src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/kernelutil_x86_avx_128_fma_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_avx_128_fma_single_h_
  36 #define _kernelutil_x86_avx_128_fma_single_h_
  37
  38
  39 #include <math.h>
  40
  41 #include "gromacs/simd/general_x86_avx_128_fma.h"
  42
  43 /* Normal sum of four xmm registers */
  44 #define gmx_mm_sum4_ps(t0, t1, t2, t3)  _mm_add_ps(_mm_add_ps(t0, t1), _mm_add_ps(t2, t3))
  45
  46 static gmx_inline int
  47 gmx_mm_any_lt(__m128 a, __m128 b)
  48 {
  49     return _mm_movemask_ps(_mm_cmplt_ps(a, b));
  50 }
  51
  52 static gmx_inline __m128
  53 gmx_mm_calc_rsq_ps(__m128 dx, __m128 dy, __m128 dz)
  54 {
  55     return _mm_macc_ps(dx, dx, _mm_macc_ps(dy, dy, _mm_mul_ps(dz, dz)));
  56 }
  57
  58 /* Load a single value from 1-4 places, merge into xmm register */
  59
  60 static gmx_inline __m128
  61 gmx_mm_load_4real_swizzle_ps(const float * gmx_restrict ptrA,
  62                              const float * gmx_restrict ptrB,
  63                              const float * gmx_restrict ptrC,
  64                              const float * gmx_restrict ptrD)
  65 {
  66     __m128 t1, t2;
  67
  68     t1 = _mm_unpacklo_ps(_mm_load_ss(ptrA), _mm_load_ss(ptrC));
  69     t2 = _mm_unpacklo_ps(_mm_load_ss(ptrB), _mm_load_ss(ptrD));
  70     return _mm_unpacklo_ps(t1, t2);
  71 }
  72
  73
  74 static gmx_inline void
  75 gmx_mm_store_4real_swizzle_ps(float * gmx_restrict ptrA,
  76                               float * gmx_restrict ptrB,
  77                               float * gmx_restrict ptrC,
  78                               float * gmx_restrict ptrD, __m128 xmm1)
  79 {
  80     __m128 t2, t3, t4;
  81
  82     t2       = _mm_permute_ps(xmm1, _MM_SHUFFLE(1, 1, 1, 1));
  83     t3       = _mm_permute_ps(xmm1, _MM_SHUFFLE(2, 2, 2, 2));
  84     t4       = _mm_permute_ps(xmm1, _MM_SHUFFLE(3, 3, 3, 3));
  85     _mm_store_ss(ptrA, xmm1);
  86     _mm_store_ss(ptrB, t2);
  87     _mm_store_ss(ptrC, t3);
  88     _mm_store_ss(ptrD, t4);
  89 }
  90
  91
  92 static gmx_inline void
  93 gmx_mm_increment_4real_swizzle_ps(float * gmx_restrict ptrA,
  94                                   float * gmx_restrict ptrB,
  95                                   float * gmx_restrict ptrC,
  96                                   float * gmx_restrict ptrD, __m128 xmm1)
  97 {
  98     __m128 tmp;
  99
 100     tmp = gmx_mm_load_4real_swizzle_ps(ptrA, ptrB, ptrC, ptrD);
 101     tmp = _mm_add_ps(tmp, xmm1);
 102     gmx_mm_store_4real_swizzle_ps(ptrA, ptrB, ptrC, ptrD, tmp);
 103 }
 104
 105
 106 static gmx_inline void
 107 gmx_mm_load_4pair_swizzle_ps(const float * gmx_restrict p1,
 108                              const float * gmx_restrict p2,
 109                              const float * gmx_restrict p3,
 110                              const float * gmx_restrict p4,
 111                              __m128 * gmx_restrict c6, __m128 * gmx_restrict c12)
 112 {
 113     __m128 t1, t2, t3, t4;
 114     t1   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)p1);
 115     t2   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)p2);
 116     t3   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)p3);
 117     t4   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)p4);
 118     t1   = _mm_unpacklo_ps(t1, t3);
 119     t2   = _mm_unpacklo_ps(t2, t4);
 120     *c6  = _mm_unpacklo_ps(t1, t2);
 121     *c12 = _mm_unpackhi_ps(t1, t2);
 122 }
 123
 124
 125
 126
 127 static gmx_inline void
 128 gmx_mm_load_shift_and_1rvec_broadcast_ps(const float * gmx_restrict xyz_shift,
 129                                          const float * gmx_restrict xyz,
 130                                          __m128 * gmx_restrict      x1,
 131                                          __m128 * gmx_restrict      y1,
 132                                          __m128 * gmx_restrict      z1)
 133 {
 134     __m128 t1, t2, t3, t4;
 135
 136     t1   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)xyz_shift);
 137     t2   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)xyz);
 138     t3   = _mm_load_ss(xyz_shift+2);
 139     t4   = _mm_load_ss(xyz+2);
 140     t1   = _mm_add_ps(t1, t2);
 141     t3   = _mm_add_ss(t3, t4);
 142
 143     *x1  = _mm_permute_ps(t1, _MM_SHUFFLE(0, 0, 0, 0));
 144     *y1  = _mm_permute_ps(t1, _MM_SHUFFLE(1, 1, 1, 1));
 145     *z1  = _mm_permute_ps(t3, _MM_SHUFFLE(0, 0, 0, 0));
 146 }
 147
 148
 149 static gmx_inline void
 150 gmx_mm_load_shift_and_3rvec_broadcast_ps(const float * gmx_restrict xyz_shift,
