src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/kernelutil_x86_avx_128_fma_double.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_avx_128_fma_double_h_
  30 #define _kernelutil_x86_avx_128_fma_double_h_
  31
  32 #include "gmx_x86_avx_128_fma.h"
  33
  34
  35 static int
  36 gmx_mm_any_lt(__m128d a, __m128d b)
  37 {
  38     return _mm_movemask_pd(_mm_cmplt_pd(a,b));
  39 }
  40
  41
  42 static gmx_inline __m128d
  43 gmx_mm_calc_rsq_pd(__m128d dx, __m128d dy, __m128d dz)
  44 {
  45     return _mm_macc_pd(dx,dx,_mm_macc_pd(dy,dy,_mm_mul_pd(dz,dz)));
  46 }
  47
  48 /* Normal sum of four ymm registers */
  49 #define gmx_mm_sum4_pd(t0,t1,t2,t3)  _mm_add_pd(_mm_add_pd(t0,t1),_mm_add_pd(t2,t3))
  50
  51
  52
  53 /* Load a double value from 1-2 places, merge into xmm register */
  54
  55
  56 static __m128d
  57 gmx_mm_load_2real_swizzle_pd(const double * gmx_restrict ptrA,
  58                              const double * gmx_restrict ptrB)
  59 {
  60     return _mm_unpacklo_pd(_mm_load_sd(ptrA),_mm_load_sd(ptrB));
  61 }
  62
  63 static __m128d
  64 gmx_mm_load_1real_pd(const double * gmx_restrict ptrA)
  65 {
  66     return _mm_load_sd(ptrA);
  67 }
  68
  69
  70 static void
  71 gmx_mm_store_2real_swizzle_pd(double * gmx_restrict ptrA,
  72                               double * gmx_restrict ptrB,
  73                               __m128d xmm1)
  74 {
  75     __m128d t2;
  76
  77     t2       = _mm_unpackhi_pd(xmm1,xmm1);
  78     _mm_store_sd(ptrA,xmm1);
  79     _mm_store_sd(ptrB,t2);
  80 }
  81
  82 static void
  83 gmx_mm_store_1real_pd(double * gmx_restrict ptrA, __m128d xmm1)
  84 {
  85     _mm_store_sd(ptrA,xmm1);
  86 }
  87
  88
  89 /* Similar to store, but increments value in memory */
  90 static void
  91 gmx_mm_increment_2real_swizzle_pd(double * gmx_restrict ptrA,
  92                                   double * gmx_restrict ptrB, __m128d xmm1)
  93 {
  94     __m128d t1;
  95
  96     t1   = _mm_unpackhi_pd(xmm1,xmm1);
  97     xmm1 = _mm_add_sd(xmm1,_mm_load_sd(ptrA));
  98     t1   = _mm_add_sd(t1,_mm_load_sd(ptrB));
  99     _mm_store_sd(ptrA,xmm1);
 100     _mm_store_sd(ptrB,t1);
 101 }
 102
 103 static void
 104 gmx_mm_increment_1real_pd(double * gmx_restrict ptrA, __m128d xmm1)
 105 {
 106     __m128d tmp;
 107
 108     tmp = gmx_mm_load_1real_pd(ptrA);
 109     tmp = _mm_add_sd(tmp,xmm1);
 110     gmx_mm_store_1real_pd(ptrA,tmp);
 111 }
 112
 113
 114
 115 static gmx_inline void
 116 gmx_mm_load_2pair_swizzle_pd(const double * gmx_restrict p1,
 117                              const double * gmx_restrict p2,
 118                              __m128d * gmx_restrict c6,
 119                              __m128d * gmx_restrict c12)
 120 {
 121     __m128d t1,t2,t3;
 122
 123     /* The c6/c12 array should be aligned */
 124     t1   = _mm_loadu_pd(p1);
 125     t2   = _mm_loadu_pd(p2);
 126     *c6  = _mm_unpacklo_pd(t1,t2);
 127     *c12 = _mm_unpackhi_pd(t1,t2);
 128 }
 129
 130 static gmx_inline void
 131 gmx_mm_load_1pair_swizzle_pd(const double * gmx_restrict p1,
 132                              __m128d * gmx_restrict c6,
 133                              __m128d * gmx_restrict c12)
 134 {
 135     *c6     = _mm_load_sd(p1);
 136     *c12    = _mm_load_sd(p1+1);
 137 }
 138
 139
 140 static gmx_inline void
 141 gmx_mm_load_shift_and_1rvec_broadcast_pd(const double * gmx_restrict xyz_shift,
 142                                          const double * gmx_restrict xyz,
 143                                          __m128d * gmx_restrict x1,
 144                                          __m128d * gmx_restrict y1,
 145                                          __m128d * gmx_restrict z1)
 146 {
 147     __m128d mem_xy,mem_z,mem_sxy,mem_sz;
 148
 149     mem_xy  = _mm_loadu_pd(xyz);
