src/gmxlib/nonbonded/nb_kernel_sse4_1_single/kernelutil_x86_sse4_1_single.h

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 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_sse4_1_single_h_
  37 #define _kernelutil_x86_sse4_1_single_h_
  38
  39 #include <math.h>
  40
  41 #include "gmx_x86_sse4_1.h"
  42
  43 #undef gmx_restrict
  44 #define gmx_restrict
  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 gmx_inline 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 gmx_inline __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 gmx_inline 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 gmx_inline 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 gmx_inline 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
 130 static gmx_inline void
 131 gmx_mm_load_shift_and_1rvec_broadcast_ps(const float * gmx_restrict xyz_shift,
 132                                          const float * gmx_restrict xyz,
 133                                          __m128 * gmx_restrict x1,
 134                                          __m128 * gmx_restrict y1,
 135                                          __m128 * gmx_restrict z1)
 136 {
 137     __m128 t1,t2,t3,t4;
 138
 139     t1   = _mm_loadl_pi(_mm_setzero_ps(),(__m64 *)xyz_shift);
 140     t2   = _mm_loadl_pi(_mm_setzero_ps(),(__m64 *)xyz);
 141     t3   = _mm_load_ss(xyz_shift+2);
 142     t4   = _mm_load_ss(xyz+2);
 143     t1   = _mm_add_ps(t1,t2);
 144     t3   = _mm_add_ss(t3,t4);
 145
 146     *x1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(0,0,0,0));
 147     *y1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(1,1,1,1));
 148     *z1  = _mm_shuffle_ps(t3,t3,_MM_SHUFFLE(0,0,0,0));
 149 }
 150
 151
 152 static gmx_inline void
 153 gmx_mm_load_shift_and_3rvec_broadcast_ps(const float * gmx_restrict xyz_shift,
 154                                          const float * gmx_restrict xyz,
 155                                          __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 156                                          __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 157                                          __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3)
 158 {
 159     __m128 tA,tB;
 160     __m128 t1,t2,t3,t4,t5,t6;
 161
 162     tA   = _mm_loadl_pi(_mm_setzero_ps(),(__m64 *)xyz_shift);
 163     tB   = _mm_load_ss(xyz_shift+2);
 164
 165     t1   = _mm_loadu_ps(xyz);
 166     t2   = _mm_loadu_ps(xyz+4);
 167     t3   = _mm_load_ss(xyz+8);
 168
 169     tA   = _mm_movelh_ps(tA,tB);
 170     t4   = _mm_shuffle_ps(tA,tA,_MM_SHUFFLE(0,2,1,0));
 171     t5   = _mm_shuffle_ps(tA,tA,_MM_SHUFFLE(1,0,2,1));
 172     t6   = _mm_shuffle_ps(tA,tA,_MM_SHUFFLE(2,1,0,2));
 173
 174     t1   = _mm_add_ps(t1,t4);
 175     t2   = _mm_add_ps(t2,t5);
 176     t3   = _mm_add_ss(t3,t6);
 177
 178     *x1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(0,0,0,0));
 179     *y1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(1,1,1,1));
 180     *z1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(2,2,2,2));
 181     *x2  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(3,3,3,3));
 182     *y2  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(0,0,0,0));