 151                                          const float * gmx_restrict xyz,
 152                                          __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 153                                          __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 154                                          __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3)
 155 {
 156     __m128 tA, tB;
 157     __m128 t1, t2, t3, t4, t5, t6;
 158
 159     tA   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)xyz_shift);
 160     tB   = _mm_load_ss(xyz_shift+2);
 161
 162     t1   = _mm_loadu_ps(xyz);
 163     t2   = _mm_loadu_ps(xyz+4);
 164     t3   = _mm_load_ss(xyz+8);
 165
 166     tA   = _mm_movelh_ps(tA, tB);
 167     t4   = _mm_permute_ps(tA, _MM_SHUFFLE(0, 2, 1, 0));
 168     t5   = _mm_permute_ps(tA, _MM_SHUFFLE(1, 0, 2, 1));
 169     t6   = _mm_permute_ps(tA, _MM_SHUFFLE(2, 1, 0, 2));
 170
 171     t1   = _mm_add_ps(t1, t4);
 172     t2   = _mm_add_ps(t2, t5);
 173     t3   = _mm_add_ss(t3, t6);
 174
 175     *x1  = _mm_permute_ps(t1, _MM_SHUFFLE(0, 0, 0, 0));
 176     *y1  = _mm_permute_ps(t1, _MM_SHUFFLE(1, 1, 1, 1));
 177     *z1  = _mm_permute_ps(t1, _MM_SHUFFLE(2, 2, 2, 2));
 178     *x2  = _mm_permute_ps(t1, _MM_SHUFFLE(3, 3, 3, 3));
 179     *y2  = _mm_permute_ps(t2, _MM_SHUFFLE(0, 0, 0, 0));
 180     *z2  = _mm_permute_ps(t2, _MM_SHUFFLE(1, 1, 1, 1));
 181     *x3  = _mm_permute_ps(t2, _MM_SHUFFLE(2, 2, 2, 2));
 182     *y3  = _mm_permute_ps(t2, _MM_SHUFFLE(3, 3, 3, 3));
 183     *z3  = _mm_permute_ps(t3, _MM_SHUFFLE(0, 0, 0, 0));
 184 }
 185
 186
 187 static gmx_inline void
 188 gmx_mm_load_shift_and_4rvec_broadcast_ps(const float * gmx_restrict xyz_shift,
 189                                          const float * gmx_restrict xyz,
 190                                          __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 191                                          __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 192                                          __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3,
 193                                          __m128 * gmx_restrict x4, __m128 * gmx_restrict y4, __m128 * gmx_restrict z4)
 194 {
 195     __m128 tA, tB;
 196     __m128 t1, t2, t3, t4, t5, t6;
 197
 198     tA   = _mm_loadl_pi(_mm_setzero_ps(), (__m64 *)xyz_shift);
 199     tB   = _mm_load_ss(xyz_shift+2);
 200
 201     t1   = _mm_loadu_ps(xyz);
 202     t2   = _mm_loadu_ps(xyz+4);
 203     t3   = _mm_loadu_ps(xyz+8);
 204
 205     tA   = _mm_movelh_ps(tA, tB);
 206     t4   = _mm_permute_ps(tA, _MM_SHUFFLE(0, 2, 1, 0));
 207     t5   = _mm_permute_ps(tA, _MM_SHUFFLE(1, 0, 2, 1));
 208     t6   = _mm_permute_ps(tA, _MM_SHUFFLE(2, 1, 0, 2));
 209
 210     t1   = _mm_add_ps(t1, t4);
 211     t2   = _mm_add_ps(t2, t5);
 212     t3   = _mm_add_ps(t3, t6);
 213
 214     *x1  = _mm_permute_ps(t1, _MM_SHUFFLE(0, 0, 0, 0));
 215     *y1  = _mm_permute_ps(t1, _MM_SHUFFLE(1, 1, 1, 1));
 216     *z1  = _mm_permute_ps(t1, _MM_SHUFFLE(2, 2, 2, 2));
 217     *x2  = _mm_permute_ps(t1, _MM_SHUFFLE(3, 3, 3, 3));
 218     *y2  = _mm_permute_ps(t2, _MM_SHUFFLE(0, 0, 0, 0));
 219     *z2  = _mm_permute_ps(t2, _MM_SHUFFLE(1, 1, 1, 1));
 220     *x3  = _mm_permute_ps(t2, _MM_SHUFFLE(2, 2, 2, 2));
 221     *y3  = _mm_permute_ps(t2, _MM_SHUFFLE(3, 3, 3, 3));
 222     *z3  = _mm_permute_ps(t3, _MM_SHUFFLE(0, 0, 0, 0));
 223     *x4  = _mm_permute_ps(t3, _MM_SHUFFLE(1, 1, 1, 1));
 224     *y4  = _mm_permute_ps(t3, _MM_SHUFFLE(2, 2, 2, 2));
 225     *z4  = _mm_permute_ps(t3, _MM_SHUFFLE(3, 3, 3, 3));
 226 }
 227
 228
 229 static gmx_inline void
 230 gmx_mm_load_1rvec_4ptr_swizzle_ps(const float * gmx_restrict ptrA, const float * gmx_restrict ptrB,
 231                                   const float * gmx_restrict ptrC, const float * gmx_restrict ptrD,
 232                                   __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1)
 233 {
 234     __m128  t1, t2, t3, t4;
 235     __m128i mask = _mm_set_epi32(0, -1, -1, -1);
 236     t1             = gmx_mm_maskload_ps(ptrA, mask);
 237     t2             = gmx_mm_maskload_ps(ptrB, mask);
 238     t3             = gmx_mm_maskload_ps(ptrC, mask);
 239     t4             = gmx_mm_maskload_ps(ptrD, mask);
 240     _MM_TRANSPOSE4_PS(t1, t2, t3, t4);
 241     *x1           = t1;
 242     *y1           = t2;
 243     *z1           = t3;
 244 }
 245
 246
 247 static gmx_inline void
 248 gmx_mm_load_3rvec_4ptr_swizzle_ps(const float * gmx_restrict ptrA, const float * gmx_restrict ptrB,
