          = _mm_sub_ps(t4, z1);
 632     t5          = _mm_sub_ps(t5, x3);
 633     t6          = _mm_sub_ps(t6, y4);
 634     _mm_storeu_ps(ptrB, t4);
 635     _mm_storeu_ps(ptrB+4, t5);
 636     _mm_storeu_ps(ptrB+8, t6);
 637     t7          = _mm_loadu_ps(ptrC);
 638     t8          = _mm_loadu_ps(ptrC+4);
 639     t9          = _mm_loadu_ps(ptrC+8);
 640     t7          = _mm_sub_ps(t7, t20);
 641     t8          = _mm_sub_ps(t8, t22);
 642     t9          = _mm_sub_ps(t9, t24);
 643     _mm_storeu_ps(ptrC, t7);
 644     _mm_storeu_ps(ptrC+4, t8);
 645     _mm_storeu_ps(ptrC+8, t9);
 646     t10         = _mm_loadu_ps(ptrD);
 647     t11         = _mm_loadu_ps(ptrD+4);
 648     t12         = _mm_loadu_ps(ptrD+8);
 649     t10         = _mm_sub_ps(t10, t14);
 650     t11         = _mm_sub_ps(t11, t16);
 651     t12         = _mm_sub_ps(t12, t18);
 652     _mm_storeu_ps(ptrD, t10);
 653     _mm_storeu_ps(ptrD+4, t11);
 654     _mm_storeu_ps(ptrD+8, t12);
 655 }
 656 #endif
 657
 658
 659 static gmx_inline void
 660 gmx_mm_update_iforce_1atom_swizzle_ps(__m128 fix1, __m128 fiy1, __m128 fiz1,
 661                                       float * gmx_restrict fptr,
 662                                       float * gmx_restrict fshiftptr)
 663 {
 664     __m128 t2, t3;
 665
 666     fix1 = _mm_hadd_ps(fix1, fix1);
 667     fiy1 = _mm_hadd_ps(fiy1, fiz1);
 668
 669     fix1 = _mm_hadd_ps(fix1, fiy1); /* fiz1 fiy1 fix1 fix1 */
 670
 671     t2 = _mm_load_ss(fptr);
 672     t2 = _mm_loadh_pi(t2, (__m64 *)(fptr+1));
 673     t3 = _mm_load_ss(fshiftptr);
 674     t3 = _mm_loadh_pi(t3, (__m64 *)(fshiftptr+1));
 675
 676     t2 = _mm_add_ps(t2, fix1);
 677     t3 = _mm_add_ps(t3, fix1);
 678
 679     _mm_store_ss(fptr, t2);
 680     _mm_storeh_pi((__m64 *)(fptr+1), t2);
 681     _mm_store_ss(fshiftptr, t3);
 682     _mm_storeh_pi((__m64 *)(fshiftptr+1), t3);
 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_update_iforce_3atom_swizzle_ps(fix1, fiy1, fiz1, fix2, fiy2, fiz2, fix3, fiy3, fiz3, \
 688                                               fptr, fshiftptr) \
 689     { \
 690         __m128 _t1, _t2, _t3, _t4; \
 691 \
 692         fix1 = _mm_hadd_ps(fix1, fiy1); \
 693         fiz1 = _mm_hadd_ps(fiz1, fix2); \
 694         fiy2 = _mm_hadd_ps(fiy2, fiz2); \
 695         fix3 = _mm_hadd_ps(fix3, fiy3); \
 696         fiz3 = _mm_hadd_ps(fiz3, fiz3); \
 697         fix1 = _mm_hadd_ps(fix1, fiz1); \
 698         fiy2 = _mm_hadd_ps(fiy2, fix3); \
 699         fiz3 = _mm_hadd_ps(fiz3, fiz3); \
