src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/kernelutil_x86_sse4_1_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_sse4_1_single_h_
  30 #define _kernelutil_x86_sse4_1_single_h_
  31
  32 #include <math.h>
  33
  34 #include "gmx_x86_sse4_1.h"
  35
  36 #undef gmx_restrict
  37 #define gmx_restrict
  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 gmx_inline 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 gmx_inline __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 gmx_inline 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 gmx_inline 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 gmx_inline 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
 123 static gmx_inline void
 124 gmx_mm_load_shift_and_1rvec_broadcast_ps(const float * gmx_restrict xyz_shift,
 125                                          const float * gmx_restrict xyz,
 126                                          __m128 * gmx_restrict x1,
 127                                          __m128 * gmx_restrict y1,
 128                                          __m128 * gmx_restrict z1)
 129 {
 130     __m128 t1,t2,t3,t4;
 131
 132     t1   = _mm_loadl_pi(_mm_setzero_ps(),(__m64 *)xyz_shift);
 133     t2   = _mm_loadl_pi(_mm_setzero_ps(),(__m64 *)xyz);
 134     t3   = _mm_load_ss(xyz_shift+2);
 135     t4   = _mm_load_ss(xyz+2);
 136     t1   = _mm_add_ps(t1,t2);
 137     t3   = _mm_add_ss(t3,t4);
 138
 139     *x1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(0,0,0,0));
 140     *y1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(1,1,1,1));
 141     *z1  = _mm_shuffle_ps(t3,t3,_MM_SHUFFLE(0,0,0,0));
 142 }
 143
 144
 145 static gmx_inline void
 146 gmx_mm_load_shift_and_3rvec_broadcast_ps(const float * gmx_restrict xyz_shift,
 147                                          const float * gmx_restrict xyz,
 148                                          __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 149                                          __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 150                                          __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3)
 151 {
 152     __m128 tA,tB;
 153     __m128 t1,t2,t3,t4,t5,t6;
 154
 155     tA   = _mm_loadl_pi(_mm_setzero_ps(),(__m64 *)xyz_shift);
 156     tB   = _mm_load_ss(xyz_shift+2);
 157
 158     t1   = _mm_loadu_ps(xyz);
 159     t2   = _mm_loadu_ps(xyz+4);
 160     t3   = _mm_load_ss(xyz+8);
 161
 162     tA   = _mm_movelh_ps(tA,tB);
 163     t4   = _mm_shuffle_ps(tA,tA,_MM_SHUFFLE(0,2,1,0));
 164     t5   = _mm_shuffle_ps(tA,tA,_MM_SHUFFLE(1,0,2,1));
 165     t6   = _mm_shuffle_ps(tA,tA,_MM_SHUFFLE(2,1,0,2));
 166
 167     t1   = _mm_add_ps(t1,t4);
 168     t2   = _mm_add_ps(t2,t5);
 169     t3   = _mm_add_ss(t3,t6);
 170
 171     *x1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(0,0,0,0));
 172     *y1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(1,1,1,1));
 173     *z1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(2,2,2,2));
 174     *x2  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(3,3,3,3));
 175     *y2  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(0,0,0,0));
 176     *z2  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(1,1,1,1));