 183     *z2  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(1,1,1,1));
 184     *x3  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(2,2,2,2));
 185     *y3  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(3,3,3,3));
 186     *z3  = _mm_shuffle_ps(t3,t3,_MM_SHUFFLE(0,0,0,0));
 187 }
 188
 189
 190 static gmx_inline void
 191 gmx_mm_load_shift_and_4rvec_broadcast_ps(const float * gmx_restrict xyz_shift,
 192                                          const float * gmx_restrict xyz,
 193                                          __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 194                                          __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 195                                          __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3,
 196                                          __m128 * gmx_restrict x4, __m128 * gmx_restrict y4, __m128 * gmx_restrict z4)
 197 {
 198     __m128 tA,tB;
 199     __m128 t1,t2,t3,t4,t5,t6;
 200
 201     tA   = _mm_castpd_ps(_mm_load_sd((const double *)xyz_shift));
 202     tB   = _mm_load_ss(xyz_shift+2);
 203
 204     t1   = _mm_loadu_ps(xyz);
 205     t2   = _mm_loadu_ps(xyz+4);
 206     t3   = _mm_loadu_ps(xyz+8);
 207
 208     tA   = _mm_movelh_ps(tA,tB);
 209     t4   = _mm_shuffle_ps(tA,tA,_MM_SHUFFLE(0,2,1,0));
 210     t5   = _mm_shuffle_ps(tA,tA,_MM_SHUFFLE(1,0,2,1));
 211     t6   = _mm_shuffle_ps(tA,tA,_MM_SHUFFLE(2,1,0,2));
 212
 213     t1   = _mm_add_ps(t1,t4);
 214     t2   = _mm_add_ps(t2,t5);
 215     t3   = _mm_add_ps(t3,t6);
 216
 217     *x1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(0,0,0,0));
 218     *y1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(1,1,1,1));
 219     *z1  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(2,2,2,2));
 220     *x2  = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(3,3,3,3));
 221     *y2  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(0,0,0,0));
 222     *z2  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(1,1,1,1));
 223     *x3  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(2,2,2,2));
 224     *y3  = _mm_shuffle_ps(t2,t2,_MM_SHUFFLE(3,3,3,3));
 225     *z3  = _mm_shuffle_ps(t3,t3,_MM_SHUFFLE(0,0,0,0));
 226     *x4  = _mm_shuffle_ps(t3,t3,_MM_SHUFFLE(1,1,1,1));
 227     *y4  = _mm_shuffle_ps(t3,t3,_MM_SHUFFLE(2,2,2,2));
 228     *z4  = _mm_shuffle_ps(t3,t3,_MM_SHUFFLE(3,3,3,3));
 229 }
 230
 231
 232 static gmx_inline void
 233 gmx_mm_load_1rvec_4ptr_swizzle_ps(const float * gmx_restrict ptrA,
 234                                   const float * gmx_restrict ptrB,
 235                                   const float * gmx_restrict ptrC,
 236                                   const float * gmx_restrict ptrD,
 237                                   __m128 *      gmx_restrict x1,
 238                                   __m128 *      gmx_restrict y1,
 239                                   __m128 *      gmx_restrict z1)
 240 {
 241     __m128 t1,t2,t3,t4,t5,t6,t7,t8;
 242     t1   = _mm_castpd_ps(_mm_load_sd((const double *)ptrA));
 243     t2   = _mm_castpd_ps(_mm_load_sd((const double *)ptrB));
 244     t3   = _mm_castpd_ps(_mm_load_sd((const double *)ptrC));
 245     t4   = _mm_castpd_ps(_mm_load_sd((const double *)ptrD));