 249                                   const float * gmx_restrict ptrC, const float * gmx_restrict ptrD,
 250                                   __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 251                                   __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 252                                   __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3)
 253 {
 254     __m128 t1, t2, t3, t4;
 255     t1            = _mm_loadu_ps(ptrA);
 256     t2            = _mm_loadu_ps(ptrB);
 257     t3            = _mm_loadu_ps(ptrC);
 258     t4            = _mm_loadu_ps(ptrD);
 259     _MM_TRANSPOSE4_PS(t1, t2, t3, t4);
 260     *x1           = t1;
 261     *y1           = t2;
 262     *z1           = t3;
 263     *x2           = t4;
 264     t1            = _mm_loadu_ps(ptrA+4);
 265     t2            = _mm_loadu_ps(ptrB+4);
 266     t3            = _mm_loadu_ps(ptrC+4);
 267     t4            = _mm_loadu_ps(ptrD+4);
 268     _MM_TRANSPOSE4_PS(t1, t2, t3, t4);
 269     *y2           = t1;
 270     *z2           = t2;
 271     *x3           = t3;
 272     *y3           = t4;
 273     t1            = _mm_load_ss(ptrA+8);
 274     t2            = _mm_load_ss(ptrB+8);
 275     t3            = _mm_load_ss(ptrC+8);
 276     t4            = _mm_load_ss(ptrD+8);
 277     t1            = _mm_unpacklo_ps(t1, t3);
 278     t3            = _mm_unpacklo_ps(t2, t4);
 279     *z3           = _mm_unpacklo_ps(t1, t3);
 280 }
 281
 282
 283 static gmx_inline void
 284 gmx_mm_load_4rvec_4ptr_swizzle_ps(const float * gmx_restrict ptrA, const float * gmx_restrict ptrB,
 285                                   const float * gmx_restrict ptrC, const float * gmx_restrict ptrD,
 286                                   __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 287                                   __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 288                                   __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3,
 289                                   __m128 * gmx_restrict x4, __m128 * gmx_restrict y4, __m128 * gmx_restrict z4)
 290 {
 291     __m128 t1, t2, t3, t4;
 292     t1            = _mm_loadu_ps(ptrA);
 293     t2            = _mm_loadu_ps(ptrB);
 294     t3            = _mm_loadu_ps(ptrC);
 295     t4            = _mm_loadu_ps(ptrD);
 296     _MM_TRANSPOSE4_PS(t1, t2, t3, t4);
 297     *x1           = t1;
 298     *y1           = t2;
 299     *z1           = t3;
 300     *x2           = t4;
 301     t1            = _mm_loadu_ps(ptrA+4);
 302     t2            = _mm_loadu_ps(ptrB+4);
 303     t3            = _mm_loadu_ps(ptrC+4);
 304     t4            = _mm_loadu_ps(ptrD+4);
 305     _MM_TRANSPOSE4_PS(t1, t2, t3, t4);
 306     *y2           = t1;
 307     *z2           = t2;
 308     *x3           = t3;
 309     *y3           = t4;
 310     t1            = _mm_loadu_ps(ptrA+8);
 311     t2            = _mm_loadu_ps(ptrB+8);
 312     t3            = _mm_loadu_ps(ptrC+8);
 313     t4            = _mm_loadu_ps(ptrD+8);
 314     _MM_TRANSPOSE4_PS(t1, t2, t3, t4);
 315     *z3           = t1;
 316     *x4           = t2;
 317     *y4           = t3;
 318     *z4           = t4;
 319 }
 320
 321
 322 static gmx_inline void
 323 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(float * gmx_restrict ptrA, float * gmx_restrict ptrB,
 324                                        float * gmx_restrict ptrC, float * gmx_restrict ptrD,
 325                                        __m128 x1, __m128 y1, __m128 z1)
 326 {
 327     __m128 t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12;
 328     t5          = _mm_unpacklo_ps(y1, z1);
 329     t6          = _mm_unpackhi_ps(y1, z1);
 330     t7          = _mm_shuffle_ps(x1, t5, _MM_SHUFFLE(1, 0, 0, 0));
 331     t8          = _mm_shuffle_ps(x1, t5, _MM_SHUFFLE(3, 2, 0, 1));
 332     t9          = _mm_shuffle_ps(x1, t6, _MM_SHUFFLE(1, 0, 0, 2));
 333     t10         = _mm_shuffle_ps(x1, t6, _MM_SHUFFLE(3, 2, 0, 3));
 334     t1          = _mm_load_ss(ptrA);
 335     t1          = _mm_loadh_pi(t1, (__m64 *)(ptrA+1));
 336     t1          = _mm_sub_ps(t1, t7);
 337     _mm_store_ss(ptrA, t1);
 338     _mm_storeh_pi((__m64 *)(ptrA+1), t1);
 339     t2          = _mm_load_ss(ptrB);
 340     t2          = _mm_loadh_pi(t2, (__m64 *)(ptrB+1));
 341     t2          = _mm_sub_ps(t2, t8);
 342     _mm_store_ss(ptrB, t2);
 343     _mm_storeh_pi((__m64 *)(ptrB+1), t2);
 344     t3          = _mm_load_ss(ptrC);
 345     t3          = _mm_loadh_pi(t3, (__m64 *)(ptrC+1));