 177     *x3  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(2,2,2,2));
 178     *y3  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(3,3,3,3));
 179     *z3  = _mm_shuffle_ps(t3,t3,_MM_SHUFFLE(0,0,0,0));
 180 }
 181
 182
 183 static gmx_inline void
 184 gmx_mm_load_shift_and_4rvec_broadcast_ps(const float * gmx_restrict xyz_shift,
 185                                          const float * gmx_restrict xyz,
 186                                          __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 187                                          __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 188                                          __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3,
 189                                          __m128 * gmx_restrict x4, __m128 * gmx_restrict y4, __m128 * gmx_restrict z4)
 190 {
 191     __m128 tA,tB;
 192     __m128 t1,t2,t3,t4,t5,t6;
 193
 194     tA   = _mm_castpd_ps(_mm_load_sd((const double *)xyz_shift));
 195     tB   = _mm_load_ss(xyz_shift+2);
 196
 197     t1   = _mm_loadu_ps(xyz);
 198     t2   = _mm_loadu_ps(xyz+4);
 199     t3   = _mm_loadu_ps(xyz+8);
 200
 201     tA   = _mm_movelh_ps(tA,tB);
 202     t4   = _mm_shuffle_ps(tA,tA,_MM_SHUFFLE(0,2,1,0));
 203     t5   = _mm_shuffle_ps(tA,tA,_MM_SHUFFLE(1,0,2,1));
 204     t6   = _mm_shuffle_ps(tA,tA,_MM_SHUFFLE(2,1,0,2));
 205
 206     t1   = _mm_add_ps(t1,t4);
 207     t2   = _mm_add_ps(t2,t5);
 208     t3   = _mm_add_ps(t3,t6);
 209
 210     *x1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(0,0,0,0));
 211     *y1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(1,1,1,1));
 212     *z1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(2,2,2,2));
 213     *x2  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(3,3,3,3));
 214     *y2  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(0,0,0,0));
 215     *z2  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(1,1,1,1));
 216     *x3  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(2,2,2,2));
 217     *y3  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(3,3,3,3));
 218     *z3  = _mm_shuffle_ps(t3,t3,_MM_SHUFFLE(0,0,0,0));
 219     *x4  = _mm_shuffle_ps(t3,t3,_MM_SHUFFLE(1,1,1,1));
 220     *y4  = _mm_shuffle_ps(t3,t3,_MM_SHUFFLE(2,2,2,2));
 221     *z4  = _mm_shuffle_ps(t3,t3,_MM_SHUFFLE(3,3,3,3));
 222 }
 223
 224
 225 static gmx_inline void
 226 gmx_mm_load_1rvec_4ptr_swizzle_ps(const float * gmx_restrict ptrA,
 227                                   const float * gmx_restrict ptrB,
 228                                   const float * gmx_restrict ptrC,
 229                                   const float * gmx_restrict ptrD,
 230                                   __m128 *      gmx_restrict x1,
 231                                   __m128 *      gmx_restrict y1,
 232                                   __m128 *      gmx_restrict z1)
 233 {
 234     __m128 t1,t2,t3,t4,t5,t6,t7,t8;
 235     t1   = _mm_castpd_ps(_mm_load_sd((const double *)ptrA));
 236     t2   = _mm_castpd_ps(_mm_load_sd((const double *)ptrB));
 237     t3   = _mm_castpd_ps(_mm_load_sd((const double *)ptrC));
 238     t4   = _mm_castpd_ps(_mm_load_sd((const double *)ptrD));