 246     t5 = _mm_load_ss(ptrA+2);
 247     t6 = _mm_load_ss(ptrB+2);
 248     t7 = _mm_load_ss(ptrC+2);
 249     t8 = _mm_load_ss(ptrD+2);
 250     t1 = _mm_unpacklo_ps(t1,t2);
 251     t3 = _mm_unpacklo_ps(t3,t4);
 252     *x1 = _mm_movelh_ps(t1,t3);
 253     *y1 = _mm_movehl_ps(t3,t1);
 254     t5  = _mm_unpacklo_ps(t5,t6);
 255     t7  = _mm_unpacklo_ps(t7,t8);
 256     *z1 = _mm_movelh_ps(t5,t7);
 257 }
 258
 259
 260 static gmx_inline void
 261 gmx_mm_load_3rvec_4ptr_swizzle_ps(const float * gmx_restrict ptrA,
 262                                   const float * gmx_restrict ptrB,
 263                                   const float * gmx_restrict ptrC,
 264                                   const float * gmx_restrict ptrD,
 265                                   __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 266                                   __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 267                                   __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3)
 268 {
 269     __m128 t1,t2,t3,t4;
 270     t1            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)ptrA ) );
 271     t2            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)ptrB ) );
 272     t3            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)ptrC ) );
 273     t4            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)ptrD ) );
 274     _MM_TRANSPOSE4_PS(t1,t2,t3,t4);
 275     *x1           = t1;
 276     *y1           = t2;
 277     *z1           = t3;
 278     *x2           = t4;
 279     t1            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrA+4) ) );
 280     t2            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrB+4) ) );
 281     t3            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrC+4) ) );
 282     t4            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrD+4) ) );
 283     _MM_TRANSPOSE4_PS(t1,t2,t3,t4);
 284     *y2           = t1;
 285     *z2           = t2;
 286     *x3           = t3;
 287     *y3           = t4;
 288     t1            = _mm_load_ss(ptrA+8);
 289     t2            = _mm_load_ss(ptrB+8);
 290     t3            = _mm_load_ss(ptrC+8);
 291     t4            = _mm_load_ss(ptrD+8);
 292     t1            = _mm_unpacklo_ps(t1,t3);
 293     t3            = _mm_unpacklo_ps(t2,t4);
 294     *z3           = _mm_unpacklo_ps(t1,t3);
 295 }
 296
 297
 298 static gmx_inline void
 299 gmx_mm_load_4rvec_4ptr_swizzle_ps(const float * gmx_restrict ptrA,
 300                                   const float * gmx_restrict ptrB,
 301                                   const float * gmx_restrict ptrC,
 302                                   const float * gmx_restrict ptrD,
 303                                   __m128 * gmx_restrict x1, __m128 * gmx_restrict y1, __m128 * gmx_restrict z1,
 304                                   __m128 * gmx_restrict x2, __m128 * gmx_restrict y2, __m128 * gmx_restrict z2,
 305                                   __m128 * gmx_restrict x3, __m128 * gmx_restrict y3, __m128 * gmx_restrict z3,
 306                                   __m128 * gmx_restrict x4, __m128 * gmx_restrict y4, __m128 * gmx_restrict z4)
 307 {
 308     __m128 t1,t2,t3,t4;