 346     t3          = _mm_sub_ps(t3, t9);
 347     _mm_store_ss(ptrC, t3);
 348     _mm_storeh_pi((__m64 *)(ptrC+1), t3);
 349     t4          = _mm_load_ss(ptrD);
 350     t4          = _mm_loadh_pi(t4, (__m64 *)(ptrD+1));
 351     t4          = _mm_sub_ps(t4, t10);
 352     _mm_store_ss(ptrD, t4);
 353     _mm_storeh_pi((__m64 *)(ptrD+1), t4);
 354 }
 355
 356
 357 #if defined (_MSC_VER) && defined(_M_IX86)
 358 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 359 #define gmx_mm_decrement_3rvec_4ptr_swizzle_ps(ptrA, ptrB, ptrC, ptrD, \
 360                                                _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3) \
 361     { \
 362         __m128 _t1, _t2, _t3, _t4, _t5, _t6, _t7, _t8, _t9, _t10; \
 363         __m128 _t11, _t12, _t13, _t14, _t15, _t16, _t17, _t18, _t19; \
 364         __m128 _t20, _t21, _t22, _t23, _t24, _t25; \
 365         _t13         = _mm_unpackhi_ps(_x1, _y1); \
 366         _x1          = _mm_unpacklo_ps(_x1, _y1); \
 367         _t14         = _mm_unpackhi_ps(_z1, _x2); \
 368         _z1          = _mm_unpacklo_ps(_z1, _x2); \
 369         _t15         = _mm_unpackhi_ps(_y2, _z2); \
 370         _y2          = _mm_unpacklo_ps(_y2, _z2); \
 371         _t16         = _mm_unpackhi_ps(_x3, _y3); \
 372         _x3          = _mm_unpacklo_ps(_x3, _y3); \
 373         _t17         = _mm_permute_ps(_z3, _MM_SHUFFLE(0, 0, 0, 1)); \
 374         _t18         = _mm_movehl_ps(_z3, _z3); \
 375         _t19         = _mm_permute_ps(_t18, _MM_SHUFFLE(0, 0, 0, 1)); \
 376         _t20         = _mm_movelh_ps(_x1, _z1); \
 377         _t21         = _mm_movehl_ps(_z1, _x1); \
 378         _t22         = _mm_movelh_ps(_t13, _t14); \
 379         _t14         = _mm_movehl_ps(_t14, _t13); \
 380         _t23         = _mm_movelh_ps(_y2, _x3); \
 381         _t24         = _mm_movehl_ps(_x3, _y2); \
 382         _t25         = _mm_movelh_ps(_t15, _t16); \
 383         _t16         = _mm_movehl_ps(_t16, _t15); \
 384         _t1          = _mm_loadu_ps(ptrA); \
 385         _t2          = _mm_loadu_ps(ptrA+4); \
 386         _t3          = _mm_load_ss(ptrA+8); \
 387         _t1          = _mm_sub_ps(_t1, _t20); \
 388         _t2          = _mm_sub_ps(_t2, _t23); \
 389         _t3          = _mm_sub_ss(_t3, _z3); \
 390         _mm_storeu_ps(ptrA, _t1); \
 391         _mm_storeu_ps(ptrA+4, _t2); \
 392         _mm_store_ss(ptrA+8, _t3); \
 393         _t4          = _mm_loadu_ps(ptrB); \
 394         _t5          = _mm_loadu_ps(ptrB+4); \
 395         _t6          = _mm_load_ss(ptrB+8); \
 396         _t4          = _mm_sub_ps(_t4, _t21); \
 397         _t5          = _mm_sub_ps(_t5, _t24); \
 398         _t6          = _mm_sub_ss(_t6, _t17); \
 399         _mm_storeu_ps(ptrB, _t4); \
 400         _mm_storeu_ps(ptrB+4, _t5); \
 401         _mm_store_ss(ptrB+8, _t6); \
 402         _t7          = _mm_loadu_ps(ptrC); \
 403         _t8          = _mm_loadu_ps(ptrC+4); \
 404         _t9          = _mm_load_ss(ptrC+8); \
 405         _t7          = _mm_sub_ps(_t7, _t22); \
 406         _t8          = _mm_sub_ps(_t8, _t25); \
 407         _t9          = _mm_sub_ss(_t9, _t18); \
 408         _mm_storeu_ps(ptrC, _t7); \
 409         _mm_storeu_ps(ptrC+4, _t8); \
 410         _mm_store_ss(ptrC+8, _t9); \
 411         _t10         = _mm_loadu_ps(ptrD); \
 412         _t11         = _mm_loadu_ps(ptrD+4); \
 413         _t12         = _mm_load_ss(ptrD+8); \
 414         _t10         = _mm_sub_ps(_t10, _t14); \
 415         _t11         = _mm_sub_ps(_t11, _t16); \
 416         _t12         = _mm_sub_ss(_t12, _t19); \
 417         _mm_storeu_ps(ptrD, _t10); \
 418         _mm_storeu_ps(ptrD+4, _t11); \
 419         _mm_store_ss(ptrD+8, _t12); \
 420     }
 421 #else
 422 /* Real function for sane compilers */
 423 static gmx_inline void
 424 gmx_mm_decrement_3rvec_4ptr_swizzle_ps(float * gmx_restrict ptrA, float * gmx_restrict ptrB,
 425                                        float * gmx_restrict ptrC, float * gmx_restrict ptrD,
 426                                        __m128 x1, __m128 y1, __m128 z1,
 427                                        __m128 x2, __m128 y2, __m128 z2,
 428                                        __m128 x3, __m128 y3, __m128 z3)
 429 {
 430     __m128 t1, t2, t3, t4, t5, t6, t7, t8, t9, t10;
 431     __m128 t11, t12, t13, t14, t15, t16, t17, t18, t19;
 432     __m128 t20, t21, t22, t23, t24, t25;
 433     t13         = _mm_unpackhi_ps(x1, y1);
 434     x1          = _mm_unpacklo_ps(x1, y1);
 435     t14         = _mm_unpackhi_ps(z1, x2);
