src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/kernelutil_x86_sse2_single.h

   1 /*
   2  *                This source code is part of
   3  *
   4  *                 G   R   O   M   A   C   S
   5  *
   6  * Copyright (c) 2011-2012, The GROMACS Development Team
   7  *
   8  * Gromacs is a library for molecular simulation and trajectory analysis,
   9  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  10  * a full list of developers and information, check out http://www.gromacs.org
  11  *
  12  * This program is free software; you can redistribute it and/or modify it under
  13  * the terms of the GNU Lesser General Public License as published by the Free
  14  * Software Foundation; either version 2 of the License, or (at your option) any
  15  * later version.
  16  * As a special exception, you may use this file as part of a free software
  17  * library without restriction.  Specifically, if other files instantiate
  18  * templates or use macros or inline functions from this file, or you compile
  19  * this file and link it with other files to produce an executable, this
  20  * file does not by itself cause the resulting executable to be covered by
  21  * the GNU Lesser General Public License.
  22  *
  23  * In plain-speak: do not worry about classes/macros/templates either - only
  24  * changes to the library have to be LGPL, not an application linking with it.
  25  *
  26  * To help fund GROMACS development, we humbly ask that you cite
  27  * the papers people have written on it - you can find them on the website!
  28  */
  29 #ifndef _kernelutil_x86_sse2_single_h_
  30 #define _kernelutil_x86_sse2_single_h_
  31
  32 /* We require SSE2 now! */
  33
  34 #include <math.h>
  35
  36 #include "gmx_x86_sse2.h"
  37
  38
  39 /* Normal sum of four xmm registers */
  40 #define gmx_mm_sum4_ps(t0, t1, t2, t3)  _mm_add_ps(_mm_add_ps(t0, t1), _mm_add_ps(t2, t3))
  41
  42 static gmx_inline __m128
  43 gmx_mm_calc_rsq_ps(__m128 dx, __m128 dy, __m128 dz)
  44 {
  45     return _mm_add_ps( _mm_add_ps( _mm_mul_ps(dx, dx), _mm_mul_ps(dy, dy) ), _mm_mul_ps(dz, dz) );
  46 }
  47
  48 static int
  49 gmx_mm_any_lt(__m128 a, __m128 b)
  50 {
  51     return _mm_movemask_ps(_mm_cmplt_ps(a, b));
  52 }
  53
  54 /* Load a single value from 1-4 places, merge into xmm register */
  55
  56 static __m128
  57 gmx_mm_load_4real_swizzle_ps(const float * gmx_restrict ptrA,
  58                              const float * gmx_restrict ptrB,
  59                              const float * gmx_restrict ptrC,
  60                              const float * gmx_restrict ptrD)
  61 {
  62     __m128 t1, t2;
  63
  64     t1 = _mm_unpacklo_ps(_mm_load_ss(ptrA), _mm_load_ss(ptrC));
  65     t2 = _mm_unpacklo_ps(_mm_load_ss(ptrB), _mm_load_ss(ptrD));
  66     return _mm_unpacklo_ps(t1, t2);
  67 }
  68
  69 static void
  70 gmx_mm_store_4real_swizzle_ps(float * gmx_restrict ptrA,
  71                               float * gmx_restrict ptrB,
  72                               float * gmx_restrict ptrC,
  73                               float * gmx_restrict ptrD,
  74                               __m128               xmm1)
  75 {
  76     __m128 t2, t3, t4;
  77
  78     t3       = _mm_movehl_ps(_mm_setzero_ps(), xmm1);
  79     t2       = _mm_shuffle_ps(xmm1, xmm1, _MM_SHUFFLE(1, 1, 1, 1));
  80     t4       = _mm_shuffle_ps(t3, t3, _MM_SHUFFLE(1, 1, 1, 1));
  81     _mm_store_ss(ptrA, xmm1);
  82     _mm_store_ss(ptrB, t2);
  83     _mm_store_ss(ptrC, t3);
  84     _mm_store_ss(ptrD, t4);
  85 }
  86
  87 /* Similar to store, but increments value in memory */
  88 static void
  89 gmx_mm_increment_4real_swizzle_ps(float * gmx_restrict ptrA,
  90                                   float * gmx_restrict ptrB,
  91                                   float * gmx_restrict ptrC,
  92                                   float * gmx_restrict ptrD, __m128 xmm1)
  93 {
  94     __m128 tmp;
  95
  96     tmp = gmx_mm_load_4real_swizzle_ps(ptrA, ptrB, ptrC, ptrD);
  97     tmp = _mm_add_ps(tmp, xmm1);
  98     gmx_mm_store_4real_swizzle_ps(ptrA, ptrB, ptrC, ptrD, tmp);
  99 }
 100
 101
 102 static void
 103 gmx_mm_load_4pair_swizzle_ps(const float * gmx_restrict p1,
 104                              const float * gmx_restrict p2,
 105                              const float * gmx_restrict p3,
 106                              const float * gmx_restrict p4,
 107                              __m128 * gmx_restrict      c6,
 108                              __m128 * gmx_restrict      c12)
 109 {
 110     __m128 t1, t2, t3, t4;
 111
 112     t1   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)p1);   /* - - c12a  c6a */
 113     t2   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)p2);   /* - - c12b  c6b */
 114     t3   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)p3);   /* - - c12c  c6c */
 115     t4   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)p4);   /* - - c12d  c6d */
 116     t1   = _mm_unpacklo_ps(t1, t2);
 117     t2   = _mm_unpacklo_ps(t3, t4);
 118     *c6  = _mm_movelh_ps(t1, t2);
 119     *c12 = _mm_movehl_ps(t2, t1);
 120 }
 121
 122 /* Routines to load 1-4 rvec from 4 places.