 309     t1            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrA) ) );
 310     t2            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrB) ) );
 311     t3            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrC) ) );
 312     t4            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrD) ) );
 313     _MM_TRANSPOSE4_PS(t1,t2,t3,t4);
 314     *x1           = t1;
 315     *y1           = t2;
 316     *z1           = t3;
 317     *x2           = t4;
 318     t1            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrA+4) ) );
 319     t2            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrB+4) ) );
 320     t3            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrC+4) ) );
 321     t4            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrD+4) ) );
 322     _MM_TRANSPOSE4_PS(t1,t2,t3,t4);
 323     *y2           = t1;
 324     *z2           = t2;
 325     *x3           = t3;
 326     *y3           = t4;
 327     t1            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrA+8) ) );
 328     t2            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrB+8) ) );
 329     t3            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrC+8) ) );
 330     t4            = gmx_mm_castsi128_ps( _mm_lddqu_si128( (void *)(ptrD+8) ) );
 331     _MM_TRANSPOSE4_PS(t1,t2,t3,t4);
 332     *z3           = t1;
 333     *x4           = t2;
 334     *y4           = t3;
 335     *z4           = t4;
 336 }
 337
 338
 339
 340 static gmx_inline void
 341 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(float * ptrA,
 342                                        float * ptrB,
 343                                        float * ptrC,
 344                                        float * ptrD,
 345                                        __m128 x1, __m128 y1, __m128 z1)
 346 {
 347     __m128 t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,t11,t12;
 348     t5          = _mm_unpacklo_ps(y1,z1);
 349     t6          = _mm_unpackhi_ps(y1,z1);
 350     t7          = _mm_shuffle_ps(x1,t5,_MM_SHUFFLE(1,0,0,0));
 351     t8          = _mm_shuffle_ps(x1,t5,_MM_SHUFFLE(3,2,0,1));
 352     t9          = _mm_shuffle_ps(x1,t6,_MM_SHUFFLE(1,0,0,2));
 353     t10         = _mm_shuffle_ps(x1,t6,_MM_SHUFFLE(3,2,0,3));
 354     t1          = _mm_load_ss(ptrA);
 355     t1          = _mm_loadh_pi(t1,(__m64 *)(ptrA+1));
 356     t1          = _mm_sub_ps(t1,t7);
 357     _mm_store_ss(ptrA,t1);
 358     _mm_storeh_pi((__m64 *)(ptrA+1),t1);
 359     t2          = _mm_load_ss(ptrB);
 360     t2          = _mm_loadh_pi(t2,(__m64 *)(ptrB+1));
 361     t2          = _mm_sub_ps(t2,t8);
 362     _mm_store_ss(ptrB,t2);
 363     _mm_storeh_pi((__m64 *)(ptrB+1),t2);
 364     t3          = _mm_load_ss(ptrC);
 365     t3          = _mm_loadh_pi(t3,(__m64 *)(ptrC+1));
 366     t3          = _mm_sub_ps(t3,t9);
 367     _mm_store_ss(ptrC,t3);
 368     _mm_storeh_pi((__m64 *)(ptrC+1),t3);
 369     t4          = _mm_load_ss(ptrD);
 370     t4          = _mm_loadh_pi(t4,(__m64 *)(ptrD+1));
 371     t4          = _mm_sub_ps(t4,t10);
 372     _mm_store_ss(ptrD,t4);
 373     _mm_storeh_pi((__m64 *)(ptrD+1),t4);
 374 }
 375
 376
 377
