src/gmxlib/nonbonded/nb_kernel_sse2_single/kernelutil_x86_sse2_single.h

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2011-2012, The GROMACS Development Team
   5  * Copyright (c) 2012, by the GROMACS development team, led by
   6  * David van der Spoel, Berk Hess, Erik Lindahl, and including many
   7  * others, as listed in the AUTHORS file in the top-level source
   8  * directory and at http://www.gromacs.org.
   9  *
  10  * GROMACS is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public License
  12  * as published by the Free Software Foundation; either version 2.1
  13  * of the License, or (at your option) any later version.
  14  *
  15  * GROMACS is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * Lesser General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU Lesser General Public
  21  * License along with GROMACS; if not, see
  22  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  23  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  24  *
  25  * If you want to redistribute modifications to GROMACS, please
  26  * consider that scientific software is very special. Version
  27  * control is crucial - bugs must be traceable. We will be happy to
  28  * consider code for inclusion in the official distribution, but
  29  * derived work must not be called official GROMACS. Details are found
  30  * in the README & COPYING files - if they are missing, get the
  31  * official version at http://www.gromacs.org.
  32  *
  33  * To help us fund GROMACS development, we humbly ask that you cite
  34  * the research papers on the package. Check out http://www.gromacs.org.
  35  */
  36 #ifndef _kernelutil_x86_sse2_single_h_
  37 #define _kernelutil_x86_sse2_single_h_
  38
  39 /* We require SSE2 now! */
  40
  41 #include <math.h>
  42
  43 #include "gmx_x86_sse2.h"
  44
  45
  46 /* Normal sum of four xmm registers */
  47 #define gmx_mm_sum4_ps(t0,t1,t2,t3)  _mm_add_ps(_mm_add_ps(t0,t1),_mm_add_ps(t2,t3))
  48
  49 static gmx_inline __m128
  50 gmx_mm_calc_rsq_ps(__m128 dx, __m128 dy, __m128 dz)
  51 {
  52     return _mm_add_ps( _mm_add_ps( _mm_mul_ps(dx,dx), _mm_mul_ps(dy,dy) ), _mm_mul_ps(dz,dz) );
  53 }
  54
  55 static int
  56 gmx_mm_any_lt(__m128 a, __m128 b)
  57 {
  58     return _mm_movemask_ps(_mm_cmplt_ps(a,b));
  59 }
  60
  61 /* Load a single value from 1-4 places, merge into xmm register */
  62
  63 static __m128
  64 gmx_mm_load_4real_swizzle_ps(const float * gmx_restrict ptrA,
  65                              const float * gmx_restrict ptrB,
  66                              const float * gmx_restrict ptrC,
  67                              const float * gmx_restrict ptrD)
  68 {
  69     __m128 t1,t2;
  70
  71     t1 = _mm_unpacklo_ps(_mm_load_ss(ptrA),_mm_load_ss(ptrC));
  72     t2 = _mm_unpacklo_ps(_mm_load_ss(ptrB),_mm_load_ss(ptrD));
  73     return _mm_unpacklo_ps(t1,t2);
  74 }
  75
  76 static void
  77 gmx_mm_store_4real_swizzle_ps(float * gmx_restrict ptrA,
  78                               float * gmx_restrict ptrB,
  79                               float * gmx_restrict ptrC,
  80                               float * gmx_restrict ptrD,
  81                               __m128 xmm1)
  82 {
  83     __m128 t2,t3,t4;
  84
  85     t3       = _mm_movehl_ps(_mm_setzero_ps(),xmm1);
  86     t2       = _mm_shuffle_ps(xmm1,xmm1,_MM_SHUFFLE(1,1,1,1));
  87     t4       = _mm_shuffle_ps(t3,t3,_MM_SHUFFLE(1,1,1,1));
  88     _mm_store_ss(ptrA,xmm1);
  89     _mm_store_ss(ptrB,t2);
  90     _mm_store_ss(ptrC,t3);
  91     _mm_store_ss(ptrD,t4);
  92 }
  93
  94 /* Similar to store, but increments value in memory */
  95 static void
  96 gmx_mm_increment_4real_swizzle_ps(float * gmx_restrict ptrA,
  97                                   float * gmx_restrict ptrB,
  98                                   float * gmx_restrict ptrC,
  99                                   float * gmx_restrict ptrD, __m128 xmm1)
 100 {
 101     __m128 tmp;
 102
 103     tmp = gmx_mm_load_4real_swizzle_ps(ptrA,ptrB,ptrC,ptrD);
 104     tmp = _mm_add_ps(tmp,xmm1);
 105     gmx_mm_store_4real_swizzle_ps(ptrA,ptrB,ptrC,ptrD,tmp);
 106 }
 107
 108
 109 static void
 110 gmx_mm_load_4pair_swizzle_ps(const float * gmx_restrict p1,
 111                              const float * gmx_restrict p2,
 112                              const float * gmx_restrict p3,
 113                              const float * gmx_restrict p4,
 114                              __m128 * gmx_restrict c6,
 115                              __m128 * gmx_restrict c12)
 116 {
 117     __m128 t1,t2,t3,t4;
 118
 119     t1   = _mm_loadl_pi(_mm_setzero_ps(),(__m64 *)p1);   /* - - c12a  c6a */
 120     t2   = _mm_loadl_pi(_mm_setzero_ps(),(__m64 *)p2);   /* - - c12b  c6b */
 121     t3   = _mm_loadl_pi(_mm_setzero_ps(),(__m64 *)p3);   /* - - c12c  c6c */
 122     t4   = _mm_loadl_pi(_mm_setzero_ps(),(__m64 *)p4);   /* - - c12d  c6d */
 123     t1   = _mm_unpacklo_ps(t1,t2);
 124     t2   = _mm_unpacklo_ps(t3,t4);
 125     *c6  = _mm_movelh_ps(t1,t2);
 126     *c12 = _mm_movehl_ps(t2,t1);
 127 }
 128
 129 /* Routines to load 1-4 rvec from 4 places.