 123  * We mainly use these to load coordinates. The extra routines
 124  * are very efficient for the water-water loops, since we e.g.
 125  * know that a TIP4p water has 4 atoms, so we should load 12 floats+shuffle.
 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 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 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            = _mm_loadu_ps(ptrA);
 270     t2            = _mm_loadu_ps(ptrB);
 271     t3            = _mm_loadu_ps(ptrC);
 272     t4            = _mm_loadu_ps(ptrD);
 273     _MM_TRANSPOSE4_PS(t1, t2, t3, t4);
 274     *x1           = t1;
 275     *y1           = t2;
 276     *z1           = t3;
 277     *x2           = t4;
 278     t1            = _mm_loadu_ps(ptrA+4);
 279     t2            = _mm_loadu_ps(ptrB+4);
 280     t3            = _mm_loadu_ps(ptrC+4);
 281     t4            = _mm_loadu_ps(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 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            = _mm_loadu_ps(ptrA);
 309     t2            = _mm_loadu_ps(ptrB);
 310     t3            = _mm_loadu_ps(ptrC);
 311     t4            = _mm_loadu_ps(ptrD);
 312     _MM_TRANSPOSE4_PS(t1, t2, t3, t4);
 313     *x1           = t1;
 314     *y1           = t2;
 315     *z1           = t3;
 316     *x2           = t4;
 317     t1            = _mm_loadu_ps(ptrA+4);
 318     t2            = _mm_loadu_ps(ptrB+4);
 319     t3            = _mm_loadu_ps(ptrC+4);
 320     t4            = _mm_loadu_ps(ptrD+4);
 321     _MM_TRANSPOSE4_PS(t1, t2, t3, t4);
 322     *y2           = t1;
 323     *z2           = t2;
 324     *x3           = t3;
 325     *y3           = t4;
 326     t1            = _mm_loadu_ps(ptrA+8);
 327     t2            = _mm_loadu_ps(ptrB+8);
 328     t3            = _mm_loadu_ps(ptrC+8);
 329     t4            = _mm_loadu_ps(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 static void
 339 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(float * gmx_restrict ptrA,
 340                                        float * gmx_restrict ptrB,
 341                                        float * gmx_restrict ptrC,
 342                                        float * gmx_restrict ptrD,
 343                                        __m128 x1, __m128 y1, __m128 z1)
 344 {
 345     __m128 t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12;
 346     t5          = _mm_unpacklo_ps(y1, z1);
 347     t6          = _mm_unpackhi_ps(y1, z1);
 348     t7          = _mm_shuffle_ps(x1, t5, _MM_SHUFFLE(1, 0, 0, 0));
 349     t8          = _mm_shuffle_ps(x1, t5, _MM_SHUFFLE(3, 2, 0, 1));
 350     t9          = _mm_shuffle_ps(x1, t6, _MM_SHUFFLE(1, 0, 0, 2));
 351     t10         = _mm_shuffle_ps(x1, t6, _MM_SHUFFLE(3, 2, 0, 3));
 352     t1          = _mm_load_ss(ptrA);
 353     t1          = _mm_loadh_pi(t1, (__m64 *)(ptrA+1));
 354     t1          = _mm_sub_ps(t1, t7);
 355     _mm_store_ss(ptrA, t1);
 356     _mm_storeh_pi((__m64 *)(ptrA+1), t1);
 357     t2          = _mm_load_ss(ptrB);
 358     t2          = _mm_loadh_pi(t2, (__m64 *)(ptrB+1));
 359     t2          = _mm_sub_ps(t2, t8);
 360     _mm_store_ss(ptrB, t2);
 361     _mm_storeh_pi((__m64 *)(ptrB+1), t2);
 362     t3          = _mm_load_ss(ptrC);
 363     t3          = _mm_loadh_pi(t3, (__m64 *)(ptrC+1));
 364     t3          = _mm_sub_ps(t3, t9);
 365     _mm_store_ss(ptrC, t3);
 366     _mm_storeh_pi((__m64 *)(ptrC+1), t3);
 367     t4          = _mm_load_ss(ptrD);
 368     t4          = _mm_loadh_pi(t4, (__m64 *)(ptrD+1));
 369     t4          = _mm_sub_ps(t4, t10);
 370     _mm_store_ss(ptrD, t4);
 371     _mm_storeh_pi((__m64 *)(ptrD+1), t4);