 150     mem_z   = _mm_load_sd(xyz+2);
 151     mem_sxy = _mm_loadu_pd(xyz_shift);
 152     mem_sz  = _mm_load_sd(xyz_shift+2);
 153
 154     mem_xy  = _mm_add_pd(mem_xy,mem_sxy);
 155     mem_z   = _mm_add_pd(mem_z,mem_sz);
 156
 157     *x1  = _mm_shuffle_pd(mem_xy,mem_xy,_MM_SHUFFLE2(0,0));
 158     *y1  = _mm_shuffle_pd(mem_xy,mem_xy,_MM_SHUFFLE2(1,1));
 159     *z1  = _mm_shuffle_pd(mem_z,mem_z,_MM_SHUFFLE2(0,0));
 160 }
 161
 162
 163 static gmx_inline void
 164 gmx_mm_load_shift_and_3rvec_broadcast_pd(const double * gmx_restrict xyz_shift,
 165                                          const double * gmx_restrict xyz,
 166                                          __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 167                                          __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 168                                          __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3)
 169 {
 170     __m128d t1,t2,t3,t4,t5,sxy,sz,szx,syz;
 171
 172     t1  = _mm_loadu_pd(xyz);
 173     t2  = _mm_loadu_pd(xyz+2);
 174     t3  = _mm_loadu_pd(xyz+4);
 175     t4  = _mm_loadu_pd(xyz+6);
 176     t5  = _mm_load_sd(xyz+8);
 177
 178     sxy = _mm_loadu_pd(xyz_shift);
 179     sz  = _mm_load_sd(xyz_shift+2);
 180     szx = _mm_shuffle_pd(sz,sxy,_MM_SHUFFLE2(0,0));
 181     syz = _mm_shuffle_pd(sxy,sz,_MM_SHUFFLE2(0,1));
 182
 183     t1  = _mm_add_pd(t1,sxy);
 184     t2  = _mm_add_pd(t2,szx);
 185     t3  = _mm_add_pd(t3,syz);
 186     t4  = _mm_add_pd(t4,sxy);
 187     t5  = _mm_add_sd(t5,sz);
 188
 189     *x1  = _mm_shuffle_pd(t1,t1,_MM_SHUFFLE2(0,0));
 190     *y1  = _mm_shuffle_pd(t1,t1,_MM_SHUFFLE2(1,1));
 191     *z1  = _mm_shuffle_pd(t2,t2,_MM_SHUFFLE2(0,0));
 192     *x2  = _mm_shuffle_pd(t2,t2,_MM_SHUFFLE2(1,1));
 193     *y2  = _mm_shuffle_pd(t3,t3,_MM_SHUFFLE2(0,0));
 194     *z2  = _mm_shuffle_pd(t3,t3,_MM_SHUFFLE2(1,1));
 195     *x3  = _mm_shuffle_pd(t4,t4,_MM_SHUFFLE2(0,0));
 196     *y3  = _mm_shuffle_pd(t4,t4,_MM_SHUFFLE2(1,1));
 197     *z3  = _mm_shuffle_pd(t5,t5,_MM_SHUFFLE2(0,0));
 198 }
 199
 200
 201 static gmx_inline void
 202 gmx_mm_load_shift_and_4rvec_broadcast_pd(const double * gmx_restrict xyz_shift,
 203                                          const double * gmx_restrict xyz,
 204                                          __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 205                                          __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 206                                          __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3,
 207                                          __m128d * gmx_restrict x4, __m128d * gmx_restrict y4, __m128d * gmx_restrict z4)
 208 {
 209     __m128d t1,t2,t3,t4,t5,t6,sxy,sz,szx,syz;
 210
 211     t1  = _mm_loadu_pd(xyz);
 212     t2  = _mm_loadu_pd(xyz+2);
 213     t3  = _mm_loadu_pd(xyz+4);
 214     t4  = _mm_loadu_pd(xyz+6);
 215     t5  = _mm_loadu_pd(xyz+8);
 216     t6  = _mm_loadu_pd(xyz+10);
 217
 218     sxy = _mm_loadu_pd(xyz_shift);
 219     sz  = _mm_load_sd(xyz_shift+2);
 220     szx = _mm_shuffle_pd(sz,sxy,_MM_SHUFFLE2(0,0));
 221     syz = _mm_shuffle_pd(sxy,sz,_MM_SHUFFLE2(0,1));
 222
 223     t1  = _mm_add_pd(t1,sxy);
 224     t2  = _mm_add_pd(t2,szx);
 225     t3  = _mm_add_pd(t3,syz);
 226     t4  = _mm_add_pd(t4,sxy);
 227     t5  = _mm_add_pd(t5,szx);
 228     t6  = _mm_add_pd(t6,syz);
 229
 230     *x1  = _mm_shuffle_pd(t1,t1,_MM_SHUFFLE2(0,0));
 231     *y1  = _mm_shuffle_pd(t1,t1,_MM_SHUFFLE2(1,1));
 232     *z1  = _mm_shuffle_pd(t2,t2,_MM_SHUFFLE2(0,0));
 233     *x2  = _mm_shuffle_pd(t2,t2,_MM_SHUFFLE2(1,1));
 234     *y2  = _mm_shuffle_pd(t3,t3,_MM_SHUFFLE2(0,0));
 235     *z2  = _mm_shuffle_pd(t3,t3,_MM_SHUFFLE2(1,1));