 130  * We mainly use these to load coordinates. The extra routines
 131  * are very efficient for the water-water loops, since we e.g.
 132  * know that a TIP4p water has 4 atoms, so we should load 12 floats+shuffle.
 133  */
 134
 135
 136 static gmx_inline void
 137 gmx_mm_load_shift_and_1rvec_broadcast_ps(const float * gmx_restrict xyz_shift,
 138                                          const float * gmx_restrict xyz,
 139                                          __m128 * gmx_restrict x1,
 140                                          __m128 * gmx_restrict y1,
 141                                          __m128 * gmx_restrict z1)
 142 {
 143     __m128 t1,t2,t3,t4;
 144
 145     t1   = _mm_loadl_pi(_mm_setzero_ps(),(__m64 *)xyz_shift);
 146     t2   = _mm_loadl_pi(_mm_setzero_ps(),(__m64 *)xyz);
 147     t3   = _mm_load_ss(xyz_shift+2);
 148     t4   = _mm_load_ss(xyz+2);
 149     t1   = _mm_add_ps(t1,t2);
 150     t3   = _mm_add_ss(t3,t4);
 151
 152     *x1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(0,0,0,0));
 153     *y1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(1,1,1,1));
 154     *z1  = _mm_shuffle_ps(t3,t3,_MM_SHUFFLE(0,0,0,0));
 155 }
 156
 157
 158 static gmx_inline void
 159 gmx_mm_load_shift_and_3rvec_broadcast_ps(const float * gmx_restrict xyz_shift,
 160                                          const float * gmx_restrict xyz,
 161                                          __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 162                                          __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 163                                          __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3)
 164 {
 165     __m128 tA,tB;
 166     __m128 t1,t2,t3,t4,t5,t6;
 167
 168     tA   = _mm_loadl_pi(_mm_setzero_ps(),(__m64 *)xyz_shift);
 169     tB   = _mm_load_ss(xyz_shift+2);
 170
 171     t1   = _mm_loadu_ps(xyz);
 172     t2   = _mm_loadu_ps(xyz+4);
 173     t3   = _mm_load_ss(xyz+8);
 174
 175     tA   = _mm_movelh_ps(tA,tB);
 176     t4   = _mm_shuffle_ps(tA,tA,_MM_SHUFFLE(0,2,1,0));
 177     t5   = _mm_shuffle_ps(tA,tA,_MM_SHUFFLE(1,0,2,1));
 178     t6   = _mm_shuffle_ps(tA,tA,_MM_SHUFFLE(2,1,0,2));
 179
 180     t1   = _mm_add_ps(t1,t4);
 181     t2   = _mm_add_ps(t2,t5);
 182     t3   = _mm_add_ss(t3,t6);
 183
 184     *x1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(0,0,0,0));
 185     *y1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(1,1,1,1));
 186     *z1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(2,2,2,2));
 187     *x2  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(3,3,3,3));
 188     *y2  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(0,0,0,0));
 189     *z2  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(1,1,1,1));
 190     *x3  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(2,2,2,2));
 191     *y3  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(3,3,3,3));
 192     *z3  = _mm_shuffle_ps(t3,t3,_MM_SHUFFLE(0,0,0,0));
 193 }
 194
 195
 196 static gmx_inline void
 197 gmx_mm_load_shift_and_4rvec_broadcast_ps(const float * gmx_restrict xyz_shift,
 198                                          const float * gmx_restrict xyz,
 199                                          __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 200                                          __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 201                                          __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3,