 372 }
 373
 374
 375
 376 #if defined (_MSC_VER) && defined(_M_IX86)
 377 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 378 #define gmx_mm_decrement_3rvec_4ptr_swizzle_ps(ptrA, ptrB, ptrC, ptrD, \
 379                                                _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3) \
 380     { \
 381         __m128 _t1, _t2, _t3, _t4, _t5, _t6, _t7, _t8, _t9, _t10; \
 382         __m128 _t11, _t12, _t13, _t14, _t15, _t16, _t17, _t18, _t19; \
 383         __m128 _t20, _t21, _t22, _t23, _t24, _t25; \
 384         _t13         = _mm_unpackhi_ps(_x1, _y1); \
 385         _x1          = _mm_unpacklo_ps(_x1, _y1); \
 386         _t14         = _mm_unpackhi_ps(_z1, _x2); \
 387         _z1          = _mm_unpacklo_ps(_z1, _x2); \
 388         _t15         = _mm_unpackhi_ps(_y2, _z2); \
 389         _y2          = _mm_unpacklo_ps(_y2, _z2); \
 390         _t16         = _mm_unpackhi_ps(_x3, _y3); \
 391         _x3          = _mm_unpacklo_ps(_x3, _y3); \
 392         _t17         = _mm_shuffle_ps(_z3, _z3, _MM_SHUFFLE(0, 0, 0, 1)); \
 393         _t18         = _mm_movehl_ps(_z3, _z3); \
 394         _t19         = _mm_shuffle_ps(_t18, _t18, _MM_SHUFFLE(0, 0, 0, 1)); \
 395         _t20         = _mm_movelh_ps(_x1, _z1); \
 396         _t21         = _mm_movehl_ps(_z1, _x1); \
 397         _t22         = _mm_movelh_ps(_t13, _t14); \
 398         _t14         = _mm_movehl_ps(_t14, _t13); \
 399         _t23         = _mm_movelh_ps(_y2, _x3); \
 400         _t24         = _mm_movehl_ps(_x3, _y2); \
 401         _t25         = _mm_movelh_ps(_t15, _t16); \
 402         _t16         = _mm_movehl_ps(_t16, _t15); \
 403         _t1          = _mm_loadu_ps(ptrA); \
 404         _t2          = _mm_loadu_ps(ptrA+4); \
 405         _t3          = _mm_load_ss(ptrA+8); \
 406         _t1          = _mm_sub_ps(_t1, _t20); \
 407         _t2          = _mm_sub_ps(_t2, _t23); \
 408         _t3          = _mm_sub_ss(_t3, _z3); \
 409         _mm_storeu_ps(ptrA, _t1); \
 410         _mm_storeu_ps(ptrA+4, _t2); \
 411         _mm_store_ss(ptrA+8, _t3); \
 412         _t4          = _mm_loadu_ps(ptrB); \
 413         _t5          = _mm_loadu_ps(ptrB+4); \
 414         _t6          = _mm_load_ss(ptrB+8); \
 415         _t4          = _mm_sub_ps(_t4, _t21); \
 416         _t5          = _mm_sub_ps(_t5, _t24); \
 417         _t6          = _mm_sub_ss(_t6, _t17); \
 418         _mm_storeu_ps(ptrB, _t4); \
 419         _mm_storeu_ps(ptrB+4, _t5); \
 420         _mm_store_ss(ptrB+8, _t6); \
 421         _t7          = _mm_loadu_ps(ptrC); \
 422         _t8          = _mm_loadu_ps(ptrC+4); \
 423         _t9          = _mm_load_ss(ptrC+8); \
 424         _t7          = _mm_sub_ps(_t7, _t22); \
 425         _t8          = _mm_sub_ps(_t8, _t25); \
 426         _t9          = _mm_sub_ss(_t9, _t18); \
 427         _mm_storeu_ps(ptrC, _t7); \
 428         _mm_storeu_ps(ptrC+4, _t8); \
 429         _mm_store_ss(ptrC+8, _t9); \
 430         _t10         = _mm_loadu_ps(ptrD); \
 431         _t11         = _mm_loadu_ps(ptrD+4); \
 432         _t12         = _mm_load_ss(ptrD+8); \
 433         _t10         = _mm_sub_ps(_t10, _t14); \
 434         _t11         = _mm_sub_ps(_t11, _t16); \
 435         _t12         = _mm_sub_ss(_t12, _t19); \
 436         _mm_storeu_ps(ptrD, _t10); \
 437         _mm_storeu_ps(ptrD+4, _t11); \
 438         _mm_store_ss(ptrD+8, _t12); \
 439     }
 440 #else
 441 /* Real function for sane compilers */
 442 static void
 443 gmx_mm_decrement_3rvec_4ptr_swizzle_ps(float * gmx_restrict ptrA, float * gmx_restrict ptrB,