 378 static gmx_inline void
 379 gmx_mm_decrement_3rvec_4ptr_swizzle_ps(float * gmx_restrict ptrA, float * gmx_restrict ptrB,
 380                                        float * gmx_restrict ptrC, float * gmx_restrict ptrD,
 381                                        __m128 x1, __m128 y1, __m128 z1,
 382                                        __m128 x2, __m128 y2, __m128 z2,
 383                                        __m128 x3, __m128 y3, __m128 z3)
 384 {
 385     __m128 t1,t2,t3,t4,t5,t6,t7,t8,t9,t10;
 386     __m128 t11,t12,t13,t14,t15,t16,t17,t18,t19;
 387     __m128 t20,t21,t22,t23,t24,t25;
 388
 389     t13         = _mm_unpackhi_ps(x1,y1);
 390     x1          = _mm_unpacklo_ps(x1,y1);
 391     t14         = _mm_unpackhi_ps(z1,x2);
 392     z1          = _mm_unpacklo_ps(z1,x2);
 393     t15         = _mm_unpackhi_ps(y2,z2);
 394     y2          = _mm_unpacklo_ps(y2,z2);
 395     t16         = _mm_unpackhi_ps(x3,y3);
 396     x3          = _mm_unpacklo_ps(x3,y3);
 397     t17         = _mm_shuffle_ps(z3,z3,_MM_SHUFFLE(0,0,0,1));
 398     t18         = _mm_movehl_ps(z3,z3);
 399     t19         = _mm_shuffle_ps(t18,t18,_MM_SHUFFLE(0,0,0,1));
 400     t20         = _mm_movelh_ps(x1,z1);
 401     t21         = _mm_movehl_ps(z1,x1);
 402     t22         = _mm_movelh_ps(t13,t14);
 403     t14         = _mm_movehl_ps(t14,t13);
 404     t23         = _mm_movelh_ps(y2,x3);
 405     t24         = _mm_movehl_ps(x3,y2);
 406     t25         = _mm_movelh_ps(t15,t16);
 407     t16         = _mm_movehl_ps(t16,t15);
 408     t1          = _mm_loadu_ps(ptrA);
 409     t2          = _mm_loadu_ps(ptrA+4);
 410     t3          = _mm_load_ss(ptrA+8);
 411     t4          = _mm_loadu_ps(ptrB);
 412     t5          = _mm_loadu_ps(ptrB+4);
 413     t6          = _mm_load_ss(ptrB+8);
 414     t7          = _mm_loadu_ps(ptrC);
 415     t8          = _mm_loadu_ps(ptrC+4);
 416     t9          = _mm_load_ss(ptrC+8);
 417     t10         = _mm_loadu_ps(ptrD);
 418     t11         = _mm_loadu_ps(ptrD+4);
 419     t12         = _mm_load_ss(ptrD+8);
 420
 421     t1          = _mm_sub_ps(t1,t20);
 422     t2          = _mm_sub_ps(t2,t23);
 423     t3          = _mm_sub_ss(t3,z3);
 424     _mm_storeu_ps(ptrA,t1);
 425     _mm_storeu_ps(ptrA+4,t2);
 426     _mm_store_ss(ptrA+8,t3);
 427     t4          = _mm_sub_ps(t4,t21);
 428     t5          = _mm_sub_ps(t5,t24);
 429     t6          = _mm_sub_ss(t6,t17);
 430     _mm_storeu_ps(ptrB,t4);
 431     _mm_storeu_ps(ptrB+4,t5);
 432     _mm_store_ss(ptrB+8,t6);
 433     t7          = _mm_sub_ps(t7,t22);
 434     t8          = _mm_sub_ps(t8,t25);
 435     t9          = _mm_sub_ss(t9,t18);
 436     _mm_storeu_ps(ptrC,t7);
 437     _mm_storeu_ps(ptrC+4,t8);
 438     _mm_store_ss(ptrC+8,t9);
 439     t10         = _mm_sub_ps(t10,t14);
 440     t11         = _mm_sub_ps(t11,t16);
 441     t12         = _mm_sub_ss(t12,t19);
 442     _mm_storeu_ps(ptrD,t10);
 443     _mm_storeu_ps(ptrD+4,t11);
 444     _mm_store_ss(ptrD+8,t12);
 445 }
 446
 447
 448 static gmx_inline void
 449 gmx_mm_decrement_4rvec_4ptr_swizzle_ps(float * gmx_restrict ptrA, float * gmx_restrict ptrB,
 450                                        float * gmx_restrict ptrC, float * gmx_restrict ptrD,