 236     *x3  = _mm_shuffle_pd(t4,t4,_MM_SHUFFLE2(0,0));
 237     *y3  = _mm_shuffle_pd(t4,t4,_MM_SHUFFLE2(1,1));
 238     *z3  = _mm_shuffle_pd(t5,t5,_MM_SHUFFLE2(0,0));
 239     *x4  = _mm_shuffle_pd(t5,t5,_MM_SHUFFLE2(1,1));
 240     *y4  = _mm_shuffle_pd(t6,t6,_MM_SHUFFLE2(0,0));
 241     *z4  = _mm_shuffle_pd(t6,t6,_MM_SHUFFLE2(1,1));
 242 }
 243
 244
 245
 246 static gmx_inline void
 247 gmx_mm_load_1rvec_1ptr_swizzle_pd(const double * gmx_restrict p1,
 248                                   __m128d * gmx_restrict x, __m128d * gmx_restrict y, __m128d * gmx_restrict z)
 249 {
 250          *x            = _mm_load_sd(p1);
 251      *y            = _mm_load_sd(p1+1);
 252      *z            = _mm_load_sd(p1+2);
 253 }
 254
 255 static gmx_inline void
 256 gmx_mm_load_3rvec_1ptr_swizzle_pd(const double * gmx_restrict p1,
 257                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 258                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 259                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3)
 260 {
 261          *x1            = _mm_load_sd(p1);
 262      *y1            = _mm_load_sd(p1+1);
 263      *z1            = _mm_load_sd(p1+2);
 264          *x2            = _mm_load_sd(p1+3);
 265      *y2            = _mm_load_sd(p1+4);
 266      *z2            = _mm_load_sd(p1+5);
 267          *x3            = _mm_load_sd(p1+6);
 268      *y3            = _mm_load_sd(p1+7);
 269      *z3            = _mm_load_sd(p1+8);
 270 }
 271
 272 static gmx_inline void
 273 gmx_mm_load_4rvec_1ptr_swizzle_pd(const double * gmx_restrict p1,
 274                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 275                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 276                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3,
 277                                   __m128d * gmx_restrict x4, __m128d * gmx_restrict y4, __m128d * gmx_restrict z4)
 278 {
 279     *x1            = _mm_load_sd(p1);
 280     *y1            = _mm_load_sd(p1+1);
 281     *z1            = _mm_load_sd(p1+2);
 282     *x2            = _mm_load_sd(p1+3);
 283     *y2            = _mm_load_sd(p1+4);
 284     *z2            = _mm_load_sd(p1+5);
 285     *x3            = _mm_load_sd(p1+6);
 286     *y3            = _mm_load_sd(p1+7);
 287     *z3            = _mm_load_sd(p1+8);
 288     *x4            = _mm_load_sd(p1+9);
 289     *y4            = _mm_load_sd(p1+10);
 290     *z4            = _mm_load_sd(p1+11);
 291 }
 292
 293
 294 static gmx_inline void
 295 gmx_mm_load_1rvec_2ptr_swizzle_pd(const double * gmx_restrict ptrA,
 296                                   const double * gmx_restrict ptrB,
 297                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1)
 298 {
 299     __m128d t1,t2,t3,t4;
 300     t1           = _mm_loadu_pd(ptrA);
 301     t2           = _mm_loadu_pd(ptrB);
 302     t3           = _mm_load_sd(ptrA+2);
 303     t4           = _mm_load_sd(ptrB+2);
 304     GMX_MM_TRANSPOSE2_PD(t1,t2);
 305     *x1          = t1;
 306     *y1          = t2;
 307     *z1          = _mm_unpacklo_pd(t3,t4);
 308 }
 309
 310 static gmx_inline void
 311 gmx_mm_load_3rvec_2ptr_swizzle_pd(const double * gmx_restrict ptrA, const double * gmx_restrict ptrB,
 312                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 313                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 314                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3)
 315 {
 316 __m128d t1,t2,t3,t4,t5,t6,t7,t8,t9,t10;
 317     t1           = _mm_loadu_pd(ptrA);
 318     t2           = _mm_loadu_pd(ptrB);
 319     t3           = _mm_loadu_pd(ptrA+2);
 320     t4           = _mm_loadu_pd(ptrB+2);
 321     t5           = _mm_loadu_pd(ptrA+4);
 322     t6           = _mm_loadu_pd(ptrB+4);