 444                                        float * gmx_restrict ptrC, float * gmx_restrict ptrD,
 445                                        __m128 x1, __m128 y1, __m128 z1,
 446                                        __m128 x2, __m128 y2, __m128 z2,
 447                                        __m128 x3, __m128 y3, __m128 z3)
 448 {
 449     __m128 t1, t2, t3, t4, t5, t6, t7, t8, t9, t10;
 450     __m128 t11, t12, t13, t14, t15, t16, t17, t18, t19;
 451     __m128 t20, t21, t22, t23, t24, t25;
 452
 453     t13         = _mm_unpackhi_ps(x1, y1);
 454     x1          = _mm_unpacklo_ps(x1, y1);
 455     t14         = _mm_unpackhi_ps(z1, x2);
 456     z1          = _mm_unpacklo_ps(z1, x2);
 457     t15         = _mm_unpackhi_ps(y2, z2);
 458     y2          = _mm_unpacklo_ps(y2, z2);
 459     t16         = _mm_unpackhi_ps(x3, y3);
 460     x3          = _mm_unpacklo_ps(x3, y3);
 461     t17         = _mm_shuffle_ps(z3, z3, _MM_SHUFFLE(0, 0, 0, 1));
 462     t18         = _mm_movehl_ps(z3, z3);
 463     t19         = _mm_shuffle_ps(t18, t18, _MM_SHUFFLE(0, 0, 0, 1));
 464     t20         = _mm_movelh_ps(x1, z1);
 465     t21         = _mm_movehl_ps(z1, x1);
 466     t22         = _mm_movelh_ps(t13, t14);
 467     t14         = _mm_movehl_ps(t14, t13);
 468     t23         = _mm_movelh_ps(y2, x3);
 469     t24         = _mm_movehl_ps(x3, y2);
 470     t25         = _mm_movelh_ps(t15, t16);
 471     t16         = _mm_movehl_ps(t16, t15);
 472     t1          = _mm_loadu_ps(ptrA);
 473     t2          = _mm_loadu_ps(ptrA+4);
 474     t3          = _mm_load_ss(ptrA+8);
 475     t1          = _mm_sub_ps(t1, t20);
 476     t2          = _mm_sub_ps(t2, t23);
 477     t3          = _mm_sub_ss(t3, z3);
 478     _mm_storeu_ps(ptrA, t1);
 479     _mm_storeu_ps(ptrA+4, t2);
 480     _mm_store_ss(ptrA+8, t3);
 481     t4          = _mm_loadu_ps(ptrB);
 482     t5          = _mm_loadu_ps(ptrB+4);
 483     t6          = _mm_load_ss(ptrB+8);
 484     t4          = _mm_sub_ps(t4, t21);
 485     t5          = _mm_sub_ps(t5, t24);
 486     t6          = _mm_sub_ss(t6, t17);
 487     _mm_storeu_ps(ptrB, t4);
 488     _mm_storeu_ps(ptrB+4, t5);
 489     _mm_store_ss(ptrB+8, t6);
 490     t7          = _mm_loadu_ps(ptrC);
 491     t8          = _mm_loadu_ps(ptrC+4);
 492     t9          = _mm_load_ss(ptrC+8);
 493     t7          = _mm_sub_ps(t7, t22);
 494     t8          = _mm_sub_ps(t8, t25);
 495     t9          = _mm_sub_ss(t9, t18);
 496     _mm_storeu_ps(ptrC, t7);
 497     _mm_storeu_ps(ptrC+4, t8);
 498     _mm_store_ss(ptrC+8, t9);
 499     t10         = _mm_loadu_ps(ptrD);
 500     t11         = _mm_loadu_ps(ptrD+4);
 501     t12         = _mm_load_ss(ptrD+8);
 502     t10         = _mm_sub_ps(t10, t14);
 503     t11         = _mm_sub_ps(t11, t16);
 504     t12         = _mm_sub_ss(t12, t19);
 505     _mm_storeu_ps(ptrD, t10);
 506     _mm_storeu_ps(ptrD+4, t11);
 507     _mm_store_ss(ptrD+8, t12);
 508 }
 509 #endif
 510
 511
 512 #if defined (_MSC_VER) && defined(_M_IX86)
 513 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 514 #define gmx_mm_decrement_4rvec_4ptr_swizzle_ps(ptrA, ptrB, ptrC, ptrD, \
 515                                                _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3, _x4, _y4, _z4) \
 516     { \
 517         __m128 _t1, _t2, _t3, _t4, _t5, _t6, _t7, _t8, _t9, _t10, _t11; \
 518         __m128 _t12, _t13, _t14, _t15, _t16, _t17, _t18, _t19, _t20, _t21, _t22; \
 519         __m128 _t23, _t24; \
 520         _t13         = _mm_unpackhi_ps(_x1, _y1); \
 521         _x1          = _mm_unpacklo_ps(_x1, _y1); \
 522         _t14         = _mm_unpackhi_ps(_z1, _x2); \