 202                                          __m128 * gmx_restrict x4, __m128 * gmx_restrict y4, __m128 * gmx_restrict z4)
 203 {
 204     __m128 tA,tB;
 205     __m128 t1,t2,t3,t4,t5,t6;
 206
 207     tA   = _mm_castpd_ps(_mm_load_sd((const double *)xyz_shift));
 208     tB   = _mm_load_ss(xyz_shift+2);
 209
 210     t1   = _mm_loadu_ps(xyz);
 211     t2   = _mm_loadu_ps(xyz+4);
 212     t3   = _mm_loadu_ps(xyz+8);
 213
 214     tA   = _mm_movelh_ps(tA,tB);
 215     t4   = _mm_shuffle_ps(tA,tA,_MM_SHUFFLE(0,2,1,0));
 216     t5   = _mm_shuffle_ps(tA,tA,_MM_SHUFFLE(1,0,2,1));
 217     t6   = _mm_shuffle_ps(tA,tA,_MM_SHUFFLE(2,1,0,2));
 218
 219     t1   = _mm_add_ps(t1,t4);
 220     t2   = _mm_add_ps(t2,t5);
 221     t3   = _mm_add_ps(t3,t6);
 222
 223     *x1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(0,0,0,0));
 224     *y1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(1,1,1,1));
 225     *z1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(2,2,2,2));
 226     *x2  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(3,3,3,3));
 227     *y2  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(0,0,0,0));
 228     *z2  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(1,1,1,1));
 229     *x3  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(2,2,2,2));
 230     *y3  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(3,3,3,3));
 231     *z3  = _mm_shuffle_ps(t3,t3,_MM_SHUFFLE(0,0,0,0));
 232     *x4  = _mm_shuffle_ps(t3,t3,_MM_SHUFFLE(1,1,1,1));
 233     *y4  = _mm_shuffle_ps(t3,t3,_MM_SHUFFLE(2,2,2,2));
 234     *z4  = _mm_shuffle_ps(t3,t3,_MM_SHUFFLE(3,3,3,3));
 235 }
 236
 237
 238 static void
 239 gmx_mm_load_1rvec_4ptr_swizzle_ps(const float * gmx_restrict ptrA,
 240                                   const float * gmx_restrict ptrB,
 241                                   const float * gmx_restrict ptrC,
 242                                   const float * gmx_restrict ptrD,
 243                                   __m128 *      gmx_restrict x1,
 244                                   __m128 *      gmx_restrict y1,
 245                                   __m128 *      gmx_restrict z1)
 246 {
 247     __m128 t1,t2,t3,t4,t5,t6,t7,t8;
 248     t1   = _mm_castpd_ps(_mm_load_sd((const double *)ptrA));
 249     t2   = _mm_castpd_ps(_mm_load_sd((const double *)ptrB));
 250     t3   = _mm_castpd_ps(_mm_load_sd((const double *)ptrC));
 251     t4   = _mm_castpd_ps(_mm_load_sd((const double *)ptrD));
 252     t5 = _mm_load_ss(ptrA+2);
 253     t6 = _mm_load_ss(ptrB+2);
 254     t7 = _mm_load_ss(ptrC+2);
 255     t8 = _mm_load_ss(ptrD+2);
 256     t1 = _mm_unpacklo_ps(t1,t2);
 257     t3 = _mm_unpacklo_ps(t3,t4);
 258     *x1 = _mm_movelh_ps(t1,t3);
 259     *y1 = _mm_movehl_ps(t3,t1);
 260     t5  = _mm_unpacklo_ps(t5,t6);
 261     t7  = _mm_unpacklo_ps(t7,t8);
 262     *z1 = _mm_movelh_ps(t5,t7);
 263 }
 264
 265
 266 static void
 267 gmx_mm_load_3rvec_4ptr_swizzle_ps(const float * gmx_restrict ptrA,
 268                                   const float * gmx_restrict ptrB,
 269                                   const float * gmx_restrict ptrC,
 270                                   const float * gmx_restrict ptrD,
 271                                   __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 272                                   __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 273                                   __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3)