 323     t7           = _mm_loadu_pd(ptrA+6);
 324     t8           = _mm_loadu_pd(ptrB+6);
 325     t9           = _mm_load_sd(ptrA+8);
 326     t10          = _mm_load_sd(ptrB+8);
 327     GMX_MM_TRANSPOSE2_PD(t1,t2);
 328     GMX_MM_TRANSPOSE2_PD(t3,t4);
 329     GMX_MM_TRANSPOSE2_PD(t5,t6);
 330     GMX_MM_TRANSPOSE2_PD(t7,t8);
 331     *x1          = t1;
 332     *y1          = t2;
 333     *z1          = t3;
 334     *x2          = t4;
 335     *y2          = t5;
 336     *z2          = t6;
 337     *x3          = t7;
 338     *y3          = t8;
 339     *z3          = _mm_unpacklo_pd(t9,t10);
 340 }
 341
 342
 343 static gmx_inline void
 344 gmx_mm_load_4rvec_2ptr_swizzle_pd(const double * gmx_restrict ptrA, const double * gmx_restrict ptrB,
 345                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 346                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 347                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3,
 348                                   __m128d * gmx_restrict x4, __m128d * gmx_restrict y4, __m128d * gmx_restrict z4)
 349 {
 350     __m128d t1,t2,t3,t4,t5,t6;
 351     t1           = _mm_loadu_pd(ptrA);
 352     t2           = _mm_loadu_pd(ptrB);
 353     t3           = _mm_loadu_pd(ptrA+2);
 354     t4           = _mm_loadu_pd(ptrB+2);
 355     t5           = _mm_loadu_pd(ptrA+4);
 356     t6           = _mm_loadu_pd(ptrB+4);
 357     GMX_MM_TRANSPOSE2_PD(t1,t2);
 358     GMX_MM_TRANSPOSE2_PD(t3,t4);
 359     GMX_MM_TRANSPOSE2_PD(t5,t6);
 360     *x1          = t1;
 361     *y1          = t2;
 362     *z1          = t3;
 363     *x2          = t4;
 364     *y2          = t5;
 365     *z2          = t6;
 366     t1           = _mm_loadu_pd(ptrA+6);
 367     t2           = _mm_loadu_pd(ptrB+6);
 368     t3           = _mm_loadu_pd(ptrA+8);
 369     t4           = _mm_loadu_pd(ptrB+8);
 370     t5           = _mm_loadu_pd(ptrA+10);
 371     t6           = _mm_loadu_pd(ptrB+10);
 372     GMX_MM_TRANSPOSE2_PD(t1,t2);
 373     GMX_MM_TRANSPOSE2_PD(t3,t4);
 374     GMX_MM_TRANSPOSE2_PD(t5,t6);
 375     *x3          = t1;
 376     *y3          = t2;
 377     *z3          = t3;
 378     *x4          = t4;
 379     *y4          = t5;
 380     *z4          = t6;
 381 }
 382
 383
 384 /* Routines to decrement rvec in memory, typically use for j particle force updates */
 385 static void
 386 gmx_mm_decrement_1rvec_1ptr_noswizzle_pd(double * gmx_restrict ptrA,
 387                                          __m128d xy, __m128d z)
 388 {
 389     __m128d t1,t2;
 390
 391     t1 = _mm_loadu_pd(ptrA);
 392     t2 = _mm_load_sd(ptrA+2);
 393
 394     t1 = _mm_sub_pd(t1,xy);
 395     t2 = _mm_sub_sd(t2,z);
 396
 397     _mm_storeu_pd(ptrA,t1);
 398     _mm_store_sd(ptrA+2,t2);
 399 }
 400
 401
 402 static void
 403 gmx_mm_decrement_3rvec_1ptr_noswizzle_pd(double * gmx_restrict ptrA,
 404                                          __m128d xy1, __m128d z1,
 405                                          __m128d xy2, __m128d z2,
 406                                          __m128d xy3, __m128d z3)
 407 {
 408     __m128d t1,t2;
 409     __m128d tA,tB,tC,tD,tE;
 410
 411     tA   = _mm_loadu_pd(ptrA);
 412     tB   = _mm_loadu_pd(ptrA+2);
 413     tC   = _mm_loadu_pd(ptrA+4);
 414     tD   = _mm_loadu_pd(ptrA+6);
 415     tE   = _mm_load_sd(ptrA+8);
 416
 417     /* xy1: y1 x1 */
 418     t1   = _mm_shuffle_pd(z1,xy2,_MM_SHUFFLE2(0,1)); /* x2 z1 */
 419     t2   = _mm_shuffle_pd(xy2,z2,_MM_SHUFFLE2(0,1)); /* z2 y2 */
 420     /* xy3: y3 x3 */
 421
 422     tA   = _mm_sub_pd(tA,xy1);
 423     tB   = _mm_sub_pd(tB,t1);
 424     tC   = _mm_sub_pd(tC,t2);
 425     tD   = _mm_sub_pd(tD,xy3);
 426     tE   = _mm_sub_sd(tE,z3);
 427
 428     _mm_storeu_pd(ptrA,tA);
 429     _mm_storeu_pd(ptrA+2,tB);