 523         _z1          = _mm_unpacklo_ps(_z1, _x2); \
 524         _t15         = _mm_unpackhi_ps(_y2, _z2); \
 525         _y2          = _mm_unpacklo_ps(_y2, _z2); \
 526         _t16         = _mm_unpackhi_ps(_x3, _y3); \
 527         _x3          = _mm_unpacklo_ps(_x3, _y3); \
 528         _t17         = _mm_unpackhi_ps(_z3, _x4); \
 529         _z3          = _mm_unpacklo_ps(_z3, _x4); \
 530         _t18         = _mm_unpackhi_ps(_y4, _z4); \
 531         _y4          = _mm_unpacklo_ps(_y4, _z4); \
 532         _t19         = _mm_movelh_ps(_x1, _z1); \
 533         _z1          = _mm_movehl_ps(_z1, _x1); \
 534         _t20         = _mm_movelh_ps(_t13, _t14); \
 535         _t14         = _mm_movehl_ps(_t14, _t13); \
 536         _t21         = _mm_movelh_ps(_y2, _x3); \
 537         _x3          = _mm_movehl_ps(_x3, _y2); \
 538         _t22         = _mm_movelh_ps(_t15, _t16); \
 539         _t16         = _mm_movehl_ps(_t16, _t15); \
 540         _t23         = _mm_movelh_ps(_z3, _y4); \
 541         _y4          = _mm_movehl_ps(_y4, _z3); \
 542         _t24         = _mm_movelh_ps(_t17, _t18); \
 543         _t18         = _mm_movehl_ps(_t18, _t17); \
 544         _t1          = _mm_loadu_ps(ptrA); \
 545         _t2          = _mm_loadu_ps(ptrA+4); \
 546         _t3          = _mm_loadu_ps(ptrA+8); \
 547         _t1          = _mm_sub_ps(_t1, _t19); \
 548         _t2          = _mm_sub_ps(_t2, _t21); \
 549         _t3          = _mm_sub_ps(_t3, _t23); \
 550         _mm_storeu_ps(ptrA, _t1); \
 551         _mm_storeu_ps(ptrA+4, _t2); \
 552         _mm_storeu_ps(ptrA+8, _t3); \
 553         _t4          = _mm_loadu_ps(ptrB); \
 554         _t5          = _mm_loadu_ps(ptrB+4); \
 555         _t6          = _mm_loadu_ps(ptrB+8); \
 556         _t4          = _mm_sub_ps(_t4, _z1); \
 557         _t5          = _mm_sub_ps(_t5, _x3); \
 558         _t6          = _mm_sub_ps(_t6, _y4); \
 559         _mm_storeu_ps(ptrB, _t4); \
 560         _mm_storeu_ps(ptrB+4, _t5); \
 561         _mm_storeu_ps(ptrB+8, _t6); \
 562         _t7          = _mm_loadu_ps(ptrC); \
 563         _t8          = _mm_loadu_ps(ptrC+4); \
 564         _t9          = _mm_loadu_ps(ptrC+8); \
 565         _t7          = _mm_sub_ps(_t7, _t20); \
 566         _t8          = _mm_sub_ps(_t8, _t22); \
 567         _t9          = _mm_sub_ps(_t9, _t24); \
 568         _mm_storeu_ps(ptrC, _t7); \
 569         _mm_storeu_ps(ptrC+4, _t8); \
 570         _mm_storeu_ps(ptrC+8, _t9); \
 571         _t10         = _mm_loadu_ps(ptrD); \
 572         _t11         = _mm_loadu_ps(ptrD+4); \
 573         _t12         = _mm_loadu_ps(ptrD+8); \
 574         _t10         = _mm_sub_ps(_t10, _t14); \
 575         _t11         = _mm_sub_ps(_t11, _t16); \
 576         _t12         = _mm_sub_ps(_t12, _t18); \
 577         _mm_storeu_ps(ptrD, _t10); \
 578         _mm_storeu_ps(ptrD+4, _t11); \
 579         _mm_storeu_ps(ptrD+8, _t12); \
 580     }
 581 #else
 582 /* Real function for sane compilers */
 583 static void
 584 gmx_mm_decrement_4rvec_4ptr_swizzle_ps(float * gmx_restrict ptrA, float * gmx_restrict ptrB,
 585                                        float * gmx_restrict ptrC, float * gmx_restrict ptrD,
 586                                        __m128 x1, __m128 y1, __m128 z1,
 587                                        __m128 x2, __m128 y2, __m128 z2,
 588                                        __m128 x3, __m128 y3, __m128 z3,
 589                                        __m128 x4, __m128 y4, __m128 z4)
 590 {
 591     __m128 t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11;