 436     z1          = _mm_unpacklo_ps(z1, x2);
 437     t15         = _mm_unpackhi_ps(y2, z2);
 438     y2          = _mm_unpacklo_ps(y2, z2);
 439     t16         = _mm_unpackhi_ps(x3, y3);
 440     x3          = _mm_unpacklo_ps(x3, y3);
 441     t17         = _mm_permute_ps(z3, _MM_SHUFFLE(0, 0, 0, 1));
 442     t18         = _mm_movehl_ps(z3, z3);
 443     t19         = _mm_permute_ps(t18, _MM_SHUFFLE(0, 0, 0, 1));
 444     t20         = _mm_movelh_ps(x1, z1);
 445     t21         = _mm_movehl_ps(z1, x1);
 446     t22         = _mm_movelh_ps(t13, t14);
 447     t14         = _mm_movehl_ps(t14, t13);
 448     t23         = _mm_movelh_ps(y2, x3);
 449     t24         = _mm_movehl_ps(x3, y2);
 450     t25         = _mm_movelh_ps(t15, t16);
 451     t16         = _mm_movehl_ps(t16, t15);
 452     t1          = _mm_loadu_ps(ptrA);
 453     t2          = _mm_loadu_ps(ptrA+4);
 454     t3          = _mm_load_ss(ptrA+8);
 455     t1          = _mm_sub_ps(t1, t20);
 456     t2          = _mm_sub_ps(t2, t23);
 457     t3          = _mm_sub_ss(t3, z3);
 458     _mm_storeu_ps(ptrA, t1);
 459     _mm_storeu_ps(ptrA+4, t2);
 460     _mm_store_ss(ptrA+8, t3);
 461     t4          = _mm_loadu_ps(ptrB);
 462     t5          = _mm_loadu_ps(ptrB+4);
 463     t6          = _mm_load_ss(ptrB+8);
 464     t4          = _mm_sub_ps(t4, t21);
 465     t5          = _mm_sub_ps(t5, t24);
 466     t6          = _mm_sub_ss(t6, t17);
 467     _mm_storeu_ps(ptrB, t4);
 468     _mm_storeu_ps(ptrB+4, t5);
 469     _mm_store_ss(ptrB+8, t6);
 470     t7          = _mm_loadu_ps(ptrC);
 471     t8          = _mm_loadu_ps(ptrC+4);
 472     t9          = _mm_load_ss(ptrC+8);
 473     t7          = _mm_sub_ps(t7, t22);
 474     t8          = _mm_sub_ps(t8, t25);
 475     t9          = _mm_sub_ss(t9, t18);
 476     _mm_storeu_ps(ptrC, t7);
 477     _mm_storeu_ps(ptrC+4, t8);
 478     _mm_store_ss(ptrC+8, t9);
 479     t10         = _mm_loadu_ps(ptrD);
 480     t11         = _mm_loadu_ps(ptrD+4);
 481     t12         = _mm_load_ss(ptrD+8);
 482     t10         = _mm_sub_ps(t10, t14);
 483     t11         = _mm_sub_ps(t11, t16);
 484     t12         = _mm_sub_ss(t12, t19);
 485     _mm_storeu_ps(ptrD, t10);
 486     _mm_storeu_ps(ptrD+4, t11);
 487     _mm_store_ss(ptrD+8, t12);
 488 }
 489 #endif
 490
 491 #if defined (_MSC_VER) && defined(_M_IX86)
 492 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 493 #define gmx_mm_decrement_4rvec_4ptr_swizzle_ps(ptrA, ptrB, ptrC, ptrD, \
 494                                                _x1, _y1, _z1, _x2, _y2, _z2, _x3, _y3, _z3, _x4, _y4, _z4) \
 495     { \
 496         __m128 _t1, _t2, _t3, _t4, _t5, _t6, _t7, _t8, _t9, _t10, _t11; \
 497         __m128 _t12, _t13, _t14, _t15, _t16, _t17, _t18, _t19, _t20, _t21, _t22; \
 498         __m128 _t23, _t24; \
 499         _t13         = _mm_unpackhi_ps(_x1, _y1); \
 500         _x1          = _mm_unpacklo_ps(_x1, _y1); \
 501         _t14         = _mm_unpackhi_ps(_z1, _x2); \
 502         _z1          = _mm_unpacklo_ps(_z1, _x2); \
 503         _t15         = _mm_unpackhi_ps(_y2, _z2); \
 504         _y2          = _mm_unpacklo_ps(_y2, _z2); \
 505         _t16         = _mm_unpackhi_ps(_x3, _y3); \
 506         _x3          = _mm_unpacklo_ps(_x3, _y3); \
 507         _t17         = _mm_unpackhi_ps(_z3, _x4); \
 508         _z3          = _mm_unpacklo_ps(_z3, _x4); \
 509         _t18         = _mm_unpackhi_ps(_y4, _z4); \
 510         _y4          = _mm_unpacklo_ps(_y4, _z4); \
 511         _t19         = _mm_movelh_ps(_x1, _z1); \
 512         _z1          = _mm_movehl_ps(_z1, _x1); \
 513         _t20         = _mm_movelh_ps(_t13, _t14); \
 514         _t14         = _mm_movehl_ps(_t14, _t13); \
 515         _t21         = _mm_movelh_ps(_y2, _x3); \
 516         _x3          = _mm_movehl_ps(_x3, _y2); \
 517         _t22         = _mm_movelh_ps(_t15, _t16); \
 518         _t16         = _mm_movehl_ps(_t16, _t15); \
 519         _t23         = _mm_movelh_ps(_z3, _y4); \
 520         _y4          = _mm_movehl_ps(_y4, _z3); \
 521         _t24         = _mm_movelh_ps(_t17, _t18); \
 522         _t18         = _mm_movehl_ps(_t18, _t17); \
 523         _t1          = _mm_loadu_ps(ptrA); \
 524         _t2          = _mm_loadu_ps(ptrA+4); \
 525         _t3          = _mm_loadu_ps(ptrA+8); \
 526         _t1          = _mm_sub_ps(_t1, _t19); \
 527         _t2          = _mm_sub_ps(_t2, _t21); \
 528         _t3          = _mm_sub_ps(_t3, _t23); \
 529         _mm_storeu_ps(ptrA, _t1); \
 530         _mm_storeu_ps(ptrA+4, _t2); \