 274 {
 275     __m128 t1,t2,t3,t4;
 276     t1            = _mm_loadu_ps(ptrA);
 277     t2            = _mm_loadu_ps(ptrB);
 278     t3            = _mm_loadu_ps(ptrC);
 279     t4            = _mm_loadu_ps(ptrD);
 280     _MM_TRANSPOSE4_PS(t1,t2,t3,t4);
 281     *x1           = t1;
 282     *y1           = t2;
 283     *z1           = t3;
 284     *x2           = t4;
 285     t1            = _mm_loadu_ps(ptrA+4);
 286     t2            = _mm_loadu_ps(ptrB+4);
 287     t3            = _mm_loadu_ps(ptrC+4);
 288     t4            = _mm_loadu_ps(ptrD+4);
 289     _MM_TRANSPOSE4_PS(t1,t2,t3,t4);
 290     *y2           = t1;
 291     *z2           = t2;
 292     *x3           = t3;
 293     *y3           = t4;
 294     t1            = _mm_load_ss(ptrA+8);
 295     t2            = _mm_load_ss(ptrB+8);
 296     t3            = _mm_load_ss(ptrC+8);
 297     t4            = _mm_load_ss(ptrD+8);
 298     t1            = _mm_unpacklo_ps(t1,t3);
 299     t3            = _mm_unpacklo_ps(t2,t4);
 300     *z3           = _mm_unpacklo_ps(t1,t3);
 301 }
 302
 303
 304 static void
 305 gmx_mm_load_4rvec_4ptr_swizzle_ps(const float * gmx_restrict ptrA,
 306                                   const float * gmx_restrict ptrB,
 307                                   const float * gmx_restrict ptrC,
 308                                   const float * gmx_restrict ptrD,
 309                                   __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 310                                   __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 311                                   __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3,
 312                                   __m128 * gmx_restrict x4, __m128 * gmx_restrict y4, __m128 * gmx_restrict z4)
 313 {
 314     __m128 t1,t2,t3,t4;
 315     t1            = _mm_loadu_ps(ptrA);
 316     t2            = _mm_loadu_ps(ptrB);
 317     t3            = _mm_loadu_ps(ptrC);
 318     t4            = _mm_loadu_ps(ptrD);
 319     _MM_TRANSPOSE4_PS(t1,t2,t3,t4);
 320     *x1           = t1;
 321     *y1           = t2;
 322     *z1           = t3;
 323     *x2           = t4;
 324     t1            = _mm_loadu_ps(ptrA+4);
 325     t2            = _mm_loadu_ps(ptrB+4);
 326     t3            = _mm_loadu_ps(ptrC+4);
 327     t4            = _mm_loadu_ps(ptrD+4);
 328     _MM_TRANSPOSE4_PS(t1,t2,t3,t4);
 329     *y2           = t1;
 330     *z2           = t2;
 331     *x3           = t3;
 332     *y3           = t4;
 333     t1            = _mm_loadu_ps(ptrA+8);
 334     t2            = _mm_loadu_ps(ptrB+8);
 335     t3            = _mm_loadu_ps(ptrC+8);
 336     t4            = _mm_loadu_ps(ptrD+8);
 337     _MM_TRANSPOSE4_PS(t1,t2,t3,t4);
 338     *z3           = t1;
 339     *x4           = t2;
 340     *y4           = t3;
 341     *z4           = t4;
 342 }
 343
 344
 345 static void
 346 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(float * gmx_restrict ptrA,
 347                                        float * gmx_restrict ptrB,
 348                                        float * gmx_restrict ptrC,
 349                                        float * gmx_restrict ptrD,
 350                                        __m128 x1, __m128 y1, __m128 z1)
 351 {
 352     __m128 t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,t11,t12;
 353     t5          = _mm_unpacklo_ps(y1,z1);
 354     t6          = _mm_unpackhi_ps(y1,z1);