 430     _mm_storeu_pd(ptrA+4,tC);
 431     _mm_storeu_pd(ptrA+6,tD);
 432     _mm_store_sd(ptrA+8,tE);
 433 }
 434
 435 static void
 436 gmx_mm_decrement_4rvec_1ptr_noswizzle_pd(double * gmx_restrict ptrA,
 437                                          __m128d xy1, __m128d z1,
 438                                          __m128d xy2, __m128d z2,
 439                                          __m128d xy3, __m128d z3,
 440                                          __m128d xy4, __m128d z4)
 441 {
 442     __m128d t1,t2,t3,t4;
 443     __m128d tA,tB,tC,tD,tE,tF;
 444
 445     tA   = _mm_loadu_pd(ptrA);
 446     tB   = _mm_loadu_pd(ptrA+2);
 447     tC   = _mm_loadu_pd(ptrA+4);
 448     tD   = _mm_loadu_pd(ptrA+6);
 449     tE   = _mm_loadu_pd(ptrA+8);
 450     tF   = _mm_loadu_pd(ptrA+10);
 451
 452     /* xy1: y1 x1 */
 453     t1   = _mm_shuffle_pd(z1,xy2,_MM_SHUFFLE2(0,0)); /* x2 z1 */
 454     t2   = _mm_shuffle_pd(xy2,z2,_MM_SHUFFLE2(0,1)); /* z2 y2 */
 455     /* xy3: y3 x3 */
 456     t3   = _mm_shuffle_pd(z3,xy4,_MM_SHUFFLE2(0,0)); /* x4 z3 */
 457     t4   = _mm_shuffle_pd(xy4,z4,_MM_SHUFFLE2(0,1)); /* z4 y4 */
 458
 459     tA   = _mm_sub_pd(tA,xy1);
 460     tB   = _mm_sub_pd(tB,t1);
 461     tC   = _mm_sub_pd(tC,t2);
 462     tD   = _mm_sub_pd(tD,xy3);
 463     tE   = _mm_sub_pd(tE,t3);
 464     tF   = _mm_sub_pd(tF,t4);
 465
 466     _mm_storeu_pd(ptrA,tA);
 467     _mm_storeu_pd(ptrA+2,tB);
 468     _mm_storeu_pd(ptrA+4,tC);
 469     _mm_storeu_pd(ptrA+6,tD);
 470     _mm_storeu_pd(ptrA+8,tE);
 471     _mm_storeu_pd(ptrA+10,tF);
 472 }
 473
 474
 475 static void
 476 gmx_mm_decrement_1rvec_1ptr_swizzle_pd(double * gmx_restrict ptrA,
 477                                        __m128d x1, __m128d y1, __m128d z1)
 478 {
 479     __m128d t1,t2,t3;
 480
 481     t1           = _mm_load_sd(ptrA);
 482     t2           = _mm_load_sd(ptrA+1);
 483     t3           = _mm_load_sd(ptrA+2);
 484
 485     t1           = _mm_sub_sd(t1,x1);
 486     t2           = _mm_sub_sd(t2,y1);
 487     t3           = _mm_sub_sd(t3,z1);
 488     _mm_store_sd(ptrA,t1);
 489     _mm_store_sd(ptrA+1,t2);
 490     _mm_store_sd(ptrA+2,t3);
 491 }
 492
 493
 494 static void
 495 gmx_mm_decrement_3rvec_1ptr_swizzle_pd(double * gmx_restrict ptrA,
 496                                        __m128d x1, __m128d y1, __m128d z1,
 497                                        __m128d x2, __m128d y2, __m128d z2,
 498                                        __m128d x3, __m128d y3, __m128d z3)
 499 {
 500     __m128d t1,t2,t3,t4,t5;
 501
 502     t1          = _mm_loadu_pd(ptrA);
 503     t2          = _mm_loadu_pd(ptrA+2);
 504     t3          = _mm_loadu_pd(ptrA+4);
 505     t4          = _mm_loadu_pd(ptrA+6);
 506     t5          = _mm_load_sd(ptrA+8);
 507
 508     x1          = _mm_unpacklo_pd(x1,y1);
 509     z1          = _mm_unpacklo_pd(z1,x2);
 510     y2          = _mm_unpacklo_pd(y2,z2);
 511     x3          = _mm_unpacklo_pd(x3,y3);
 512     /* nothing to be done for z3 */
 513
 514     t1          = _mm_sub_pd(t1,x1);
 515     t2          = _mm_sub_pd(t2,z1);
 516     t3          = _mm_sub_pd(t3,y2);
 517     t4          = _mm_sub_pd(t4,x3);
 518     t5          = _mm_sub_sd(t5,z3);
 519     _mm_storeu_pd(ptrA,t1);
 520     _mm_storeu_pd(ptrA+2,t2);
 521     _mm_storeu_pd(ptrA+4,t3);
 522     _mm_storeu_pd(ptrA+6,t4);
 523     _mm_store_sd(ptrA+8,t5);
 524 }
 525
 526
 527 static void
 528 gmx_mm_decrement_4rvec_1ptr_swizzle_pd(double * gmx_restrict ptrA,
 529                                        __m128d x1, __m128d y1, __m128d z1,
 530                                        __m128d x2, __m128d y2, __m128d z2,
 531                                        __m128d x3, __m128d y3, __m128d z3,
 532                                        __m128d x4, __m128d y4, __m128d z4)