 239     t5 = _mm_load_ss(ptrA+2);
 240     t6 = _mm_load_ss(ptrB+2);
 241     t7 = _mm_load_ss(ptrC+2);
 242     t8 = _mm_load_ss(ptrD+2);
 243     t1 = _mm_unpacklo_ps(t1,t2);
 244     t3 = _mm_unpacklo_ps(t3,t4);
 245     *x1 = _mm_movelh_ps(t1,t3);
 246     *y1 = _mm_movehl_ps(t3,t1);
 247     t5  = _mm_unpacklo_ps(t5,t6);
 248     t7  = _mm_unpacklo_ps(t7,t8);
 249     *z1 = _mm_movelh_ps(t5,t7);
 250 }
 251
 252
 253 static gmx_inline void
 254 gmx_mm_load_3rvec_4ptr_swizzle_ps(const float * gmx_restrict ptrA,
 255                                   const float * gmx_restrict ptrB,
 256                                   const float * gmx_restrict ptrC,
 257                                   const float * gmx_restrict ptrD,
 258                                   __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 259                                   __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 260                                   __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3)
 261 {
 262     __m128 t1,t2,t3,t4;
 263     t1            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)ptrA ) );
 264     t2            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)ptrB ) );
 265     t3            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)ptrC ) );
 266     t4            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)ptrD ) );
 267     _MM_TRANSPOSE4_PS(t1,t2,t3,t4);
 268     *x1           = t1;
 269     *y1           = t2;
 270     *z1           = t3;
 271     *x2           = t4;
 272     t1            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrA+4) ) );
 273     t2            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrB+4) ) );
 274     t3            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrC+4) ) );
 275     t4            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrD+4) ) );
 276     _MM_TRANSPOSE4_PS(t1,t2,t3,t4);
 277     *y2           = t1;
 278     *z2           = t2;
 279     *x3           = t3;
 280     *y3           = t4;
 281     t1            = _mm_load_ss(ptrA+8);
 282     t2            = _mm_load_ss(ptrB+8);
 283     t3            = _mm_load_ss(ptrC+8);
 284     t4            = _mm_load_ss(ptrD+8);
 285     t1            = _mm_unpacklo_ps(t1,t3);
 286     t3            = _mm_unpacklo_ps(t2,t4);
 287     *z3           = _mm_unpacklo_ps(t1,t3);
 288 }
 289
 290
 291 static gmx_inline void
 292 gmx_mm_load_4rvec_4ptr_swizzle_ps(const float * gmx_restrict ptrA,
 293                                   const float * gmx_restrict ptrB,
 294                                   const float * gmx_restrict ptrC,
 295                                   const float * gmx_restrict ptrD,
 296                                   __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 297                                   __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 298                                   __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3,
 299                                   __m128 * gmx_restrict x4, __m128 * gmx_restrict y4, __m128 * gmx_restrict z4)
 300 {
 301     __m128 t1,t2,t3,t4;
 302     t1            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrA) ) );
