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

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