 355     t7          = _mm_shuffle_ps(x1,t5,_MM_SHUFFLE(1,0,0,0));
 356     t8          = _mm_shuffle_ps(x1,t5,_MM_SHUFFLE(3,2,0,1));
 357     t9          = _mm_shuffle_ps(x1,t6,_MM_SHUFFLE(1,0,0,2));
 358     t10         = _mm_shuffle_ps(x1,t6,_MM_SHUFFLE(3,2,0,3));
 359     t1          = _mm_load_ss(ptrA);
 360     t1          = _mm_loadh_pi(t1,(__m64 *)(ptrA+1));
 361     t1          = _mm_sub_ps(t1,t7);
 362     _mm_store_ss(ptrA,t1);
 363     _mm_storeh_pi((__m64 *)(ptrA+1),t1);
 364     t2          = _mm_load_ss(ptrB);
 365     t2          = _mm_loadh_pi(t2,(__m64 *)(ptrB+1));
 366     t2          = _mm_sub_ps(t2,t8);
 367     _mm_store_ss(ptrB,t2);
 368     _mm_storeh_pi((__m64 *)(ptrB+1),t2);
 369     t3          = _mm_load_ss(ptrC);
 370     t3          = _mm_loadh_pi(t3,(__m64 *)(ptrC+1));
 371     t3          = _mm_sub_ps(t3,t9);
 372     _mm_store_ss(ptrC,t3);
 373     _mm_storeh_pi((__m64 *)(ptrC+1),t3);
 374     t4          = _mm_load_ss(ptrD);
 375     t4          = _mm_loadh_pi(t4,(__m64 *)(ptrD+1));
 376     t4          = _mm_sub_ps(t4,t10);
 377     _mm_store_ss(ptrD,t4);
 378     _mm_storeh_pi((__m64 *)(ptrD+1),t4);
 379 }
 380
 381
 382
 383 static void
 384 gmx_mm_decrement_3rvec_4ptr_swizzle_ps(float * gmx_restrict ptrA, float * gmx_restrict ptrB,
 385                                        float * gmx_restrict ptrC, float * gmx_restrict ptrD,
 386                                        __m128 x1, __m128 y1, __m128 z1,
 387                                        __m128 x2, __m128 y2, __m128 z2,
 388                                        __m128 x3, __m128 y3, __m128 z3)
 389 {
 390     __m128 t1,t2,t3,t4,t5,t6,t7,t8,t9,t10;
 391     __m128 t11,t12,t13,t14,t15,t16,t17,t18,t19;
 392     __m128 t20,t21,t22,t23,t24,t25;
 393
 394     t13         = _mm_unpackhi_ps(x1,y1);
 395     x1          = _mm_unpacklo_ps(x1,y1);
 396     t14         = _mm_unpackhi_ps(z1,x2);
 397     z1          = _mm_unpacklo_ps(z1,x2);
 398     t15         = _mm_unpackhi_ps(y2,z2);
 399     y2          = _mm_unpacklo_ps(y2,z2);
 400     t16         = _mm_unpackhi_ps(x3,y3);
 401     x3          = _mm_unpacklo_ps(x3,y3);
 402     t17         = _mm_shuffle_ps(z3,z3,_MM_SHUFFLE(0,0,0,1));
 403     t18         = _mm_movehl_ps(z3,z3);
 404     t19         = _mm_shuffle_ps(t18,t18,_MM_SHUFFLE(0,0,0,1));
 405     t20         = _mm_movelh_ps(x1,z1);
 406     t21         = _mm_movehl_ps(z1,x1);
 407     t22         = _mm_movelh_ps(t13,t14);
 408     t14         = _mm_movehl_ps(t14,t13);
 409     t23         = _mm_movelh_ps(y2,x3);
 410     t24         = _mm_movehl_ps(x3,y2);
 411     t25         = _mm_movelh_ps(t15,t16);
 412     t16         = _mm_movehl_ps(t16,t15);
 413     t1          = _mm_loadu_ps(ptrA);
 414     t2          = _mm_loadu_ps(ptrA+4);
 415     t3          = _mm_load_ss(ptrA+8);
 416     t1          = _mm_sub_ps(t1,t20);
 417     t2          = _mm_sub_ps(t2,t23);
 418     t3          = _mm_sub_ss(t3,z3);
 419     _mm_storeu_ps(ptrA,t1);
 420     _mm_storeu_ps(ptrA+4,t2);
 421     _mm_store_ss(ptrA+8,t3);
 422     t4          = _mm_loadu_ps(ptrB);
 423     t5          = _mm_loadu_ps(ptrB+4);
 424     t6          = _mm_load_ss(ptrB+8);
 425     t4          = _mm_sub_ps(t4,t21);
 426     t5          = _mm_sub_ps(t5,t24);
 427     t6          = _mm_sub_ss(t6,t17);
 428     _mm_storeu_ps(ptrB,t4);
 429     _mm_storeu_ps(ptrB+4,t5);
 430     _mm_store_ss(ptrB+8,t6);
 431     t7          = _mm_loadu_ps(ptrC);
 432     t8          = _mm_loadu_ps(ptrC+4);