 451                                        __m128 x1, __m128 y1, __m128 z1,
 452                                        __m128 x2, __m128 y2, __m128 z2,
 453                                        __m128 x3, __m128 y3, __m128 z3,
 454                                        __m128 x4, __m128 y4, __m128 z4)
 455 {
 456     __m128 t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,t11;
 457     __m128 t12,t13,t14,t15,t16,t17,t18,t19,t20,t21,t22;
 458     __m128 t23,t24;
 459     t13         = _mm_unpackhi_ps(x1,y1);
 460     x1          = _mm_unpacklo_ps(x1,y1);
 461     t14         = _mm_unpackhi_ps(z1,x2);
 462     z1          = _mm_unpacklo_ps(z1,x2);
 463     t15         = _mm_unpackhi_ps(y2,z2);
 464     y2          = _mm_unpacklo_ps(y2,z2);
 465     t16         = _mm_unpackhi_ps(x3,y3);
 466     x3          = _mm_unpacklo_ps(x3,y3);
 467     t17         = _mm_unpackhi_ps(z3,x4);
 468     z3          = _mm_unpacklo_ps(z3,x4);
 469     t18         = _mm_unpackhi_ps(y4,z4);
 470     y4          = _mm_unpacklo_ps(y4,z4);
 471     t19         = _mm_movelh_ps(x1,z1);
 472     z1          = _mm_movehl_ps(z1,x1);
 473     t20         = _mm_movelh_ps(t13,t14);
 474     t14         = _mm_movehl_ps(t14,t13);
 475     t21         = _mm_movelh_ps(y2,x3);
 476     x3          = _mm_movehl_ps(x3,y2);
 477     t22         = _mm_movelh_ps(t15,t16);
 478     t16         = _mm_movehl_ps(t16,t15);
 479     t23         = _mm_movelh_ps(z3,y4);
 480     y4          = _mm_movehl_ps(y4,z3);
 481     t24         = _mm_movelh_ps(t17,t18);
 482     t18         = _mm_movehl_ps(t18,t17);
 483     t1          = _mm_loadu_ps(ptrA);
 484     t2          = _mm_loadu_ps(ptrA+4);
 485     t3          = _mm_loadu_ps(ptrA+8);
 486     t1          = _mm_sub_ps(t1,t19);
 487     t2          = _mm_sub_ps(t2,t21);
 488     t3          = _mm_sub_ps(t3,t23);
 489     _mm_storeu_ps(ptrA,t1);
 490     _mm_storeu_ps(ptrA+4,t2);
 491     _mm_storeu_ps(ptrA+8,t3);
 492     t4          = _mm_loadu_ps(ptrB);
 493     t5          = _mm_loadu_ps(ptrB+4);
 494     t6          = _mm_loadu_ps(ptrB+8);
 495     t4          = _mm_sub_ps(t4,z1);
 496     t5          = _mm_sub_ps(t5,x3);
 497     t6          = _mm_sub_ps(t6,y4);
 498     _mm_storeu_ps(ptrB,t4);
 499     _mm_storeu_ps(ptrB+4,t5);
 500     _mm_storeu_ps(ptrB+8,t6);
 501     t7          = _mm_loadu_ps(ptrC);
 502     t8          = _mm_loadu_ps(ptrC+4);
 503     t9          = _mm_loadu_ps(ptrC+8);
 504     t7          = _mm_sub_ps(t7,t20);
 505     t8          = _mm_sub_ps(t8,t22);
 506     t9          = _mm_sub_ps(t9,t24);
 507     _mm_storeu_ps(ptrC,t7);
 508     _mm_storeu_ps(ptrC+4,t8);
 509     _mm_storeu_ps(ptrC+8,t9);
 510     t10         = _mm_loadu_ps(ptrD);
 511     t11         = _mm_loadu_ps(ptrD+4);
 512     t12         = _mm_loadu_ps(ptrD+8);
 513     t10         = _mm_sub_ps(t10,t14);
 514     t11         = _mm_sub_ps(t11,t16);
 515     t12         = _mm_sub_ps(t12,t18);
 516     _mm_storeu_ps(ptrD,t10);
 517     _mm_storeu_ps(ptrD+4,t11);
 518     _mm_storeu_ps(ptrD+8,t12);
 519 }
 520
 521
 522
 523 static gmx_inline void
 524 gmx_mm_update_iforce_1atom_swizzle_ps(__m128 fix1, __m128 fiy1, __m128 fiz1,
 525                                       float * gmx_restrict fptr,