 303     t2            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrB) ) );
 304     t3            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrC) ) );
 305     t4            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrD) ) );
 306     _MM_TRANSPOSE4_PS(t1,t2,t3,t4);
 307     *x1           = t1;
 308     *y1           = t2;
 309     *z1           = t3;
 310     *x2           = t4;
 311     t1            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrA+4) ) );
 312     t2            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrB+4) ) );
 313     t3            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrC+4) ) );
 314     t4            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrD+4) ) );
 315     _MM_TRANSPOSE4_PS(t1,t2,t3,t4);
 316     *y2           = t1;
 317     *z2           = t2;
 318     *x3           = t3;
 319     *y3           = t4;
 320     t1            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrA+8) ) );
 321     t2            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrB+8) ) );
 322     t3            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrC+8) ) );
 323     t4            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrD+8) ) );
 324     _MM_TRANSPOSE4_PS(t1,t2,t3,t4);
 325     *z3           = t1;
 326     *x4           = t2;
 327     *y4           = t3;
 328     *z4           = t4;
 329 }
 330
 331
 332
 333 static gmx_inline void
 334 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(float * ptrA,
 335                                        float * ptrB,
 336                                        float * ptrC,
 337                                        float * ptrD,
 338                                        __m128 x1, __m128 y1, __m128 z1)
 339 {
 340     __m128 t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,t11,t12;
 341     t5          = _mm_unpacklo_ps(y1,z1);
 342     t6          = _mm_unpackhi_ps(y1,z1);
 343     t7          = _mm_shuffle_ps(x1,t5,_MM_SHUFFLE(1,0,0,0));
 344     t8          = _mm_shuffle_ps(x1,t5,_MM_SHUFFLE(3,2,0,1));
 345     t9          = _mm_shuffle_ps(x1,t6,_MM_SHUFFLE(1,0,0,2));
 346     t10         = _mm_shuffle_ps(x1,t6,_MM_SHUFFLE(3,2,0,3));
 347     t1          = _mm_load_ss(ptrA);
 348     t1          = _mm_loadh_pi(t1,(__m64 *)(ptrA+1));
 349     t1          = _mm_sub_ps(t1,t7);
 350     _mm_store_ss(ptrA,t1);
 351     _mm_storeh_pi((__m64 *)(ptrA+1),t1);
 352     t2          = _mm_load_ss(ptrB);
 353     t2          = _mm_loadh_pi(t2,(__m64 *)(ptrB+1));
 354     t2          = _mm_sub_ps(t2,t8);
 355     _mm_store_ss(ptrB,t2);
 356     _mm_storeh_pi((__m64 *)(ptrB+1),t2);
 357     t3          = _mm_load_ss(ptrC);
 358     t3          = _mm_loadh_pi(t3,(__m64 *)(ptrC+1));
 359     t3          = _mm_sub_ps(t3,t9);
 360     _mm_store_ss(ptrC,t3);
 361     _mm_storeh_pi((__m64 *)(ptrC+1),t3);
 362     t4          = _mm_load_ss(ptrD);
 363     t4          = _mm_loadh_pi(t4,(__m64 *)(ptrD+1));
 364     t4          = _mm_sub_ps(t4,t10);
 365     _mm_store_ss(ptrD,t4);
 366     _mm_storeh_pi((__m64 *)(ptrD+1),t4);
 367 }
 368
 369
 370
 371 static gmx_inline void