 531         _mm_storeu_ps(ptrA+8, _t3); \
 532         _t4          = _mm_loadu_ps(ptrB); \
 533         _t5          = _mm_loadu_ps(ptrB+4); \
 534         _t6          = _mm_loadu_ps(ptrB+8); \
 535         _t4          = _mm_sub_ps(_t4, _z1); \
 536         _t5          = _mm_sub_ps(_t5, _x3); \
 537         _t6          = _mm_sub_ps(_t6, _y4); \
 538         _mm_storeu_ps(ptrB, _t4); \
 539         _mm_storeu_ps(ptrB+4, _t5); \
 540         _mm_storeu_ps(ptrB+8, _t6); \
 541         _t7          = _mm_loadu_ps(ptrC); \
 542         _t8          = _mm_loadu_ps(ptrC+4); \
 543         _t9          = _mm_loadu_ps(ptrC+8); \
 544         _t7          = _mm_sub_ps(_t7, _t20); \
 545         _t8          = _mm_sub_ps(_t8, _t22); \
 546         _t9          = _mm_sub_ps(_t9, _t24); \
 547         _mm_storeu_ps(ptrC, _t7); \
 548         _mm_storeu_ps(ptrC+4, _t8); \
 549         _mm_storeu_ps(ptrC+8, _t9); \
 550         _t10         = _mm_loadu_ps(ptrD); \
 551         _t11         = _mm_loadu_ps(ptrD+4); \
 552         _t12         = _mm_loadu_ps(ptrD+8); \
 553         _t10         = _mm_sub_ps(_t10, _t14); \
 554         _t11         = _mm_sub_ps(_t11, _t16); \
 555         _t12         = _mm_sub_ps(_t12, _t18); \
 556         _mm_storeu_ps(ptrD, _t10); \
 557         _mm_storeu_ps(ptrD+4, _t11); \
 558         _mm_storeu_ps(ptrD+8, _t12); \
 559     }
 560 #else
 561 /* Real function for sane compilers */
 562 static gmx_inline void
 563 gmx_mm_decrement_4rvec_4ptr_swizzle_ps(float * gmx_restrict ptrA, float * gmx_restrict ptrB,
 564                                        float * gmx_restrict ptrC, float * gmx_restrict ptrD,
 565                                        __m128 x1, __m128 y1, __m128 z1,
 566                                        __m128 x2, __m128 y2, __m128 z2,
 567                                        __m128 x3, __m128 y3, __m128 z3,
 568                                        __m128 x4, __m128 y4, __m128 z4)
 569 {
 570     __m128 t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11;
 571     __m128 t12, t13, t14, t15, t16, t17, t18, t19, t20, t21, t22;
 572     __m128 t23, t24;
 573     t13         = _mm_unpackhi_ps(x1, y1);
 574     x1          = _mm_unpacklo_ps(x1, y1);
 575     t14         = _mm_unpackhi_ps(z1, x2);
 576     z1          = _mm_unpacklo_ps(z1, x2);
 577     t15         = _mm_unpackhi_ps(y2, z2);
 578     y2          = _mm_unpacklo_ps(y2, z2);
 579     t16         = _mm_unpackhi_ps(x3, y3);
 580     x3          = _mm_unpacklo_ps(x3, y3);
 581     t17         = _mm_unpackhi_ps(z3, x4);
 582     z3          = _mm_unpacklo_ps(z3, x4);
 583     t18         = _mm_unpackhi_ps(y4, z4);
 584     y4          = _mm_unpacklo_ps(y4, z4);
 585     t19         = _mm_movelh_ps(x1, z1);
 586     z1          = _mm_movehl_ps(z1, x1);
 587     t20         = _mm_movelh_ps(t13, t14);
 588     t14         = _mm_movehl_ps(t14, t13);
 589     t21         = _mm_movelh_ps(y2, x3);
 590     x3          = _mm_movehl_ps(x3, y2);
 591     t22         = _mm_movelh_ps(t15, t16);
 592     t16         = _mm_movehl_ps(t16, t15);
 593     t23         = _mm_movelh_ps(z3, y4);
 594     y4          = _mm_movehl_ps(y4, z3);
 595     t24         = _mm_movelh_ps(t17, t18);
 596     t18         = _mm_movehl_ps(t18, t17);
 597     t1          = _mm_loadu_ps(ptrA);
 598     t2          = _mm_loadu_ps(ptrA+4);
 599     t3          = _mm_loadu_ps(ptrA+8);
 600     t1          = _mm_sub_ps(t1, t19);
 601     t2          = _mm_sub_ps(t2, t21);
 602     t3          = _mm_sub_ps(t3, t23);
 603     _mm_storeu_ps(ptrA, t1);
 604     _mm_storeu_ps(ptrA+4, t2);
 605     _mm_storeu_ps(ptrA+8, t3);
 606     t4          = _mm_loadu_ps(ptrB);
 607     t5          = _mm_loadu_ps(ptrB+4);
 608     t6          = _mm_loadu_ps(ptrB+8);
 609     t4          = _mm_sub_ps(t4, z1);
 610     t5          = _mm_sub_ps(t5, x3);
 611     t6          = _mm_sub_ps(t6, y4);
 612     _mm_storeu_ps(ptrB, t4);
 613     _mm_storeu_ps(ptrB+4, t5);
 614     _mm_storeu_ps(ptrB+8, t6);
 615     t7          = _mm_loadu_ps(ptrC);
 616     t8          = _mm_loadu_ps(ptrC+4);
 617     t9          = _mm_loadu_ps(ptrC+8);
 618     t7          = _mm_sub_ps(t7, t20);
 619     t8          = _mm_sub_ps(t8, t22);
 620     t9          = _mm_sub_ps(t9, t24);
 621     _mm_storeu_ps(ptrC, t7);
 622     _mm_storeu_ps(ptrC+4, t8);
 623     _mm_storeu_ps(ptrC+8, t9);
 624     t10         = _mm_loadu_ps(ptrD);
 625     t11         = _mm_loadu_ps(ptrD+4);
 626     t12         = _mm_loadu_ps(ptrD+8);
 627     t10         = _mm_sub_ps(t10, t14);