 592     __m128 t12, t13, t14, t15, t16, t17, t18, t19, t20, t21, t22;
 593     __m128 t23, t24;
 594     t13         = _mm_unpackhi_ps(x1, y1);
 595     x1          = _mm_unpacklo_ps(x1, y1);
 596     t14         = _mm_unpackhi_ps(z1, x2);
 597     z1          = _mm_unpacklo_ps(z1, x2);
 598     t15         = _mm_unpackhi_ps(y2, z2);
 599     y2          = _mm_unpacklo_ps(y2, z2);
 600     t16         = _mm_unpackhi_ps(x3, y3);
 601     x3          = _mm_unpacklo_ps(x3, y3);
 602     t17         = _mm_unpackhi_ps(z3, x4);
 603     z3          = _mm_unpacklo_ps(z3, x4);
 604     t18         = _mm_unpackhi_ps(y4, z4);
 605     y4          = _mm_unpacklo_ps(y4, z4);
 606     t19         = _mm_movelh_ps(x1, z1);
 607     z1          = _mm_movehl_ps(z1, x1);
 608     t20         = _mm_movelh_ps(t13, t14);
 609     t14         = _mm_movehl_ps(t14, t13);
 610     t21         = _mm_movelh_ps(y2, x3);
 611     x3          = _mm_movehl_ps(x3, y2);
 612     t22         = _mm_movelh_ps(t15, t16);
 613     t16         = _mm_movehl_ps(t16, t15);
 614     t23         = _mm_movelh_ps(z3, y4);
 615     y4          = _mm_movehl_ps(y4, z3);
 616     t24         = _mm_movelh_ps(t17, t18);
 617     t18         = _mm_movehl_ps(t18, t17);
 618     t1          = _mm_loadu_ps(ptrA);
 619     t2          = _mm_loadu_ps(ptrA+4);
 620     t3          = _mm_loadu_ps(ptrA+8);
 621     t1          = _mm_sub_ps(t1, t19);
 622     t2          = _mm_sub_ps(t2, t21);
 623     t3          = _mm_sub_ps(t3, t23);
 624     _mm_storeu_ps(ptrA, t1);
 625     _mm_storeu_ps(ptrA+4, t2);
 626     _mm_storeu_ps(ptrA+8, t3);
 627     t4          = _mm_loadu_ps(ptrB);
 628     t5          = _mm_loadu_ps(ptrB+4);
 629     t6          = _mm_loadu_ps(ptrB+8);
 630     t4          = _mm_sub_ps(t4, z1);
 631     t5          = _mm_sub_ps(t5, x3);
 632     t6          = _mm_sub_ps(t6, y4);
 633     _mm_storeu_ps(ptrB, t4);
 634     _mm_storeu_ps(ptrB+4, t5);
 635     _mm_storeu_ps(ptrB+8, t6);
 636     t7          = _mm_loadu_ps(ptrC);
 637     t8          = _mm_loadu_ps(ptrC+4);
 638     t9          = _mm_loadu_ps(ptrC+8);
 639     t7          = _mm_sub_ps(t7, t20);
 640     t8          = _mm_sub_ps(t8, t22);
 641     t9          = _mm_sub_ps(t9, t24);
 642     _mm_storeu_ps(ptrC, t7);
 643     _mm_storeu_ps(ptrC+4, t8);
 644     _mm_storeu_ps(ptrC+8, t9);
 645     t10         = _mm_loadu_ps(ptrD);
 646     t11         = _mm_loadu_ps(ptrD+4);
 647     t12         = _mm_loadu_ps(ptrD+8);
 648     t10         = _mm_sub_ps(t10, t14);
 649     t11         = _mm_sub_ps(t11, t16);
 650     t12         = _mm_sub_ps(t12, t18);
 651     _mm_storeu_ps(ptrD, t10);
 652     _mm_storeu_ps(ptrD+4, t11);
 653     _mm_storeu_ps(ptrD+8, t12);
 654 }
 655 #endif
 656
 657
 658 static gmx_inline void
 659 gmx_mm_update_iforce_1atom_swizzle_ps(__m128 fix1, __m128 fiy1, __m128 fiz1,
 660                                       float * gmx_restrict fptr,
 661                                       float * gmx_restrict fshiftptr)
 662 {
 663     __m128 t1, t2, t3;
 664
 665     /* transpose data */
 666     t1 = fix1;
 667     _MM_TRANSPOSE4_PS(fix1, t1, fiy1, fiz1);
 668     fix1 = _mm_add_ps(_mm_add_ps(fix1, t1), _mm_add_ps(fiy1, fiz1));
 669
 670     t2 = _mm_load_ss(fptr);
 671     t2 = _mm_loadh_pi(t2, (__m64 *)(fptr+1));
 672     t3 = _mm_load_ss(fshiftptr);
 673     t3 = _mm_loadh_pi(t3, (__m64 *)(fshiftptr+1));
 674
 675     t2 = _mm_add_ps(t2, fix1);
 676     t3 = _mm_add_ps(t3, fix1);
 677
 678     _mm_store_ss(fptr, t2);
 679     _mm_storeh_pi((__m64 *)(fptr+1), t2);
 680     _mm_store_ss(fshiftptr, t3);
 681     _mm_storeh_pi((__m64 *)(fshiftptr+1), t3);