 372 gmx_mm_decrement_3rvec_4ptr_swizzle_ps(float * gmx_restrict ptrA, float * gmx_restrict ptrB,
 373                                        float * gmx_restrict ptrC, float * gmx_restrict ptrD,
 374                                        __m128 x1, __m128 y1, __m128 z1,
 375                                        __m128 x2, __m128 y2, __m128 z2,
 376                                        __m128 x3, __m128 y3, __m128 z3)
 377 {
 378     __m128 t1,t2,t3,t4,t5,t6,t7,t8,t9,t10;
 379     __m128 t11,t12,t13,t14,t15,t16,t17,t18,t19;
 380     __m128 t20,t21,t22,t23,t24,t25;
 381
 382     t13         = _mm_unpackhi_ps(x1,y1);
 383     x1          = _mm_unpacklo_ps(x1,y1);
 384     t14         = _mm_unpackhi_ps(z1,x2);
 385     z1          = _mm_unpacklo_ps(z1,x2);
 386     t15         = _mm_unpackhi_ps(y2,z2);
 387     y2          = _mm_unpacklo_ps(y2,z2);
 388     t16         = _mm_unpackhi_ps(x3,y3);
 389     x3          = _mm_unpacklo_ps(x3,y3);
 390     t17         = _mm_shuffle_ps(z3,z3,_MM_SHUFFLE(0,0,0,1));
 391     t18         = _mm_movehl_ps(z3,z3);
 392     t19         = _mm_shuffle_ps(t18,t18,_MM_SHUFFLE(0,0,0,1));
 393     t20         = _mm_movelh_ps(x1,z1);
 394     t21         = _mm_movehl_ps(z1,x1);
 395     t22         = _mm_movelh_ps(t13,t14);
 396     t14         = _mm_movehl_ps(t14,t13);
 397     t23         = _mm_movelh_ps(y2,x3);
 398     t24         = _mm_movehl_ps(x3,y2);
 399     t25         = _mm_movelh_ps(t15,t16);
 400     t16         = _mm_movehl_ps(t16,t15);
 401     t1          = _mm_loadu_ps(ptrA);
 402     t2          = _mm_loadu_ps(ptrA+4);
 403     t3          = _mm_load_ss(ptrA+8);
 404     t4          = _mm_loadu_ps(ptrB);
 405     t5          = _mm_loadu_ps(ptrB+4);
 406     t6          = _mm_load_ss(ptrB+8);
 407     t7          = _mm_loadu_ps(ptrC);
 408     t8          = _mm_loadu_ps(ptrC+4);
 409     t9          = _mm_load_ss(ptrC+8);
 410     t10         = _mm_loadu_ps(ptrD);
 411     t11         = _mm_loadu_ps(ptrD+4);
 412     t12         = _mm_load_ss(ptrD+8);
 413
 414     t1          = _mm_sub_ps(t1,t20);
 415     t2          = _mm_sub_ps(t2,t23);
 416     t3          = _mm_sub_ss(t3,z3);
 417     _mm_storeu_ps(ptrA,t1);
 418     _mm_storeu_ps(ptrA+4,t2);
 419     _mm_store_ss(ptrA+8,t3);
 420     t4          = _mm_sub_ps(t4,t21);
 421     t5          = _mm_sub_ps(t5,t24);
 422     t6          = _mm_sub_ss(t6,t17);
 423     _mm_storeu_ps(ptrB,t4);
 424     _mm_storeu_ps(ptrB+4,t5);
 425     _mm_store_ss(ptrB+8,t6);
 426     t7          = _mm_sub_ps(t7,t22);
 427     t8          = _mm_sub_ps(t8,t25);
 428     t9          = _mm_sub_ss(t9,t18);
 429     _mm_storeu_ps(ptrC,t7);
 430     _mm_storeu_ps(ptrC+4,t8);
 431     _mm_store_ss(ptrC+8,t9);
 432     t10         = _mm_sub_ps(t10,t14);
 433     t11         = _mm_sub_ps(t11,t16);
 434     t12         = _mm_sub_ss(t12,t19);
 435     _mm_storeu_ps(ptrD,t10);
 436     _mm_storeu_ps(ptrD+4,t11);
 437     _mm_store_ss(ptrD+8,t12);
 438 }
 439
 440
 441 static gmx_inline void
 442 gmx_mm_decrement_4rvec_4ptr_swizzle_ps(float * gmx_restrict ptrA, float * gmx_restrict ptrB,
 443                                        float * gmx_restrict ptrC, float * gmx_restrict ptrD,
 444                                        __m128 x1, __m128 y1, __m128 z1,