 433     t9          = _mm_load_ss(ptrC+8);
 434     t7          = _mm_sub_ps(t7,t22);
 435     t8          = _mm_sub_ps(t8,t25);
 436     t9          = _mm_sub_ss(t9,t18);
 437     _mm_storeu_ps(ptrC,t7);
 438     _mm_storeu_ps(ptrC+4,t8);
 439     _mm_store_ss(ptrC+8,t9);
 440     t10         = _mm_loadu_ps(ptrD);
 441     t11         = _mm_loadu_ps(ptrD+4);
 442     t12         = _mm_load_ss(ptrD+8);
 443     t10         = _mm_sub_ps(t10,t14);
 444     t11         = _mm_sub_ps(t11,t16);
 445     t12         = _mm_sub_ss(t12,t19);
 446     _mm_storeu_ps(ptrD,t10);
 447     _mm_storeu_ps(ptrD+4,t11);
 448     _mm_store_ss(ptrD+8,t12);
 449 }
 450
 451
 452 static void
 453 gmx_mm_decrement_4rvec_4ptr_swizzle_ps(float * gmx_restrict ptrA, float * gmx_restrict ptrB,
 454                                        float * gmx_restrict ptrC, float * gmx_restrict ptrD,
 455                                        __m128 x1, __m128 y1, __m128 z1,
 456                                        __m128 x2, __m128 y2, __m128 z2,
 457                                        __m128 x3, __m128 y3, __m128 z3,
 458                                        __m128 x4, __m128 y4, __m128 z4)
 459 {
 460     __m128 t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,t11;
 461     __m128 t12,t13,t14,t15,t16,t17,t18,t19,t20,t21,t22;
 462     __m128 t23,t24;
 463     t13         = _mm_unpackhi_ps(x1,y1);
 464     x1          = _mm_unpacklo_ps(x1,y1);
 465     t14         = _mm_unpackhi_ps(z1,x2);
 466     z1          = _mm_unpacklo_ps(z1,x2);
 467     t15         = _mm_unpackhi_ps(y2,z2);
 468     y2          = _mm_unpacklo_ps(y2,z2);
 469     t16         = _mm_unpackhi_ps(x3,y3);
 470     x3          = _mm_unpacklo_ps(x3,y3);
 471     t17         = _mm_unpackhi_ps(z3,x4);
 472     z3          = _mm_unpacklo_ps(z3,x4);
 473     t18         = _mm_unpackhi_ps(y4,z4);
 474     y4          = _mm_unpacklo_ps(y4,z4);
 475     t19         = _mm_movelh_ps(x1,z1);
 476     z1          = _mm_movehl_ps(z1,x1);
 477     t20         = _mm_movelh_ps(t13,t14);
 478     t14         = _mm_movehl_ps(t14,t13);
 479     t21         = _mm_movelh_ps(y2,x3);
 480     x3          = _mm_movehl_ps(x3,y2);
 481     t22         = _mm_movelh_ps(t15,t16);
 482     t16         = _mm_movehl_ps(t16,t15);
 483     t23         = _mm_movelh_ps(z3,y4);
 484     y4          = _mm_movehl_ps(y4,z3);
 485     t24         = _mm_movelh_ps(t17,t18);
 486     t18         = _mm_movehl_ps(t18,t17);
 487     t1          = _mm_loadu_ps(ptrA);
 488     t2          = _mm_loadu_ps(ptrA+4);
 489     t3          = _mm_loadu_ps(ptrA+8);
 490     t1          = _mm_sub_ps(t1,t19);
 491     t2          = _mm_sub_ps(t2,t21);
 492     t3          = _mm_sub_ps(t3,t23);
 493     _mm_storeu_ps(ptrA,t1);
 494     _mm_storeu_ps(ptrA+4,t2);
 495     _mm_storeu_ps(ptrA+8,t3);
 496     t4          = _mm_loadu_ps(ptrB);
 497     t5          = _mm_loadu_ps(ptrB+4);
 498     t6          = _mm_loadu_ps(ptrB+8);
 499     t4          = _mm_sub_ps(t4,z1);
 500     t5          = _mm_sub_ps(t5,x3);
 501     t6          = _mm_sub_ps(t6,y4);
 502     _mm_storeu_ps(ptrB,t4);
 503     _mm_storeu_ps(ptrB+4,t5);
 504     _mm_storeu_ps(ptrB+8,t6);
 505     t7          = _mm_loadu_ps(ptrC);
 506     t8          = _mm_loadu_ps(ptrC+4);
 507     t9          = _mm_loadu_ps(ptrC+8);
 508     t7          = _mm_sub_ps(t7,t20);
 509     t8          = _mm_sub_ps(t8,t22);
 510     t9          = _mm_sub_ps(t9,t24);
 511     _mm_storeu_ps(ptrC,t7);