 682 }
 683
 684 #if defined (_MSC_VER) && defined(_M_IX86)
 685 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 686 #define gmx_mm_update_iforce_3atom_swizzle_ps(fix1, fiy1, fiz1, fix2, fiy2, fiz2, fix3, fiy3, fiz3, \
 687                                               fptr, fshiftptr) \
 688     { \
 689         __m128 _t1, _t2, _t3, _t4; \
 690 \
 691         _MM_TRANSPOSE4_PS(fix1, fiy1, fiz1, fix2); \
 692         _MM_TRANSPOSE4_PS(fiy2, fiz2, fix3, fiy3); \
 693         _t2   = _mm_movehl_ps(_mm_setzero_ps(), fiz3); \
 694         _t1   = _mm_shuffle_ps(fiz3, fiz3, _MM_SHUFFLE(0, 0, 0, 1)); \
 695         _t3   = _mm_shuffle_ps(_t2, _t2, _MM_SHUFFLE(0, 0, 0, 1)); \
 696         fix1  = _mm_add_ps(_mm_add_ps(fix1, fiy1), _mm_add_ps(fiz1, fix2)); \
 697         fiy2  = _mm_add_ps(_mm_add_ps(fiy2, fiz2), _mm_add_ps(fix3, fiy3)); \
 698         fiz3  = _mm_add_ss(_mm_add_ps(fiz3, _t1), _mm_add_ps(_t2, _t3)); \
 699         _mm_storeu_ps(fptr,  _mm_add_ps(fix1, _mm_loadu_ps(fptr)  )); \
 700         _mm_storeu_ps(fptr+4, _mm_add_ps(fiy2, _mm_loadu_ps(fptr+4))); \
 701         _mm_store_ss (fptr+8, _mm_add_ss(fiz3, _mm_load_ss(fptr+8) )); \
 702         _t4 = _mm_load_ss(fshiftptr+2); \
 703         _t4 = _mm_loadh_pi(_t4, (__m64 *)(fshiftptr)); \
 704         _t1 = _mm_shuffle_ps(fiz3, fix1, _MM_SHUFFLE(1, 0, 0, 0)); \
 705         _t2 = _mm_shuffle_ps(fix1, fiy2, _MM_SHUFFLE(3, 2, 2, 2)); \
 706         _t3 = _mm_shuffle_ps(fiy2, fix1, _MM_SHUFFLE(3, 3, 0, 1)); \
 707         _t3 = _mm_shuffle_ps(_t3, _t3, _MM_SHUFFLE(1, 2, 0, 0)); \
 708         _t1 = _mm_add_ps(_t1, _t2); \
 709         _t3 = _mm_add_ps(_t3, _t4); \
 710         _t1 = _mm_add_ps(_t1, _t3); \
 711         _mm_store_ss(fshiftptr+2, _t1); \
 712         _mm_storeh_pi((__m64 *)(fshiftptr), _t1); \
 713     }
 714 #else
 715 /* Real function for sane compilers */
 716 static gmx_inline void
 717 gmx_mm_update_iforce_3atom_swizzle_ps(__m128 fix1, __m128 fiy1, __m128 fiz1,
 718                                       __m128 fix2, __m128 fiy2, __m128 fiz2,
 719                                       __m128 fix3, __m128 fiy3, __m128 fiz3,
 720                                       float * gmx_restrict fptr,
 721                                       float * gmx_restrict fshiftptr)
 722 {
 723     __m128 t1, t2, t3, t4;
 724
 725     /* transpose data */
 726     _MM_TRANSPOSE4_PS(fix1, fiy1, fiz1, fix2);
 727     _MM_TRANSPOSE4_PS(fiy2, fiz2, fix3, fiy3);
 728     t2   = _mm_movehl_ps(_mm_setzero_ps(), fiz3);
 729     t1   = _mm_shuffle_ps(fiz3, fiz3, _MM_SHUFFLE(0, 0, 0, 1));
 730     t3   = _mm_shuffle_ps(t2, t2, _MM_SHUFFLE(0, 0, 0, 1));
 731
 732     fix1 = _mm_add_ps(_mm_add_ps(fix1, fiy1), _mm_add_ps(fiz1, fix2));
 733     fiy2 = _mm_add_ps(_mm_add_ps(fiy2, fiz2), _mm_add_ps(fix3, fiy3));
 734     fiz3 = _mm_add_ss(_mm_add_ps(fiz3, t1), _mm_add_ps(t2, t3));
 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         _MM_TRANSPOSE4_PS(fix1, fiy1, fiz1, fix2); \
 764         _MM_TRANSPOSE4_PS(fiy2, fiz2, fix3, fiy3); \
 765         _MM_TRANSPOSE4_PS(fiz3, fix4, fiy4, fiz4); \
 766         fix1 = _mm_add_ps(_mm_add_ps(fix1, fiy1), _mm_add_ps(fiz1, fix2)); \
 767         fiy2 = _mm_add_ps(_mm_add_ps(fiy2, fiz2), _mm_add_ps(fix3, fiy3)); \
 768         fiz3 = _mm_add_ps(_mm_add_ps(fiz3, fix4), _mm_add_ps(fiy4, fiz4)); \
 769         _mm_storeu_ps(fptr,  _mm_add_ps(fix1, _mm_loadu_ps(fptr)  )); \
 770         _mm_storeu_ps(fptr+4, _mm_add_ps(fiy2, _mm_loadu_ps(fptr+4))); \
 771         _mm_storeu_ps(fptr+8, _mm_add_ps(fiz3, _mm_loadu_ps(fptr+8))); \
 772         _t5 = _mm_load_ss(fshiftptr+2); \