 445                                        __m128 x2, __m128 y2, __m128 z2,
 446                                        __m128 x3, __m128 y3, __m128 z3,
 447                                        __m128 x4, __m128 y4, __m128 z4)
 448 {
 449     __m128 t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,t11;
 450     __m128 t12,t13,t14,t15,t16,t17,t18,t19,t20,t21,t22;
 451     __m128 t23,t24;
 452     t13         = _mm_unpackhi_ps(x1,y1);
 453     x1          = _mm_unpacklo_ps(x1,y1);
 454     t14         = _mm_unpackhi_ps(z1,x2);
 455     z1          = _mm_unpacklo_ps(z1,x2);
 456     t15         = _mm_unpackhi_ps(y2,z2);
 457     y2          = _mm_unpacklo_ps(y2,z2);
 458     t16         = _mm_unpackhi_ps(x3,y3);
 459     x3          = _mm_unpacklo_ps(x3,y3);
 460     t17         = _mm_unpackhi_ps(z3,x4);
 461     z3          = _mm_unpacklo_ps(z3,x4);
 462     t18         = _mm_unpackhi_ps(y4,z4);
 463     y4          = _mm_unpacklo_ps(y4,z4);
 464     t19         = _mm_movelh_ps(x1,z1);
 465     z1          = _mm_movehl_ps(z1,x1);
 466     t20         = _mm_movelh_ps(t13,t14);
 467     t14         = _mm_movehl_ps(t14,t13);
 468     t21         = _mm_movelh_ps(y2,x3);
 469     x3          = _mm_movehl_ps(x3,y2);
 470     t22         = _mm_movelh_ps(t15,t16);
 471     t16         = _mm_movehl_ps(t16,t15);
 472     t23         = _mm_movelh_ps(z3,y4);
 473     y4          = _mm_movehl_ps(y4,z3);
 474     t24         = _mm_movelh_ps(t17,t18);
 475     t18         = _mm_movehl_ps(t18,t17);
 476     t1          = _mm_loadu_ps(ptrA);
 477     t2          = _mm_loadu_ps(ptrA+4);
 478     t3          = _mm_loadu_ps(ptrA+8);
 479     t1          = _mm_sub_ps(t1,t19);
 480     t2          = _mm_sub_ps(t2,t21);
 481     t3          = _mm_sub_ps(t3,t23);
 482     _mm_storeu_ps(ptrA,t1);
 483     _mm_storeu_ps(ptrA+4,t2);
 484     _mm_storeu_ps(ptrA+8,t3);
 485     t4          = _mm_loadu_ps(ptrB);
 486     t5          = _mm_loadu_ps(ptrB+4);
 487     t6          = _mm_loadu_ps(ptrB+8);
 488     t4          = _mm_sub_ps(t4,z1);
 489     t5          = _mm_sub_ps(t5,x3);
 490     t6          = _mm_sub_ps(t6,y4);
 491     _mm_storeu_ps(ptrB,t4);
 492     _mm_storeu_ps(ptrB+4,t5);
 493     _mm_storeu_ps(ptrB+8,t6);
 494     t7          = _mm_loadu_ps(ptrC);
 495     t8          = _mm_loadu_ps(ptrC+4);
 496     t9          = _mm_loadu_ps(ptrC+8);
 497     t7          = _mm_sub_ps(t7,t20);
 498     t8          = _mm_sub_ps(t8,t22);
 499     t9          = _mm_sub_ps(t9,t24);
 500     _mm_storeu_ps(ptrC,t7);
 501     _mm_storeu_ps(ptrC+4,t8);
 502     _mm_storeu_ps(ptrC+8,t9);
 503     t10         = _mm_loadu_ps(ptrD);
 504     t11         = _mm_loadu_ps(ptrD+4);
 505     t12         = _mm_loadu_ps(ptrD+8);
 506     t10         = _mm_sub_ps(t10,t14);
 507     t11         = _mm_sub_ps(t11,t16);
 508     t12         = _mm_sub_ps(t12,t18);
 509     _mm_storeu_ps(ptrD,t10);
 510     _mm_storeu_ps(ptrD+4,t11);
 511     _mm_storeu_ps(ptrD+8,t12);
 512 }
 513
 514
 515
 516 static gmx_inline void
 517 gmx_mm_update_iforce_1atom_swizzle_ps(__m128 fix1, __m128 fiy1, __m128 fiz1,
 518                                       float * gmx_restrict fptr,
 519                                       float * gmx_restrict fshiftptr)