 533 {
 534     __m128d t1,t2,t3,t4,t5,t6;
 535
 536     t1          = _mm_loadu_pd(ptrA);
 537     t2          = _mm_loadu_pd(ptrA+2);
 538     t3          = _mm_loadu_pd(ptrA+4);
 539     t4          = _mm_loadu_pd(ptrA+6);
 540     t5          = _mm_loadu_pd(ptrA+8);
 541     t6          = _mm_loadu_pd(ptrA+10);
 542
 543     x1          = _mm_unpacklo_pd(x1,y1);
 544     z1          = _mm_unpacklo_pd(z1,x2);
 545     y2          = _mm_unpacklo_pd(y2,z2);
 546     x3          = _mm_unpacklo_pd(x3,y3);
 547     z3          = _mm_unpacklo_pd(z3,x4);
 548     y4          = _mm_unpacklo_pd(y4,z4);
 549
 550     _mm_storeu_pd(ptrA,    _mm_sub_pd( t1,x1 ));
 551     _mm_storeu_pd(ptrA+2,  _mm_sub_pd( t2,z1 ));
 552     _mm_storeu_pd(ptrA+4,  _mm_sub_pd( t3,y2 ));
 553     _mm_storeu_pd(ptrA+6,  _mm_sub_pd( t4,x3 ));
 554     _mm_storeu_pd(ptrA+8,  _mm_sub_pd( t5,z3 ));
 555     _mm_storeu_pd(ptrA+10, _mm_sub_pd( t6,y4 ));
 556 }
 557
 558 static void
 559 gmx_mm_decrement_1rvec_2ptr_swizzle_pd(double * gmx_restrict ptrA, double * gmx_restrict ptrB,
 560                                        __m128d x1, __m128d y1, __m128d z1)
 561 {
 562     __m128d t1,t2,t3,t4,t5,t6,t7;
 563
 564     t1          = _mm_loadu_pd(ptrA);
 565     t2          = _mm_load_sd(ptrA+2);
 566     t3          = _mm_loadu_pd(ptrB);
 567     t4          = _mm_load_sd(ptrB+2);
 568
 569     t5          = _mm_unpacklo_pd(x1,y1);
 570     t6          = _mm_unpackhi_pd(x1,y1);
 571     t7          = _mm_unpackhi_pd(z1,z1);
 572
 573     t1          = _mm_sub_pd(t1,t5);
 574     t2          = _mm_sub_sd(t2,z1);
 575
 576     t3          = _mm_sub_pd(t3,t6);
 577     t4          = _mm_sub_sd(t4,t7);
 578
 579     _mm_storeu_pd(ptrA,t1);
 580     _mm_store_sd(ptrA+2,t2);
 581     _mm_storeu_pd(ptrB,t3);
 582     _mm_store_sd(ptrB+2,t4);
 583 }
 584
 585
 586 static void
 587 gmx_mm_decrement_3rvec_2ptr_swizzle_pd(double * gmx_restrict ptrA, double * gmx_restrict ptrB,
 588                                        __m128d x1, __m128d y1, __m128d z1,
 589                                        __m128d x2, __m128d y2, __m128d z2,
 590                                        __m128d x3, __m128d y3, __m128d z3)
 591 {
 592     __m128d t1,t2,t3,t4,t5,t6,t7,t8,t9,t10;
 593     __m128d tA,tB,tC,tD,tE,tF,tG,tH,tI;
 594
 595     t1          = _mm_loadu_pd(ptrA);
 596     t2          = _mm_loadu_pd(ptrA+2);
 597     t3          = _mm_loadu_pd(ptrA+4);
 598     t4          = _mm_loadu_pd(ptrA+6);
 599     t5          = _mm_load_sd(ptrA+8);
 600     t6          = _mm_loadu_pd(ptrB);
 601     t7          = _mm_loadu_pd(ptrB+2);
 602     t8          = _mm_loadu_pd(ptrB+4);
 603     t9          = _mm_loadu_pd(ptrB+6);
 604     t10         = _mm_load_sd(ptrB+8);
 605
 606     tA          = _mm_unpacklo_pd(x1,y1);
 607     tB          = _mm_unpackhi_pd(x1,y1);
 608     tC          = _mm_unpacklo_pd(z1,x2);
 609     tD          = _mm_unpackhi_pd(z1,x2);
 610     tE          = _mm_unpacklo_pd(y2,z2);
 611     tF          = _mm_unpackhi_pd(y2,z2);
 612     tG          = _mm_unpacklo_pd(x3,y3);
 613     tH          = _mm_unpackhi_pd(x3,y3);
 614     tI          = _mm_unpackhi_pd(z3,z3);
 615
 616     t1          = _mm_sub_pd(t1,tA);
 617     t2          = _mm_sub_pd(t2,tC);
 618     t3          = _mm_sub_pd(t3,tE);
 619     t4          = _mm_sub_pd(t4,tG);
 620     t5          = _mm_sub_sd(t5,z3);
 621
 622     t6          = _mm_sub_pd(t6,tB);
 623     t7          = _mm_sub_pd(t7,tD);
 624     t8          = _mm_sub_pd(t8,tF);
 625     t9          = _mm_sub_pd(t9,tH);
 626     t10         = _mm_sub_sd(t10,tI);
 627
 628     _mm_storeu_pd(ptrA,t1);
 629     _mm_storeu_pd(ptrA+2,t2);
 630     _mm_storeu_pd(ptrA+4,t3);
 631     _mm_storeu_pd(ptrA+6,t4);