 526                                       float * gmx_restrict fshiftptr)
 527 {
 528         __m128 t2,t3;
 529
 530     fix1 = _mm_hadd_ps(fix1,fix1);
 531         fiy1 = _mm_hadd_ps(fiy1,fiz1);
 532
 533         fix1 = _mm_hadd_ps(fix1,fiy1); /* fiz1 fiy1 fix1 fix1 */
 534
 535         t2 = _mm_load_ss(fptr);
 536         t2 = _mm_loadh_pi(t2,(__m64 *)(fptr+1));
 537         t3 = _mm_load_ss(fshiftptr);
 538         t3 = _mm_loadh_pi(t3,(__m64 *)(fshiftptr+1));
 539
 540         t2 = _mm_add_ps(t2,fix1);
 541         t3 = _mm_add_ps(t3,fix1);
 542
 543         _mm_store_ss(fptr,t2);
 544         _mm_storeh_pi((__m64 *)(fptr+1),t2);
 545         _mm_store_ss(fshiftptr,t3);
 546         _mm_storeh_pi((__m64 *)(fshiftptr+1),t3);
 547 }
 548
 549 static gmx_inline void
 550 gmx_mm_update_iforce_3atom_swizzle_ps(__m128 fix1, __m128 fiy1, __m128 fiz1,
 551                                       __m128 fix2, __m128 fiy2, __m128 fiz2,
 552                                       __m128 fix3, __m128 fiy3, __m128 fiz3,
 553                                       float * gmx_restrict fptr,
 554                                       float * gmx_restrict fshiftptr)
 555 {
 556         __m128 t1,t2,t3,t4;
 557
 558         fix1 = _mm_hadd_ps(fix1,fiy1);
 559         fiz1 = _mm_hadd_ps(fiz1,fix2);
 560         fiy2 = _mm_hadd_ps(fiy2,fiz2);
 561         fix3 = _mm_hadd_ps(fix3,fiy3);
 562         fiz3 = _mm_hadd_ps(fiz3,fiz3);
 563
 564         fix1 = _mm_hadd_ps(fix1,fiz1); /* fix2 fiz1 fiy1 fix1 */
 565         fiy2 = _mm_hadd_ps(fiy2,fix3); /* fiy3 fix3 fiz2 fiy2 */
 566         fiz3 = _mm_hadd_ps(fiz3,fiz3); /*  -    -    -   fiz3 */
 567
 568         _mm_storeu_ps(fptr,  _mm_add_ps(fix1,_mm_loadu_ps(fptr)  ));
 569         _mm_storeu_ps(fptr+4,_mm_add_ps(fiy2,_mm_loadu_ps(fptr+4)));
 570         _mm_store_ss (fptr+8,_mm_add_ss(fiz3,_mm_load_ss(fptr+8) ));
 571
 572         t4 = _mm_load_ss(fshiftptr+2);
 573         t4 = _mm_loadh_pi(t4,(__m64 *)(fshiftptr));
 574
 575         t1 = _mm_shuffle_ps(fiz3,fix1,_MM_SHUFFLE(1,0,0,0));   /* fiy1 fix1  -   fiz3 */
 576         t2 = _mm_shuffle_ps(fix1,fiy2,_MM_SHUFFLE(3,2,2,2));   /* fiy3 fix3  -   fiz1 */
 577         t3 = _mm_shuffle_ps(fiy2,fix1,_MM_SHUFFLE(3,3,0,1));   /* fix2 fix2 fiy2 fiz2 */
 578         t3 = _mm_shuffle_ps(t3  ,t3  ,_MM_SHUFFLE(1,2,0,0));   /* fiy2 fix2  -   fiz2 */
 579
 580         t1 = _mm_add_ps(t1,t2);
 581         t3 = _mm_add_ps(t3,t4);
 582         t1 = _mm_add_ps(t1,t3); /* y x - z */
 583
 584         _mm_store_ss(fshiftptr+2,t1);
 585         _mm_storeh_pi((__m64 *)(fshiftptr),t1);
 586 }
 587
 588
 589 static gmx_inline void
 590 gmx_mm_update_iforce_4atom_swizzle_ps(__m128 fix1, __m128 fiy1, __m128 fiz1,
 591                                       __m128 fix2, __m128 fiy2, __m128 fiz2,
 592                                       __m128 fix3, __m128 fiy3, __m128 fiz3,
 593                                       __m128 fix4, __m128 fiy4, __m128 fiz4,
 594                                       float * gmx_restrict fptr,
 595                                       float * gmx_restrict fshiftptr)
 596 {
 597         __m128 t1,t2,t3,t4,t5;
 598
 599         fix1 = _mm_hadd_ps(fix1,fiy1);
 600         fiz1 = _mm_hadd_ps(fiz1,fix2);