 700         _mm_storeu_ps(fptr,  _mm_add_ps(fix1, _mm_loadu_ps(fptr)  )); \
 701         _mm_storeu_ps(fptr+4, _mm_add_ps(fiy2, _mm_loadu_ps(fptr+4))); \
 702         _mm_store_ss (fptr+8, _mm_add_ss(fiz3, _mm_load_ss(fptr+8) )); \
 703         _t4 = _mm_load_ss(fshiftptr+2); \
 704         _t4 = _mm_loadh_pi(_t4, (__m64 *)(fshiftptr)); \
 705         _t1 = _mm_shuffle_ps(fiz3, fix1, _MM_SHUFFLE(1, 0, 0, 0)); \
 706         _t2 = _mm_shuffle_ps(fix1, fiy2, _MM_SHUFFLE(3, 2, 2, 2)); \
 707         _t3 = _mm_shuffle_ps(fiy2, fix1, _MM_SHUFFLE(3, 3, 0, 1)); \
 708         _t3 = _mm_shuffle_ps(_t3, _t3, _MM_SHUFFLE(1, 2, 0, 0)); \
 709         _t1 = _mm_add_ps(_t1, _t2); \
 710         _t3 = _mm_add_ps(_t3, _t4); \
 711         _t1 = _mm_add_ps(_t1, _t3); \
 712         _mm_store_ss(fshiftptr+2, _t1); \
 713         _mm_storeh_pi((__m64 *)(fshiftptr), _t1); \
 714     }
 715 #else
 716 /* Real function for sane compilers */
 717 static gmx_inline void
 718 gmx_mm_update_iforce_3atom_swizzle_ps(__m128 fix1, __m128 fiy1, __m128 fiz1,
 719                                       __m128 fix2, __m128 fiy2, __m128 fiz2,
 720                                       __m128 fix3, __m128 fiy3, __m128 fiz3,
 721                                       float * gmx_restrict fptr,
 722                                       float * gmx_restrict fshiftptr)
 723 {
 724     __m128 t1, t2, t3, t4;
 725
 726     fix1 = _mm_hadd_ps(fix1, fiy1);
 727     fiz1 = _mm_hadd_ps(fiz1, fix2);
 728     fiy2 = _mm_hadd_ps(fiy2, fiz2);
 729     fix3 = _mm_hadd_ps(fix3, fiy3);
 730     fiz3 = _mm_hadd_ps(fiz3, fiz3);
 731
 732     fix1 = _mm_hadd_ps(fix1, fiz1); /* fix2 fiz1 fiy1 fix1 */
 733     fiy2 = _mm_hadd_ps(fiy2, fix3); /* fiy3 fix3 fiz2 fiy2 */
 734     fiz3 = _mm_hadd_ps(fiz3, fiz3); /*  -    -    -   fiz3 */
 735
 736     _mm_storeu_ps(fptr,  _mm_add_ps(fix1, _mm_loadu_ps(fptr)  ));
 737     _mm_storeu_ps(fptr+4, _mm_add_ps(fiy2, _mm_loadu_ps(fptr+4)));
 738     _mm_store_ss (fptr+8, _mm_add_ss(fiz3, _mm_load_ss(fptr+8) ));
 739
 740     t4 = _mm_load_ss(fshiftptr+2);
 741     t4 = _mm_loadh_pi(t4, (__m64 *)(fshiftptr));
 742
 743     t1 = _mm_shuffle_ps(fiz3, fix1, _MM_SHUFFLE(1, 0, 0, 0)); /* fiy1 fix1  -   fiz3 */
 744     t2 = _mm_shuffle_ps(fix1, fiy2, _MM_SHUFFLE(3, 2, 2, 2)); /* fiy3 fix3  -   fiz1 */
 745     t3 = _mm_shuffle_ps(fiy2, fix1, _MM_SHUFFLE(3, 3, 0, 1)); /* fix2 fix2 fiy2 fiz2 */
 746     t3 = _mm_shuffle_ps(t3, t3, _MM_SHUFFLE(1, 2, 0, 0));     /* fiy2 fix2  -   fiz2 */