 512     _mm_storeu_ps(ptrC+4,t8);
 513     _mm_storeu_ps(ptrC+8,t9);
 514     t10         = _mm_loadu_ps(ptrD);
 515     t11         = _mm_loadu_ps(ptrD+4);
 516     t12         = _mm_loadu_ps(ptrD+8);
 517     t10         = _mm_sub_ps(t10,t14);
 518     t11         = _mm_sub_ps(t11,t16);
 519     t12         = _mm_sub_ps(t12,t18);
 520     _mm_storeu_ps(ptrD,t10);
 521     _mm_storeu_ps(ptrD+4,t11);
 522     _mm_storeu_ps(ptrD+8,t12);
 523 }
 524
 525
 526
 527 static gmx_inline void
 528 gmx_mm_update_iforce_1atom_swizzle_ps(__m128 fix1, __m128 fiy1, __m128 fiz1,
 529                                       float * gmx_restrict fptr,
 530                                       float * gmx_restrict fshiftptr)
 531 {
 532     __m128 t1,t2,t3;
 533
 534     /* transpose data */
 535     t1 = fix1;
 536     _MM_TRANSPOSE4_PS(fix1,t1,fiy1,fiz1);
 537     fix1 = _mm_add_ps(_mm_add_ps(fix1,t1), _mm_add_ps(fiy1,fiz1));
 538
 539     t2 = _mm_load_ss(fptr);
 540     t2 = _mm_loadh_pi(t2,(__m64 *)(fptr+1));
 541     t3 = _mm_load_ss(fshiftptr);
 542     t3 = _mm_loadh_pi(t3,(__m64 *)(fshiftptr+1));
 543
 544     t2 = _mm_add_ps(t2,fix1);
 545     t3 = _mm_add_ps(t3,fix1);
 546
 547     _mm_store_ss(fptr,t2);
 548     _mm_storeh_pi((__m64 *)(fptr+1),t2);
 549     _mm_store_ss(fshiftptr,t3);
 550     _mm_storeh_pi((__m64 *)(fshiftptr+1),t3);
 551 }
 552
 553 static gmx_inline void
 554 gmx_mm_update_iforce_3atom_swizzle_ps(__m128 fix1, __m128 fiy1, __m128 fiz1,
 555                                       __m128 fix2, __m128 fiy2, __m128 fiz2,
 556                                       __m128 fix3, __m128 fiy3, __m128 fiz3,
 557                                       float * gmx_restrict fptr,
 558                                       float * gmx_restrict fshiftptr)
 559 {
 560     __m128 t1,t2,t3,t4;
 561
 562     /* transpose data */
 563     _MM_TRANSPOSE4_PS(fix1,fiy1,fiz1,fix2);
 564     _MM_TRANSPOSE4_PS(fiy2,fiz2,fix3,fiy3);
 565     t2   = _mm_movehl_ps(_mm_setzero_ps(),fiz3);
 566     t1   = _mm_shuffle_ps(fiz3,fiz3,_MM_SHUFFLE(0,0,0,1));
 567     t3   = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(0,0,0,1));
 568
 569     fix1 = _mm_add_ps(_mm_add_ps(fix1,fiy1), _mm_add_ps(fiz1,fix2));
 570     fiy2 = _mm_add_ps(_mm_add_ps(fiy2,fiz2), _mm_add_ps(fix3,fiy3));
 571     fiz3 = _mm_add_ss(_mm_add_ps(fiz3,t1)  , _mm_add_ps(t2,t3));
 572
 573     _mm_storeu_ps(fptr,  _mm_add_ps(fix1,_mm_loadu_ps(fptr)  ));
 574     _mm_storeu_ps(fptr+4,_mm_add_ps(fiy2,_mm_loadu_ps(fptr+4)));
 575     _mm_store_ss (fptr+8,_mm_add_ss(fiz3,_mm_load_ss(fptr+8) ));
 576
 577     t4 = _mm_load_ss(fshiftptr+2);
 578     t4 = _mm_loadh_pi(t4,(__m64 *)(fshiftptr));
 579
 580     t1 = _mm_shuffle_ps(fiz3,fix1,_MM_SHUFFLE(1,0,0,0));   /* fiy1 fix1  -   fiz3 */
 581     t2 = _mm_shuffle_ps(fix1,fiy2,_MM_SHUFFLE(3,2,2,2));   /* fiy3 fix3  -   fiz1 */
 582     t3 = _mm_shuffle_ps(fiy2,fix1,_MM_SHUFFLE(3,3,0,1));   /* fix2 fix2 fiy2 fiz2 */
 583     t3 = _mm_shuffle_ps(t3  ,t3  ,_MM_SHUFFLE(1,2,0,0));   /* fiy2 fix2  -   fiz2 */
 584
 585     t1 = _mm_add_ps(t1,t2);
 586     t3 = _mm_add_ps(t3,t4);
 587     t1 = _mm_add_ps(t1,t3); /* y x - z */
 588
 589     _mm_store_ss(fshiftptr+2,t1);
 590     _mm_storeh_pi((__m64 *)(fshiftptr),t1);
 591 }
 592
 593
 594 static gmx_inline void
 595 gmx_mm_update_iforce_4atom_swizzle_ps(__m128 fix1, __m128 fiy1, __m128 fiz1,
 596                                       __m128 fix2, __m128 fiy2, __m128 fiz2,