 601         fiy2 = _mm_hadd_ps(fiy2,fiz2);
 602         fix3 = _mm_hadd_ps(fix3,fiy3);
 603         fiz3 = _mm_hadd_ps(fiz3,fix4);
 604         fiy4 = _mm_hadd_ps(fiy4,fiz4);
 605
 606         fix1 = _mm_hadd_ps(fix1,fiz1); /* fix2 fiz1 fiy1 fix1 */
 607         fiy2 = _mm_hadd_ps(fiy2,fix3); /* fiy3 fix3 fiz2 fiy2 */
 608         fiz3 = _mm_hadd_ps(fiz3,fiy4); /* fiz4 fiy4 fix4 fiz3 */
 609
 610         _mm_storeu_ps(fptr,  _mm_add_ps(fix1,_mm_loadu_ps(fptr)  ));
 611         _mm_storeu_ps(fptr+4,_mm_add_ps(fiy2,_mm_loadu_ps(fptr+4)));
 612         _mm_storeu_ps(fptr+8,_mm_add_ps(fiz3,_mm_loadu_ps(fptr+8)));
 613
 614         t5 = _mm_load_ss(fshiftptr+2);
 615         t5 = _mm_loadh_pi(t5,(__m64 *)(fshiftptr));
 616
 617         t1 = _mm_shuffle_ps(fix1,fix1,_MM_SHUFFLE(1,0,2,2));
 618         t2 = _mm_shuffle_ps(fiy2,fiy2,_MM_SHUFFLE(3,2,1,1));
 619         t3 = _mm_shuffle_ps(fiz3,fiz3,_MM_SHUFFLE(2,1,0,0));
 620         t4 = _mm_shuffle_ps(fix1,fiy2,_MM_SHUFFLE(0,0,3,3));
 621         t4 = _mm_shuffle_ps(fiz3,t4  ,_MM_SHUFFLE(2,0,3,3));
 622
 623         t1 = _mm_add_ps(t1,t2);
 624         t3 = _mm_add_ps(t3,t4);
 625         t1 = _mm_add_ps(t1,t3);
 626         t5 = _mm_add_ps(t5,t1);
 627
 628         _mm_store_ss(fshiftptr+2,t5);
 629         _mm_storeh_pi((__m64 *)(fshiftptr),t5);
 630 }
 631
 632
 633
 634 static gmx_inline void
 635 gmx_mm_update_1pot_ps(__m128 pot1, float * gmx_restrict ptrA)
 636 {
 637     pot1 = _mm_add_ps(pot1,_mm_movehl_ps(_mm_setzero_ps(),pot1));
 638     pot1 = _mm_add_ps(pot1,_mm_shuffle_ps(pot1,pot1,_MM_SHUFFLE(0,0,0,1)));
 639     _mm_store_ss(ptrA,_mm_add_ss(pot1,_mm_load_ss(ptrA)));
 640 }
 641
 642 static gmx_inline void
 643 gmx_mm_update_2pot_ps(__m128 pot1, float * gmx_restrict ptrA,
 644                       __m128 pot2, float * gmx_restrict ptrB)
 645 {
 646         __m128 t1,t2;
 647         t1   = _mm_movehl_ps(pot2,pot1);
 648         t2   = _mm_movelh_ps(pot1,pot2);
 649         t1   = _mm_add_ps(t1,t2);
 650         t2   = _mm_shuffle_ps(t1,t1,_MM_SHUFFLE(3,3,1,1));
 651         pot1 = _mm_add_ps(t1,t2);
 652         pot2 = _mm_movehl_ps(t2,pot1);
 653         _mm_store_ss(ptrA,_mm_add_ss(pot1,_mm_load_ss(ptrA)));
 654         _mm_store_ss(ptrB,_mm_add_ss(pot2,_mm_load_ss(ptrB)));
 655 }
 656
 657
 658 static gmx_inline void
 659 gmx_mm_update_4pot_ps(__m128 pot1, float * gmx_restrict ptrA,
 660                       __m128 pot2, float * gmx_restrict ptrB,
 661                       __m128 pot3, float * gmx_restrict ptrC,
 662                       __m128 pot4, float * gmx_restrict ptrD)
 663 {
 664     _MM_TRANSPOSE4_PS(pot1,pot2,pot3,pot4);
 665     pot1 = _mm_add_ps(_mm_add_ps(pot1,pot2),_mm_add_ps(pot3,pot4));
 666     pot2 = _mm_shuffle_ps(pot1,pot1,_MM_SHUFFLE(1,1,1,1));
 667     pot3 = _mm_shuffle_ps(pot1,pot1,_MM_SHUFFLE(2,2,2,2));
 668     pot4 = _mm_shuffle_ps(pot1,pot1,_MM_SHUFFLE(3,3,3,3));
 669         _mm_store_ss(ptrA,_mm_add_ss(pot1,_mm_load_ss(ptrA)));
 670         _mm_store_ss(ptrB,_mm_add_ss(pot2,_mm_load_ss(ptrB)));
 671         _mm_store_ss(ptrC,_mm_add_ss(pot3,_mm_load_ss(ptrC)));
 672         _mm_store_ss(ptrD,_mm_add_ss(pot4,_mm_load_ss(ptrD)));
 673 }
 674
 675
 676 #endif /* _kernelutil_x86_sse4_1_single_h_ */