 747
 748     t1 = _mm_add_ps(t1, t2);
 749     t3 = _mm_add_ps(t3, t4);
 750     t1 = _mm_add_ps(t1, t3); /* y x - z */
 751
 752     _mm_store_ss(fshiftptr+2, t1);
 753     _mm_storeh_pi((__m64 *)(fshiftptr), t1);
 754 }
 755 #endif
 756
 757 #if defined (_MSC_VER) && defined(_M_IX86)
 758 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 759 #define gmx_mm_update_iforce_4atom_swizzle_ps(fix1, fiy1, fiz1, fix2, fiy2, fiz2, fix3, fiy3, fiz3, fix4, fiy4, fiz4, \
 760                                               fptr, fshiftptr) \
 761     { \
 762         __m128 _t1, _t2, _t3, _t4, _t5; \
 763 \
 764         fix1 = _mm_hadd_ps(fix1, fiy1); \
 765         fiz1 = _mm_hadd_ps(fiz1, fix2); \
 766         fiy2 = _mm_hadd_ps(fiy2, fiz2); \
 767         fix3 = _mm_hadd_ps(fix3, fiy3); \
 768         fiz3 = _mm_hadd_ps(fiz3, fix4); \
 769         fiy4 = _mm_hadd_ps(fiy4, fiz4); \
 770         fix1 = _mm_hadd_ps(fix1, fiz1); \
 771         fiy2 = _mm_hadd_ps(fiy2, fix3); \
 772         fiz3 = _mm_hadd_ps(fiz3, fiy4); \
 773         _mm_storeu_ps(fptr,  _mm_add_ps(fix1, _mm_loadu_ps(fptr)  )); \
 774         _mm_storeu_ps(fptr+4, _mm_add_ps(fiy2, _mm_loadu_ps(fptr+4))); \
 775         _mm_storeu_ps(fptr+8, _mm_add_ps(fiz3, _mm_loadu_ps(fptr+8))); \
 776         _t5 = _mm_load_ss(fshiftptr+2); \
 777         _t5 = _mm_loadh_pi(_t5, (__m64 *)(fshiftptr)); \
 778         _t1 = _mm_shuffle_ps(fix1, fix1, _MM_SHUFFLE(1, 0, 2, 2)); \
 779         _t2 = _mm_shuffle_ps(fiy2, fiy2, _MM_SHUFFLE(3, 2, 1, 1)); \
 780         _t3 = _mm_shuffle_ps(fiz3, fiz3, _MM_SHUFFLE(2, 1, 0, 0)); \
 781         _t4 = _mm_shuffle_ps(fix1, fiy2, _MM_SHUFFLE(0, 0, 3, 3)); \
 782         _t4 = _mm_shuffle_ps(fiz3, _t4, _MM_SHUFFLE(2, 0, 3, 3)); \
 783         _t1 = _mm_add_ps(_t1, _t2); \
 784         _t3 = _mm_add_ps(_t3, _t4); \
 785         _t1 = _mm_add_ps(_t1, _t3); \
 786         _t5 = _mm_add_ps(_t5, _t1); \
 787         _mm_store_ss(fshiftptr+2, _t5); \
 788         _mm_storeh_pi((__m64 *)(fshiftptr), _t5); \
 789     }
 790 #else
 791 /* Real function for sane compilers */
 792 static gmx_inline void
 793 gmx_mm_update_iforce_4atom_swizzle_ps(__m128 fix1, __m128 fiy1, __m128 fiz1,
 794                                       __m128 fix2, __m128 fiy2, __m128 fiz2,
 795                                       __m128 fix3, __m128 fiy3, __m128 fiz3,
 796                                       __m128 fix4, __m128 fiy4, __m128 fiz4,
 797                                       float * gmx_restrict fptr,