 597                                       __m128 fix3, __m128 fiy3, __m128 fiz3,
 598                                       __m128 fix4, __m128 fiy4, __m128 fiz4,
 599                                       float * gmx_restrict fptr,
 600                                       float * gmx_restrict fshiftptr)
 601 {
 602     __m128 t1,t2,t3,t4,t5;
 603
 604     /* transpose data */
 605     _MM_TRANSPOSE4_PS(fix1,fiy1,fiz1,fix2);
 606     _MM_TRANSPOSE4_PS(fiy2,fiz2,fix3,fiy3);
 607     _MM_TRANSPOSE4_PS(fiz3,fix4,fiy4,fiz4);
 608
 609     fix1 = _mm_add_ps(_mm_add_ps(fix1,fiy1), _mm_add_ps(fiz1,fix2));
 610     fiy2 = _mm_add_ps(_mm_add_ps(fiy2,fiz2), _mm_add_ps(fix3,fiy3));
 611     fiz3 = _mm_add_ps(_mm_add_ps(fiz3,fix4), _mm_add_ps(fiy4,fiz4));
 612
 613     _mm_storeu_ps(fptr,  _mm_add_ps(fix1,_mm_loadu_ps(fptr)  ));
 614     _mm_storeu_ps(fptr+4,_mm_add_ps(fiy2,_mm_loadu_ps(fptr+4)));
 615     _mm_storeu_ps(fptr+8,_mm_add_ps(fiz3,_mm_loadu_ps(fptr+8)));
 616
 617     t5 = _mm_load_ss(fshiftptr+2);
 618     t5 = _mm_loadh_pi(t5,(__m64 *)(fshiftptr));
 619
 620     t1 = _mm_shuffle_ps(fix1,fix1,_MM_SHUFFLE(1,0,2,2));
 621     t2 = _mm_shuffle_ps(fiy2,fiy2,_MM_SHUFFLE(3,2,1,1));
 622     t3 = _mm_shuffle_ps(fiz3,fiz3,_MM_SHUFFLE(2,1,0,0));
 623     t4 = _mm_shuffle_ps(fix1,fiy2,_MM_SHUFFLE(0,0,3,3));
 624     t4 = _mm_shuffle_ps(fiz3,t4  ,_MM_SHUFFLE(2,0,3,3));
 625
 626     t1 = _mm_add_ps(t1,t2);
 627     t3 = _mm_add_ps(t3,t4);
 628     t1 = _mm_add_ps(t1,t3);
 629     t5 = _mm_add_ps(t5,t1);
 630
 631     _mm_store_ss(fshiftptr+2,t5);
 632     _mm_storeh_pi((__m64 *)(fshiftptr),t5);
 633 }
 634
 635
 636
 637 static void
 638 gmx_mm_update_1pot_ps(__m128 pot1, float * gmx_restrict ptrA)
 639 {
 640     pot1 = _mm_add_ps(pot1,_mm_movehl_ps(_mm_setzero_ps(),pot1));
 641     pot1 = _mm_add_ps(pot1,_mm_shuffle_ps(pot1,pot1,_MM_SHUFFLE(0,0,0,1)));
 642     _mm_store_ss(ptrA,_mm_add_ss(pot1,_mm_load_ss(ptrA)));
 643 }
 644
 645 static void
 646 gmx_mm_update_2pot_ps(__m128 pot1, float * gmx_restrict ptrA,
 647                       __m128 pot2, float * gmx_restrict ptrB)
 648 {
 649     __m128 t1,t2;
 650     t1   = _mm_movehl_ps(pot2,pot1);
 651     t2   = _mm_movelh_ps(pot1,pot2);
 652     t1   = _mm_add_ps(t1,t2);
 653     t2   = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(3,3,1,1));
 654     pot1 = _mm_add_ps(t1,t2);
 655     pot2 = _mm_movehl_ps(t2,pot1);
 656     _mm_store_ss(ptrA,_mm_add_ss(pot1,_mm_load_ss(ptrA)));
 657     _mm_store_ss(ptrB,_mm_add_ss(pot2,_mm_load_ss(ptrB)));
 658 }
 659
 660
 661 static void
 662 gmx_mm_update_4pot_ps(__m128 pot1, float * gmx_restrict ptrA,
 663                       __m128 pot2, float * gmx_restrict ptrB,
 664                       __m128 pot3, float * gmx_restrict ptrC,
 665                       __m128 pot4, float * gmx_restrict ptrD)
 666 {
 667     _MM_TRANSPOSE4_PS(pot1,pot2,pot3,pot4);
 668     pot1 = _mm_add_ps(_mm_add_ps(pot1,pot2),_mm_add_ps(pot3,pot4));
 669     pot2 = _mm_shuffle_ps(pot1,pot1,_MM_SHUFFLE(1,1,1,1));
 670     pot3 = _mm_shuffle_ps(pot1,pot1,_MM_SHUFFLE(2,2,2,2));
 671     pot4 = _mm_shuffle_ps(pot1,pot1,_MM_SHUFFLE(3,3,3,3));
 672     _mm_store_ss(ptrA,_mm_add_ss(pot1,_mm_load_ss(ptrA)));
 673     _mm_store_ss(ptrB,_mm_add_ss(pot2,_mm_load_ss(ptrB)));
 674     _mm_store_ss(ptrC,_mm_add_ss(pot3,_mm_load_ss(ptrC)));
 675     _mm_store_ss(ptrD,_mm_add_ss(pot4,_mm_load_ss(ptrD)));
 676 }
 677
 678
 679 #endif /* _kernelutil_x86_sse2_single_h_ */