 632     _mm_store_sd(ptrA+8,t5);
 633     _mm_storeu_pd(ptrB,t6);
 634     _mm_storeu_pd(ptrB+2,t7);
 635     _mm_storeu_pd(ptrB+4,t8);
 636     _mm_storeu_pd(ptrB+6,t9);
 637     _mm_store_sd(ptrB+8,t10);
 638 }
 639
 640
 641 static void
 642 gmx_mm_decrement_4rvec_2ptr_swizzle_pd(double * gmx_restrict ptrA, double * gmx_restrict ptrB,
 643                                        __m128d x1, __m128d y1, __m128d z1,
 644                                        __m128d x2, __m128d y2, __m128d z2,
 645                                        __m128d x3, __m128d y3, __m128d z3,
 646                                        __m128d x4, __m128d y4, __m128d z4)
 647 {
 648     __m128d t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,t11,t12;
 649     __m128d tA,tB,tC,tD,tE,tF,tG,tH,tI,tJ,tK,tL;
 650
 651     t1          = _mm_loadu_pd(ptrA);
 652     t2          = _mm_loadu_pd(ptrA+2);
 653     t3          = _mm_loadu_pd(ptrA+4);
 654     t4          = _mm_loadu_pd(ptrA+6);
 655     t5          = _mm_loadu_pd(ptrA+8);
 656     t6          = _mm_loadu_pd(ptrA+10);
 657     t7          = _mm_loadu_pd(ptrB);
 658     t8          = _mm_loadu_pd(ptrB+2);
 659     t9          = _mm_loadu_pd(ptrB+4);
 660     t10         = _mm_loadu_pd(ptrB+6);
 661     t11         = _mm_loadu_pd(ptrB+8);
 662     t12         = _mm_loadu_pd(ptrB+10);
 663
 664     tA          = _mm_unpacklo_pd(x1,y1);
 665     tB          = _mm_unpackhi_pd(x1,y1);
 666     tC          = _mm_unpacklo_pd(z1,x2);
 667     tD          = _mm_unpackhi_pd(z1,x2);
 668     tE          = _mm_unpacklo_pd(y2,z2);
 669     tF          = _mm_unpackhi_pd(y2,z2);
 670     tG          = _mm_unpacklo_pd(x3,y3);
 671     tH          = _mm_unpackhi_pd(x3,y3);
 672     tI          = _mm_unpacklo_pd(z3,x4);
 673     tJ          = _mm_unpackhi_pd(z3,x4);
 674     tK          = _mm_unpacklo_pd(y4,z4);
 675     tL          = _mm_unpackhi_pd(y4,z4);
 676
 677     t1          = _mm_sub_pd(t1,tA);
 678     t2          = _mm_sub_pd(t2,tC);
 679     t3          = _mm_sub_pd(t3,tE);
 680     t4          = _mm_sub_pd(t4,tG);
 681     t5          = _mm_sub_pd(t5,tI);
 682     t6          = _mm_sub_pd(t6,tK);
 683
 684     t7          = _mm_sub_pd(t7,tB);
 685     t8          = _mm_sub_pd(t8,tD);
 686     t9          = _mm_sub_pd(t9,tF);
 687     t10         = _mm_sub_pd(t10,tH);
 688     t11         = _mm_sub_pd(t11,tJ);
 689     t12         = _mm_sub_pd(t12,tL);
 690
 691     _mm_storeu_pd(ptrA,  t1);
 692     _mm_storeu_pd(ptrA+2,t2);
 693     _mm_storeu_pd(ptrA+4,t3);
 694     _mm_storeu_pd(ptrA+6,t4);
 695     _mm_storeu_pd(ptrA+8,t5);
 696     _mm_storeu_pd(ptrA+10,t6);
 697     _mm_storeu_pd(ptrB,  t7);
 698     _mm_storeu_pd(ptrB+2,t8);
 699     _mm_storeu_pd(ptrB+4,t9);
 700     _mm_storeu_pd(ptrB+6,t10);
 701     _mm_storeu_pd(ptrB+8,t11);
 702     _mm_storeu_pd(ptrB+10,t12);
 703 }
 704
 705
 706
 707 static gmx_inline void
 708 gmx_mm_update_iforce_1atom_swizzle_pd(__m128d fix1, __m128d fiy1, __m128d fiz1,
 709                                       double * gmx_restrict fptr,
 710                                       double * gmx_restrict fshiftptr)
 711 {
 712     fix1 = _mm_hadd_pd(fix1,fiy1);
 713     fiz1 = _mm_hadd_pd(fiz1,fiz1);
 714
 715     _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr), fix1 ));
 716     _mm_store_sd( fptr+2, _mm_add_sd( _mm_load_sd(fptr+2), fiz1 ));
 717
 718     _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 ));
 719     _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 ));
 720 }
 721
 722 static gmx_inline void
 723 gmx_mm_update_iforce_3atom_swizzle_pd(__m128d fix1, __m128d fiy1, __m128d fiz1,
 724                                       __m128d fix2, __m128d fiy2, __m128d fiz2,
 725                                       __m128d fix3, __m128d fiy3, __m128d fiz3,