 798                                       float * gmx_restrict fshiftptr)
 799 {
 800     __m128 t1, t2, t3, t4, t5;
 801
 802     fix1 = _mm_hadd_ps(fix1, fiy1);
 803     fiz1 = _mm_hadd_ps(fiz1, fix2);
 804     fiy2 = _mm_hadd_ps(fiy2, fiz2);
 805     fix3 = _mm_hadd_ps(fix3, fiy3);
 806     fiz3 = _mm_hadd_ps(fiz3, fix4);
 807     fiy4 = _mm_hadd_ps(fiy4, fiz4);
 808
 809     fix1 = _mm_hadd_ps(fix1, fiz1); /* fix2 fiz1 fiy1 fix1 */
 810     fiy2 = _mm_hadd_ps(fiy2, fix3); /* fiy3 fix3 fiz2 fiy2 */
 811     fiz3 = _mm_hadd_ps(fiz3, fiy4); /* fiz4 fiy4 fix4 fiz3 */
 812
 813     _mm_storeu_ps(fptr,  _mm_add_ps(fix1, _mm_loadu_ps(fptr)  ));
 814     _mm_storeu_ps(fptr+4, _mm_add_ps(fiy2, _mm_loadu_ps(fptr+4)));
 815     _mm_storeu_ps(fptr+8, _mm_add_ps(fiz3, _mm_loadu_ps(fptr+8)));
 816
 817     t5 = _mm_load_ss(fshiftptr+2);
 818     t5 = _mm_loadh_pi(t5, (__m64 *)(fshiftptr));
 819
 820     t1 = _mm_shuffle_ps(fix1, fix1, _MM_SHUFFLE(1, 0, 2, 2));
 821     t2 = _mm_shuffle_ps(fiy2, fiy2, _MM_SHUFFLE(3, 2, 1, 1));
 822     t3 = _mm_shuffle_ps(fiz3, fiz3, _MM_SHUFFLE(2, 1, 0, 0));
 823     t4 = _mm_shuffle_ps(fix1, fiy2, _MM_SHUFFLE(0, 0, 3, 3));
 824     t4 = _mm_shuffle_ps(fiz3, t4, _MM_SHUFFLE(2, 0, 3, 3));
 825
 826     t1 = _mm_add_ps(t1, t2);
 827     t3 = _mm_add_ps(t3, t4);
 828     t1 = _mm_add_ps(t1, t3);
 829     t5 = _mm_add_ps(t5, t1);
 830
 831     _mm_store_ss(fshiftptr+2, t5);
 832     _mm_storeh_pi((__m64 *)(fshiftptr), t5);
 833 }
 834 #endif
 835
 836
 837 static gmx_inline void
 838 gmx_mm_update_1pot_ps(__m128 pot1, float * gmx_restrict ptrA)
 839 {
 840     pot1 = _mm_add_ps(pot1, _mm_movehl_ps(_mm_setzero_ps(), pot1));
 841     pot1 = _mm_add_ps(pot1, _mm_shuffle_ps(pot1, pot1, _MM_SHUFFLE(0, 0, 0, 1)));
 842     _mm_store_ss(ptrA, _mm_add_ss(pot1, _mm_load_ss(ptrA)));
 843 }
 844
 845 static gmx_inline void
 846 gmx_mm_update_2pot_ps(__m128 pot1, float * gmx_restrict ptrA,
 847                       __m128 pot2, float * gmx_restrict ptrB)
 848 {
 849     __m128 t1, t2;
 850     t1   = _mm_movehl_ps(pot2, pot1);
 851     t2   = _mm_movelh_ps(pot1, pot2);
 852     t1   = _mm_add_ps(t1, t2);
 853     t2   = _mm_shuffle_ps(t1, t1, _MM_SHUFFLE(3, 3, 1, 1));
 854     pot1 = _mm_add_ps(t1, t2);
 855     pot2 = _mm_movehl_ps(t2, pot1);
 856     _mm_store_ss(ptrA, _mm_add_ss(pot1, _mm_load_ss(ptrA)));
 857     _mm_store_ss(ptrB, _mm_add_ss(pot2, _mm_load_ss(ptrB)));
 858 }
 859
 860
 861 #endif /* _kernelutil_x86_sse4_1_single_h_ */