 628     t11         = _mm_sub_ps(t11, t16);
 629     t12         = _mm_sub_ps(t12, t18);
 630     _mm_storeu_ps(ptrD, t10);
 631     _mm_storeu_ps(ptrD+4, t11);
 632     _mm_storeu_ps(ptrD+8, t12);
 633 }
 634 #endif
 635
 636 static gmx_inline void
 637 gmx_mm_update_iforce_1atom_swizzle_ps(__m128 fix1, __m128 fiy1, __m128 fiz1,
 638                                       float * gmx_restrict fptr,
 639                                       float * gmx_restrict fshiftptr)
 640 {
 641     __m128 t2, t3;
 642
 643     fix1 = _mm_hadd_ps(fix1, fix1);
 644     fiy1 = _mm_hadd_ps(fiy1, fiz1);
 645
 646     fix1 = _mm_hadd_ps(fix1, fiy1); /* fiz1 fiy1 fix1 fix1 */
 647
 648     t2 = _mm_load_ss(fptr);
 649     t2 = _mm_loadh_pi(t2, (__m64 *)(fptr+1));
 650     t3 = _mm_load_ss(fshiftptr);
 651     t3 = _mm_loadh_pi(t3, (__m64 *)(fshiftptr+1));
 652
 653     t2 = _mm_add_ps(t2, fix1);
 654     t3 = _mm_add_ps(t3, fix1);
 655
 656     _mm_store_ss(fptr, t2);
 657     _mm_storeh_pi((__m64 *)(fptr+1), t2);
 658     _mm_store_ss(fshiftptr, t3);
 659     _mm_storeh_pi((__m64 *)(fshiftptr+1), t3);
 660 }
 661
 662 #if defined (_MSC_VER) && defined(_M_IX86)
 663 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 664 #define gmx_mm_update_iforce_3atom_swizzle_ps(fix1, fiy1, fiz1, fix2, fiy2, fiz2, fix3, fiy3, fiz3, \
 665                                               fptr, fshiftptr) \
 666     { \
 667         __m128 _t1, _t2, _t3, _t4; \
 668 \
 669         fix1 = _mm_hadd_ps(fix1, fiy1); \
 670         fiz1 = _mm_hadd_ps(fiz1, fix2); \
 671         fiy2 = _mm_hadd_ps(fiy2, fiz2); \
 672         fix3 = _mm_hadd_ps(fix3, fiy3); \
 673         fiz3 = _mm_hadd_ps(fiz3, fiz3); \
 674         fix1 = _mm_hadd_ps(fix1, fiz1); \
 675         fiy2 = _mm_hadd_ps(fiy2, fix3); \
 676         fiz3 = _mm_hadd_ps(fiz3, fiz3); \
 677         _mm_storeu_ps(fptr,  _mm_add_ps(fix1, _mm_loadu_ps(fptr)  )); \
 678         _mm_storeu_ps(fptr+4, _mm_add_ps(fiy2, _mm_loadu_ps(fptr+4))); \
 679         _mm_store_ss (fptr+8, _mm_add_ss(fiz3, _mm_load_ss(fptr+8) )); \
 680         _t4 = _mm_load_ss(fshiftptr+2); \
 681         _t4 = _mm_loadh_pi(_t4, (__m64 *)(fshiftptr)); \
 682         _t1 = _mm_shuffle_ps(fiz3, fix1, _MM_SHUFFLE(1, 0, 0, 0)); \
 683         _t2 = _mm_shuffle_ps(fix1, fiy2, _MM_SHUFFLE(3, 2, 2, 2)); \
 684         _t3 = _mm_shuffle_ps(fiy2, fix1, _MM_SHUFFLE(3, 3, 0, 1)); \
 685         _t3 = _mm_permute_ps(_t3, _MM_SHUFFLE(1, 2, 0, 0)); \
 686         _t1 = _mm_add_ps(_t1, _t2); \
 687         _t3 = _mm_add_ps(_t3, _t4); \
 688         _t1 = _mm_add_ps(_t1, _t3); \
 689         _mm_store_ss(fshiftptr+2, _t1); \
 690         _mm_storeh_pi((__m64 *)(fshiftptr), _t1); \
 691     }
 692 #else
 693 /* Real function for sane compilers */
 694 static gmx_inline void
 695 gmx_mm_update_iforce_3atom_swizzle_ps(__m128 fix1, __m128 fiy1, __m128 fiz1,
 696                                       __m128 fix2, __m128 fiy2, __m128 fiz2,
 697                                       __m128 fix3, __m128 fiy3, __m128 fiz3,
 698                                       float * gmx_restrict fptr,
 699                                       float * gmx_restrict fshiftptr)
 700 {
 701     __m128 t1, t2, t3, t4;
 702
 703     fix1 = _mm_hadd_ps(fix1, fiy1);
 704     fiz1 = _mm_hadd_ps(fiz1, fix2);
 705     fiy2 = _mm_hadd_ps(fiy2, fiz2);
 706     fix3 = _mm_hadd_ps(fix3, fiy3);
 707     fiz3 = _mm_hadd_ps(fiz3, fiz3);
 708
 709     fix1 = _mm_hadd_ps(fix1, fiz1); /* fix2 fiz1 fiy1 fix1 */
 710     fiy2 = _mm_hadd_ps(fiy2, fix3); /* fiy3 fix3 fiz2 fiy2 */
 711     fiz3 = _mm_hadd_ps(fiz3, fiz3); /*  -    -    -   fiz3 */
 712
 713     _mm_storeu_ps(fptr,  _mm_add_ps(fix1, _mm_loadu_ps(fptr)  ));
 714     _mm_storeu_ps(fptr+4, _mm_add_ps(fiy2, _mm_loadu_ps(fptr+4)));
 715     _mm_store_ss (fptr+8, _mm_add_ss(fiz3, _mm_load_ss(fptr+8) ));
 716
 717     t4 = _mm_load_ss(fshiftptr+2);
 718     t4 = _mm_loadh_pi(t4, (__m64 *)(fshiftptr));
 719
 720     t1 = _mm_shuffle_ps(fiz3, fix1, _MM_SHUFFLE(1, 0, 0, 0)); /* fiy1 fix1  -   fiz3 */
 721     t2 = _mm_shuffle_ps(fix1, fiy2, _MM_SHUFFLE(3, 2, 2, 2)); /* fiy3 fix3  -   fiz1 */
 722     t3 = _mm_shuffle_ps(fiy2, fix1, _MM_SHUFFLE(3, 3, 0, 1)); /* fix2 fix2 fiy2 fiz2 */
 723     t3 = _mm_permute_ps(t3, _MM_SHUFFLE(1, 2, 0, 0));         /* fiy2 fix2  -   fiz2 */
 724
 725     t1 = _mm_add_ps(t1, t2);
 726     t3 = _mm_add_ps(t3, t4);
 727     t1 = _mm_add_ps(t1, t3); /* y x - z */
 728
 729     _mm_store_ss(fshiftptr+2, t1);
 730     _mm_storeh_pi((__m64 *)(fshiftptr), t1);