 520 {
 521         __m128 t2,t3;
 522
 523     fix1 = _mm_hadd_ps(fix1,fix1);
 524         fiy1 = _mm_hadd_ps(fiy1,fiz1);
 525
 526         fix1 = _mm_hadd_ps(fix1,fiy1); /* fiz1 fiy1 fix1 fix1 */
 527
 528         t2 = _mm_load_ss(fptr);
 529         t2 = _mm_loadh_pi(t2,(__m64 *)(fptr+1));
 530         t3 = _mm_load_ss(fshiftptr);
 531         t3 = _mm_loadh_pi(t3,(__m64 *)(fshiftptr+1));
 532
 533         t2 = _mm_add_ps(t2,fix1);
 534         t3 = _mm_add_ps(t3,fix1);
 535
 536         _mm_store_ss(fptr,t2);
 537         _mm_storeh_pi((__m64 *)(fptr+1),t2);
 538         _mm_store_ss(fshiftptr,t3);
 539         _mm_storeh_pi((__m64 *)(fshiftptr+1),t3);
 540 }
 541
 542 static gmx_inline void
 543 gmx_mm_update_iforce_3atom_swizzle_ps(__m128 fix1, __m128 fiy1, __m128 fiz1,
 544                                       __m128 fix2, __m128 fiy2, __m128 fiz2,
 545                                       __m128 fix3, __m128 fiy3, __m128 fiz3,
 546                                       float * gmx_restrict fptr,
 547                                       float * gmx_restrict fshiftptr)
 548 {
 549         __m128 t1,t2,t3,t4;
 550
 551         fix1 = _mm_hadd_ps(fix1,fiy1);
 552         fiz1 = _mm_hadd_ps(fiz1,fix2);
 553         fiy2 = _mm_hadd_ps(fiy2,fiz2);
 554         fix3 = _mm_hadd_ps(fix3,fiy3);
 555         fiz3 = _mm_hadd_ps(fiz3,fiz3);
 556
 557         fix1 = _mm_hadd_ps(fix1,fiz1); /* fix2 fiz1 fiy1 fix1 */
 558         fiy2 = _mm_hadd_ps(fiy2,fix3); /* fiy3 fix3 fiz2 fiy2 */
 559         fiz3 = _mm_hadd_ps(fiz3,fiz3); /*  -    -    -   fiz3 */
 560
 561         _mm_storeu_ps(fptr,  _mm_add_ps(fix1,_mm_loadu_ps(fptr)  ));
 562         _mm_storeu_ps(fptr+4,_mm_add_ps(fiy2,_mm_loadu_ps(fptr+4)));
 563         _mm_store_ss (fptr+8,_mm_add_ss(fiz3,_mm_load_ss(fptr+8) ));
 564
 565         t4 = _mm_load_ss(fshiftptr+2);
 566         t4 = _mm_loadh_pi(t4,(__m64 *)(fshiftptr));
 567
 568         t1 = _mm_shuffle_ps(fiz3,fix1,_MM_SHUFFLE(1,0,0,0));   /* fiy1 fix1  -   fiz3 */
 569         t2 = _mm_shuffle_ps(fix1,fiy2,_MM_SHUFFLE(3,2,2,2));   /* fiy3 fix3  -   fiz1 */
 570         t3 = _mm_shuffle_ps(fiy2,fix1,_MM_SHUFFLE(3,3,0,1));   /* fix2 fix2 fiy2 fiz2 */
 571         t3 = _mm_shuffle_ps(t3  ,t3  ,_MM_SHUFFLE(1,2,0,0));   /* fiy2 fix2  -   fiz2 */
 572
 573         t1 = _mm_add_ps(t1,t2);
 574         t3 = _mm_add_ps(t3,t4);
 575         t1 = _mm_add_ps(t1,t3); /* y x - z */
 576
 577         _mm_store_ss(fshiftptr+2,t1);
 578         _mm_storeh_pi((__m64 *)(fshiftptr),t1);
 579 }
 580
 581
 582 static gmx_inline void
 583 gmx_mm_update_iforce_4atom_swizzle_ps(__m128 fix1, __m128 fiy1, __m128 fiz1,
 584                                       __m128 fix2, __m128 fiy2, __m128 fiz2,
 585                                       __m128 fix3, __m128 fiy3, __m128 fiz3,
 586                                       __m128 fix4, __m128 fiy4, __m128 fiz4,
 587                                       float * gmx_restrict fptr,
 588                                       float * gmx_restrict fshiftptr)
 589 {
 590         __m128 t1,t2,t3,t4,t5;
 591
 592         fix1 = _mm_hadd_ps(fix1,fiy1);
 593         fiz1 = _mm_hadd_ps(fiz1,fix2);
 594         fiy2 = _mm_hadd_ps(fiy2,fiz2);