 773         _t5 = _mm_loadh_pi(_t5, (__m64 *)(fshiftptr)); \
 774         _t1 = _mm_shuffle_ps(fix1, fix1, _MM_SHUFFLE(1, 0, 2, 2)); \
 775         _t2 = _mm_shuffle_ps(fiy2, fiy2, _MM_SHUFFLE(3, 2, 1, 1)); \
 776         _t3 = _mm_shuffle_ps(fiz3, fiz3, _MM_SHUFFLE(2, 1, 0, 0)); \
 777         _t4 = _mm_shuffle_ps(fix1, fiy2, _MM_SHUFFLE(0, 0, 3, 3)); \
 778         _t4 = _mm_shuffle_ps(fiz3, _t4, _MM_SHUFFLE(2, 0, 3, 3)); \
 779         _t1 = _mm_add_ps(_t1, _t2); \
 780         _t3 = _mm_add_ps(_t3, _t4); \
 781         _t1 = _mm_add_ps(_t1, _t3); \
 782         _t5 = _mm_add_ps(_t5, _t1); \
 783         _mm_store_ss(fshiftptr+2, _t5); \
 784         _mm_storeh_pi((__m64 *)(fshiftptr), _t5); \
 785     }
 786 #else
 787 /* Real function for sane compilers */
 788 static gmx_inline void
 789 gmx_mm_update_iforce_4atom_swizzle_ps(__m128 fix1, __m128 fiy1, __m128 fiz1,
 790                                       __m128 fix2, __m128 fiy2, __m128 fiz2,
 791                                       __m128 fix3, __m128 fiy3, __m128 fiz3,
 792                                       __m128 fix4, __m128 fiy4, __m128 fiz4,
 793                                       float * gmx_restrict fptr,
 794                                       float * gmx_restrict fshiftptr)
 795 {
 796     __m128 t1, t2, t3, t4, t5;
 797
 798     /* transpose data */
 799     _MM_TRANSPOSE4_PS(fix1, fiy1, fiz1, fix2);
 800     _MM_TRANSPOSE4_PS(fiy2, fiz2, fix3, fiy3);
 801     _MM_TRANSPOSE4_PS(fiz3, fix4, fiy4, fiz4);
 802
 803     fix1 = _mm_add_ps(_mm_add_ps(fix1, fiy1), _mm_add_ps(fiz1, fix2));
 804     fiy2 = _mm_add_ps(_mm_add_ps(fiy2, fiz2), _mm_add_ps(fix3, fiy3));
 805     fiz3 = _mm_add_ps(_mm_add_ps(fiz3, fix4), _mm_add_ps(fiy4, fiz4));
 806
 807     _mm_storeu_ps(fptr,  _mm_add_ps(fix1, _mm_loadu_ps(fptr)  ));
 808     _mm_storeu_ps(fptr+4, _mm_add_ps(fiy2, _mm_loadu_ps(fptr+4)));
 809     _mm_storeu_ps(fptr+8, _mm_add_ps(fiz3, _mm_loadu_ps(fptr+8)));
 810
 811     t5 = _mm_load_ss(fshiftptr+2);
 812     t5 = _mm_loadh_pi(t5, (__m64 *)(fshiftptr));
 813
 814     t1 = _mm_shuffle_ps(fix1, fix1, _MM_SHUFFLE(1, 0, 2, 2));
 815     t2 = _mm_shuffle_ps(fiy2, fiy2, _MM_SHUFFLE(3, 2, 1, 1));
 816     t3 = _mm_shuffle_ps(fiz3, fiz3, _MM_SHUFFLE(2, 1, 0, 0));
 817     t4 = _mm_shuffle_ps(fix1, fiy2, _MM_SHUFFLE(0, 0, 3, 3));
 818     t4 = _mm_shuffle_ps(fiz3, t4, _MM_SHUFFLE(2, 0, 3, 3));
 819
 820     t1 = _mm_add_ps(t1, t2);
 821     t3 = _mm_add_ps(t3, t4);
 822     t1 = _mm_add_ps(t1, t3);
 823     t5 = _mm_add_ps(t5, t1);
 824
 825     _mm_store_ss(fshiftptr+2, t5);
 826     _mm_storeh_pi((__m64 *)(fshiftptr), t5);
 827 }
 828 #endif
 829
 830
 831 static void
 832 gmx_mm_update_1pot_ps(__m128 pot1, float * gmx_restrict ptrA)
 833 {
 834     pot1 = _mm_add_ps(pot1, _mm_movehl_ps(_mm_setzero_ps(), pot1));
 835     pot1 = _mm_add_ps(pot1, _mm_shuffle_ps(pot1, pot1, _MM_SHUFFLE(0, 0, 0, 1)));
 836     _mm_store_ss(ptrA, _mm_add_ss(pot1, _mm_load_ss(ptrA)));
 837 }
 838
 839 static void
 840 gmx_mm_update_2pot_ps(__m128 pot1, float * gmx_restrict ptrA,
 841                       __m128 pot2, float * gmx_restrict ptrB)
 842 {
 843     __m128 t1, t2;
 844     t1   = _mm_movehl_ps(pot2, pot1);
 845     t2   = _mm_movelh_ps(pot1, pot2);
 846     t1   = _mm_add_ps(t1, t2);
 847     t2   = _mm_shuffle_ps(t1, t1, _MM_SHUFFLE(3, 3, 1, 1));
 848     pot1 = _mm_add_ps(t1, t2);
 849     pot2 = _mm_movehl_ps(t2, pot1);
 850     _mm_store_ss(ptrA, _mm_add_ss(pot1, _mm_load_ss(ptrA)));
 851     _mm_store_ss(ptrB, _mm_add_ss(pot2, _mm_load_ss(ptrB)));
 852 }
 853
 854
 855 #endif /* _kernelutil_x86_sse2_single_h_ */