 731 }
 732 #endif
 733
 734 #if defined (_MSC_VER) && defined(_M_IX86)
 735 /* Macro work-around since 32-bit MSVC cannot handle >3 xmm/ymm parameters */
 736 #define gmx_mm_update_iforce_4atom_swizzle_ps(fix1, fiy1, fiz1, fix2, fiy2, fiz2, fix3, fiy3, fiz3, fix4, fiy4, fiz4, \
 737                                               fptr, fshiftptr) \
 738     { \
 739         __m128 _t1, _t2, _t3, _t4, _t5; \
 740 \
 741         fix1 = _mm_hadd_ps(fix1, fiy1); \
 742         fiz1 = _mm_hadd_ps(fiz1, fix2); \
 743         fiy2 = _mm_hadd_ps(fiy2, fiz2); \
 744         fix3 = _mm_hadd_ps(fix3, fiy3); \
 745         fiz3 = _mm_hadd_ps(fiz3, fix4); \
 746         fiy4 = _mm_hadd_ps(fiy4, fiz4); \
 747         fix1 = _mm_hadd_ps(fix1, fiz1); \
 748         fiy2 = _mm_hadd_ps(fiy2, fix3); \
 749         fiz3 = _mm_hadd_ps(fiz3, fiy4); \
 750         _mm_storeu_ps(fptr,  _mm_add_ps(fix1, _mm_loadu_ps(fptr)  )); \
 751         _mm_storeu_ps(fptr+4, _mm_add_ps(fiy2, _mm_loadu_ps(fptr+4))); \
 752         _mm_storeu_ps(fptr+8, _mm_add_ps(fiz3, _mm_loadu_ps(fptr+8))); \
 753         _t5 = _mm_load_ss(fshiftptr+2); \
 754         _t5 = _mm_loadh_pi(_t5, (__m64 *)(fshiftptr)); \
 755         _t1 = _mm_permute_ps(fix1, _MM_SHUFFLE(1, 0, 2, 2)); \
 756         _t2 = _mm_permute_ps(fiy2, _MM_SHUFFLE(3, 2, 1, 1)); \
 757         _t3 = _mm_permute_ps(fiz3, _MM_SHUFFLE(2, 1, 0, 0)); \
 758         _t4 = _mm_shuffle_ps(fix1, fiy2, _MM_SHUFFLE(0, 0, 3, 3)); \
 759         _t4 = _mm_shuffle_ps(fiz3, _t4, _MM_SHUFFLE(2, 0, 3, 3)); \
 760         _t1 = _mm_add_ps(_t1, _t2); \
 761         _t3 = _mm_add_ps(_t3, _t4); \
 762         _t1 = _mm_add_ps(_t1, _t3); \
 763         _t5 = _mm_add_ps(_t5, _t1); \
 764         _mm_store_ss(fshiftptr+2, _t5); \
 765         _mm_storeh_pi((__m64 *)(fshiftptr), _t5); \
 766     }
 767 #else
 768 /* Real function for sane compilers */
 769 static gmx_inline void
 770 gmx_mm_update_iforce_4atom_swizzle_ps(__m128 fix1, __m128 fiy1, __m128 fiz1,
 771                                       __m128 fix2, __m128 fiy2, __m128 fiz2,
 772                                       __m128 fix3, __m128 fiy3, __m128 fiz3,
 773                                       __m128 fix4, __m128 fiy4, __m128 fiz4,
 774                                       float * gmx_restrict fptr,
 775                                       float * gmx_restrict fshiftptr)
 776 {
 777     __m128 t1, t2, t3, t4, t5;
 778
 779     fix1 = _mm_hadd_ps(fix1, fiy1);
 780     fiz1 = _mm_hadd_ps(fiz1, fix2);
 781     fiy2 = _mm_hadd_ps(fiy2, fiz2);
 782     fix3 = _mm_hadd_ps(fix3, fiy3);
 783     fiz3 = _mm_hadd_ps(fiz3, fix4);
 784     fiy4 = _mm_hadd_ps(fiy4, fiz4);
 785
 786     fix1 = _mm_hadd_ps(fix1, fiz1); /* fix2 fiz1 fiy1 fix1 */
 787     fiy2 = _mm_hadd_ps(fiy2, fix3); /* fiy3 fix3 fiz2 fiy2 */
 788     fiz3 = _mm_hadd_ps(fiz3, fiy4); /* fiz4 fiy4 fix4 fiz3 */
 789
 790     _mm_storeu_ps(fptr,  _mm_add_ps(fix1, _mm_loadu_ps(fptr)  ));
 791     _mm_storeu_ps(fptr+4, _mm_add_ps(fiy2, _mm_loadu_ps(fptr+4)));
 792     _mm_storeu_ps(fptr+8, _mm_add_ps(fiz3, _mm_loadu_ps(fptr+8)));
 793
 794     t5 = _mm_load_ss(fshiftptr+2);
 795     t5 = _mm_loadh_pi(t5, (__m64 *)(fshiftptr));
 796
 797     t1 = _mm_permute_ps(fix1, _MM_SHUFFLE(1, 0, 2, 2));
 798     t2 = _mm_permute_ps(fiy2, _MM_SHUFFLE(3, 2, 1, 1));
 799     t3 = _mm_permute_ps(fiz3, _MM_SHUFFLE(2, 1, 0, 0));
 800     t4 = _mm_shuffle_ps(fix1, fiy2, _MM_SHUFFLE(0, 0, 3, 3));
 801     t4 = _mm_shuffle_ps(fiz3, t4, _MM_SHUFFLE(2, 0, 3, 3));
 802
 803     t1 = _mm_add_ps(t1, t2);
 804     t3 = _mm_add_ps(t3, t4);
 805     t1 = _mm_add_ps(t1, t3);
 806     t5 = _mm_add_ps(t5, t1);
 807
 808     _mm_store_ss(fshiftptr+2, t5);
 809     _mm_storeh_pi((__m64 *)(fshiftptr), t5);
 810 }
 811 #endif
 812
 813
 814 static gmx_inline void
 815 gmx_mm_update_1pot_ps(__m128 pot1, float * gmx_restrict ptrA)
 816 {
 817     pot1 = _mm_hadd_ps(pot1, pot1);
 818     pot1 = _mm_hadd_ps(pot1, pot1);
 819     _mm_store_ss(ptrA, _mm_add_ss(pot1, _mm_load_ss(ptrA)));
 820 }
 821
 822 static gmx_inline void
 823 gmx_mm_update_2pot_ps(__m128 pot1, float * gmx_restrict ptrA,
 824                       __m128 pot2, float * gmx_restrict ptrB)
 825 {
 826     pot1 = _mm_hadd_ps(pot1, pot2);
 827     pot1 = _mm_hadd_ps(pot1, pot1);
 828     pot2 = _mm_permute_ps(pot1, _MM_SHUFFLE(0, 0, 0, 1));
 829     _mm_store_ss(ptrA, _mm_add_ss(pot1, _mm_load_ss(ptrA)));
 830     _mm_store_ss(ptrB, _mm_add_ss(pot2, _mm_load_ss(ptrB)));
 831 }
 832
 833
 834 #endif /* _kernelutil_x86_avx_128_fma_single_h_ */