 726                                       double * gmx_restrict fptr,
 727                                       double * gmx_restrict fshiftptr)
 728 {
 729     __m128d t1,t2;
 730
 731     fix1 = _mm_hadd_pd(fix1,fiy1);
 732     fiz1 = _mm_hadd_pd(fiz1,fix2);
 733     fiy2 = _mm_hadd_pd(fiy2,fiz2);
 734     fix3 = _mm_hadd_pd(fix3,fiy3);
 735     fiz3 = _mm_hadd_pd(fiz3,fiz3);
 736
 737     _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr), fix1 ));
 738     _mm_storeu_pd( fptr+2, _mm_add_pd( _mm_loadu_pd(fptr+2), fiz1 ));
 739     _mm_storeu_pd( fptr+4, _mm_add_pd( _mm_loadu_pd(fptr+4), fiy2 ));
 740     _mm_storeu_pd( fptr+6, _mm_add_pd( _mm_loadu_pd(fptr+6), fix3 ));
 741     _mm_store_sd( fptr+8, _mm_add_sd( _mm_load_sd(fptr+8), fiz3 ));
 742
 743     fix1 = _mm_add_pd(fix1,fix3);
 744     t1   = _mm_shuffle_pd(fiz1,fiy2,_MM_SHUFFLE2(0,1));
 745     fix1 = _mm_add_pd(fix1,t1); /* x and y sums */
 746
 747     t2   = _mm_shuffle_pd(fiy2,fiy2,_MM_SHUFFLE2(1,1));
 748     fiz1 = _mm_add_sd(fiz1,fiz3);
 749     fiz1 = _mm_add_sd(fiz1,t2); /* z sum */
 750
 751     _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 ));
 752     _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 ));
 753 }
 754
 755
 756 static gmx_inline void
 757 gmx_mm_update_iforce_4atom_swizzle_pd(__m128d fix1, __m128d fiy1, __m128d fiz1,
 758                                       __m128d fix2, __m128d fiy2, __m128d fiz2,
 759                                       __m128d fix3, __m128d fiy3, __m128d fiz3,
 760                                       __m128d fix4, __m128d fiy4, __m128d fiz4,
 761                                       double * gmx_restrict fptr,
 762                                       double * gmx_restrict fshiftptr)
 763 {
 764     __m128d t1,t2;
 765
 766     fix1 = _mm_hadd_pd(fix1,fiy1);
 767     fiz1 = _mm_hadd_pd(fiz1,fix2);
 768     fiy2 = _mm_hadd_pd(fiy2,fiz2);
 769     fix3 = _mm_hadd_pd(fix3,fiy3);
 770     fiz3 = _mm_hadd_pd(fiz3,fix4);
 771     fiy4 = _mm_hadd_pd(fiy4,fiz4);
 772
 773     _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr),       fix1 ));
 774     _mm_storeu_pd( fptr+2, _mm_add_pd( _mm_loadu_pd(fptr+2),   fiz1 ));
 775     _mm_storeu_pd( fptr+4, _mm_add_pd( _mm_loadu_pd(fptr+4),   fiy2 ));
 776     _mm_storeu_pd( fptr+6, _mm_add_pd( _mm_loadu_pd(fptr+6),   fix3 ));
 777     _mm_storeu_pd( fptr+8, _mm_add_pd( _mm_loadu_pd(fptr+8),   fiz3 ));
 778     _mm_storeu_pd( fptr+10, _mm_add_pd( _mm_loadu_pd(fptr+10), fiy4 ));
 779
 780     t1 = _mm_shuffle_pd(fiz1,fiy2,_MM_SHUFFLE2(0,1));
 781     fix1 = _mm_add_pd(fix1,t1);
 782     t2 = _mm_shuffle_pd(fiz3,fiy4,_MM_SHUFFLE2(0,1));
 783     fix3 = _mm_add_pd(fix3,t2);
 784     fix1 = _mm_add_pd(fix1,fix3); /* x and y sums */
 785
 786     fiz1 = _mm_add_sd(fiz1, _mm_unpackhi_pd(fiy2,fiy2));
 787     fiz3 = _mm_add_sd(fiz3, _mm_unpackhi_pd(fiy4,fiy4));
 788     fiz1 = _mm_add_sd(fiz1,fiz3); /* z sum */
 789
 790     _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 ));
 791     _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 ));
 792 }
 793
 794
 795
 796 static gmx_inline void
 797 gmx_mm_update_1pot_pd(__m128d pot1, double * gmx_restrict ptrA)
 798 {
 799     pot1 = _mm_hadd_pd(pot1,pot1);
 800     _mm_store_sd(ptrA,_mm_add_sd(pot1,_mm_load_sd(ptrA)));
 801 }
 802
 803 static gmx_inline void
 804 gmx_mm_update_2pot_pd(__m128d pot1, double * gmx_restrict ptrA,
 805                       __m128d pot2, double * gmx_restrict ptrB)
 806 {
 807     pot1 = _mm_hadd_pd(pot1,pot2);
 808     pot2 = _mm_unpackhi_pd(pot1,pot1);
 809
 810     _mm_store_sd(ptrA,_mm_add_sd(pot1,_mm_load_sd(ptrA)));
 811     _mm_store_sd(ptrB,_mm_add_sd(pot2,_mm_load_sd(ptrB)));
 812 }
 813
 814
 815 #endif /* _kernelutil_x86_avx_128_fma_double_h_ */