 595         fix3 = _mm_hadd_ps(fix3,fiy3);
 596         fiz3 = _mm_hadd_ps(fiz3,fix4);
 597         fiy4 = _mm_hadd_ps(fiy4,fiz4);
 598
 599         fix1 = _mm_hadd_ps(fix1,fiz1); /* fix2 fiz1 fiy1 fix1 */
 600         fiy2 = _mm_hadd_ps(fiy2,fix3); /* fiy3 fix3 fiz2 fiy2 */
 601         fiz3 = _mm_hadd_ps(fiz3,fiy4); /* fiz4 fiy4 fix4 fiz3 */
 602
 603         _mm_storeu_ps(fptr,  _mm_add_ps(fix1,_mm_loadu_ps(fptr)  ));
 604         _mm_storeu_ps(fptr+4,_mm_add_ps(fiy2,_mm_loadu_ps(fptr+4)));
 605         _mm_storeu_ps(fptr+8,_mm_add_ps(fiz3,_mm_loadu_ps(fptr+8)));
 606
 607         t5 = _mm_load_ss(fshiftptr+2);
 608         t5 = _mm_loadh_pi(t5,(__m64 *)(fshiftptr));
 609
 610         t1 = _mm_shuffle_ps(fix1,fix1,_MM_SHUFFLE(1,0,2,2));
 611         t2 = _mm_shuffle_ps(fiy2,fiy2,_MM_SHUFFLE(3,2,1,1));
 612         t3 = _mm_shuffle_ps(fiz3,fiz3,_MM_SHUFFLE(2,1,0,0));
 613         t4 = _mm_shuffle_ps(fix1,fiy2,_MM_SHUFFLE(0,0,3,3));
 614         t4 = _mm_shuffle_ps(fiz3,t4  ,_MM_SHUFFLE(2,0,3,3));
 615
 616         t1 = _mm_add_ps(t1,t2);
 617         t3 = _mm_add_ps(t3,t4);
 618         t1 = _mm_add_ps(t1,t3);
 619         t5 = _mm_add_ps(t5,t1);
 620
 621         _mm_store_ss(fshiftptr+2,t5);
 622         _mm_storeh_pi((__m64 *)(fshiftptr),t5);
 623 }
 624
 625
 626
 627 static gmx_inline void
 628 gmx_mm_update_1pot_ps(__m128 pot1, float * gmx_restrict ptrA)
 629 {
 630     pot1 = _mm_add_ps(pot1,_mm_movehl_ps(_mm_setzero_ps(),pot1));
 631     pot1 = _mm_add_ps(pot1,_mm_shuffle_ps(pot1,pot1,_MM_SHUFFLE(0,0,0,1)));
 632     _mm_store_ss(ptrA,_mm_add_ss(pot1,_mm_load_ss(ptrA)));
 633 }
 634
 635 static gmx_inline void
 636 gmx_mm_update_2pot_ps(__m128 pot1, float * gmx_restrict ptrA,
 637                       __m128 pot2, float * gmx_restrict ptrB)
 638 {
 639         __m128 t1,t2;
 640         t1   = _mm_movehl_ps(pot2,pot1);
 641         t2   = _mm_movelh_ps(pot1,pot2);
 642         t1   = _mm_add_ps(t1,t2);
 643         t2   = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(3,3,1,1));
 644         pot1 = _mm_add_ps(t1,t2);
 645         pot2 = _mm_movehl_ps(t2,pot1);
 646         _mm_store_ss(ptrA,_mm_add_ss(pot1,_mm_load_ss(ptrA)));
 647         _mm_store_ss(ptrB,_mm_add_ss(pot2,_mm_load_ss(ptrB)));
 648 }
 649
 650
 651 static gmx_inline void
 652 gmx_mm_update_4pot_ps(__m128 pot1, float * gmx_restrict ptrA,
 653                       __m128 pot2, float * gmx_restrict ptrB,
 654                       __m128 pot3, float * gmx_restrict ptrC,
 655                       __m128 pot4, float * gmx_restrict ptrD)
 656 {
 657     _MM_TRANSPOSE4_PS(pot1,pot2,pot3,pot4);
 658     pot1 = _mm_add_ps(_mm_add_ps(pot1,pot2),_mm_add_ps(pot3,pot4));
 659     pot2 = _mm_shuffle_ps(pot1,pot1,_MM_SHUFFLE(1,1,1,1));
 660     pot3 = _mm_shuffle_ps(pot1,pot1,_MM_SHUFFLE(2,2,2,2));
 661     pot4 = _mm_shuffle_ps(pot1,pot1,_MM_SHUFFLE(3,3,3,3));
 662         _mm_store_ss(ptrA,_mm_add_ss(pot1,_mm_load_ss(ptrA)));
 663         _mm_store_ss(ptrB,_mm_add_ss(pot2,_mm_load_ss(ptrB)));
 664         _mm_store_ss(ptrC,_mm_add_ss(pot3,_mm_load_ss(ptrC)));
 665         _mm_store_ss(ptrD,_mm_add_ss(pot4,_mm_load_ss(ptrD)));
 666 }
 667
 668
 669 #endif /* _kernelutil_x86_sse4_1_single_h_ */