src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/kernelutil_x86_avx_128_fma_double.h

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 #ifndef _kernelutil_x86_avx_128_fma_double_h_
  36 #define _kernelutil_x86_avx_128_fma_double_h_
  37
  38 #include <math.h>
  39 #include <immintrin.h>
  40 #ifdef _MSC_VER
  41 #    include <intrin.h>
  42 #else
  43 #    include <x86intrin.h>
  44 #endif
  45
  46 #include "config.h"
  47
  48 #define gmx_mm_castsi128_pd   _mm_castsi128_pd
  49 #define gmx_mm_extract_epi32  _mm_extract_epi32
  50
  51 #define GMX_MM_TRANSPOSE2_PD(row0, row1) {           \
  52         __m128d __gmx_t1 = row0;                         \
  53         row0           = _mm_unpacklo_pd(row0, row1);     \
  54         row1           = _mm_unpackhi_pd(__gmx_t1, row1); \
  55 }
  56
  57 static gmx_inline int gmx_simdcall
  58 gmx_mm_any_lt(__m128d a, __m128d b)
  59 {
  60     return _mm_movemask_pd(_mm_cmplt_pd(a, b));
  61 }
  62
  63
  64 static gmx_inline __m128d gmx_simdcall
  65 gmx_mm_calc_rsq_pd(__m128d dx, __m128d dy, __m128d dz)
  66 {
  67     return _mm_macc_pd(dx, dx, _mm_macc_pd(dy, dy, _mm_mul_pd(dz, dz)));
  68 }
  69
  70 /* Normal sum of four ymm registers */
  71 #define gmx_mm_sum4_pd(t0, t1, t2, t3)  _mm_add_pd(_mm_add_pd(t0, t1), _mm_add_pd(t2, t3))
  72
  73
  74
  75 /* Load a double value from 1-2 places, merge into xmm register */
  76
  77
  78 static gmx_inline __m128d gmx_simdcall
  79 gmx_mm_load_2real_swizzle_pd(const double * gmx_restrict ptrA,
  80                              const double * gmx_restrict ptrB)
  81 {
  82     return _mm_unpacklo_pd(_mm_load_sd(ptrA), _mm_load_sd(ptrB));
  83 }
  84
  85 static gmx_inline __m128d gmx_simdcall
  86 gmx_mm_load_1real_pd(const double * gmx_restrict ptrA)
  87 {
  88     return _mm_load_sd(ptrA);
  89 }
  90
  91
  92 static gmx_inline void gmx_simdcall
  93 gmx_mm_store_2real_swizzle_pd(double * gmx_restrict ptrA,
  94                               double * gmx_restrict ptrB,
  95                               __m128d               xmm1)
  96 {
  97     __m128d t2;
  98
  99     t2       = _mm_unpackhi_pd(xmm1, xmm1);
 100     _mm_store_sd(ptrA, xmm1);
 101     _mm_store_sd(ptrB, t2);
 102 }
 103
 104 static gmx_inline void gmx_simdcall
 105 gmx_mm_store_1real_pd(double * gmx_restrict ptrA, __m128d xmm1)
 106 {
 107     _mm_store_sd(ptrA, xmm1);
 108 }
 109
 110
 111 /* Similar to store, but increments value in memory */
 112 static gmx_inline void gmx_simdcall
 113 gmx_mm_increment_2real_swizzle_pd(double * gmx_restrict ptrA,
 114                                   double * gmx_restrict ptrB, __m128d xmm1)
 115 {
 116     __m128d t1;
 117
 118     t1   = _mm_unpackhi_pd(xmm1, xmm1);
 119     xmm1 = _mm_add_sd(xmm1, _mm_load_sd(ptrA));
 120     t1   = _mm_add_sd(t1, _mm_load_sd(ptrB));
 121     _mm_store_sd(ptrA, xmm1);
 122     _mm_store_sd(ptrB, t1);
 123 }
 124
 125 static gmx_inline void gmx_simdcall
 126 gmx_mm_increment_1real_pd(double * gmx_restrict ptrA, __m128d xmm1)
 127 {
 128     __m128d tmp;
 129
 130     tmp = gmx_mm_load_1real_pd(ptrA);
 131     tmp = _mm_add_sd(tmp, xmm1);
 132     gmx_mm_store_1real_pd(ptrA, tmp);
 133 }
 134
 135
 136
 137 static gmx_inline void gmx_simdcall
 138 gmx_mm_load_2pair_swizzle_pd(const double * gmx_restrict p1,
 139                              const double * gmx_restrict p2,
 140                              __m128d * gmx_restrict      c6,
 141                              __m128d * gmx_restrict      c12)
 142 {
 143     __m128d t1, t2, t3;
 144
 145     /* The c6/c12 array should be aligned */
 146     t1   = _mm_loadu_pd(p1);
 147     t2   = _mm_loadu_pd(p2);
 148     *c6  = _mm_unpacklo_pd(t1, t2);
 149     *c12 = _mm_unpackhi_pd(t1, t2);
 150 }
 151
 152 static gmx_inline void gmx_simdcall
 153 gmx_mm_load_1pair_swizzle_pd(const double * gmx_restrict p1,
 154                              __m128d * gmx_restrict      c6,
 155                              __m128d * gmx_restrict      c12)
 156 {
 157     *c6     = _mm_load_sd(p1);
 158     *c12    = _mm_load_sd(p1+1);
 159 }
 160
 161
 162 static gmx_inline void gmx_simdcall
 163 gmx_mm_load_shift_and_1rvec_broadcast_pd(const double * gmx_restrict xyz_shift,
 164                                          const double * gmx_restrict xyz,
 165                                          __m128d * gmx_restrict      x1,
 166                                          __m128d * gmx_restrict      y1,
 167                                          __m128d * gmx_restrict      z1)
 168 {
 169     __m128d mem_xy, mem_z, mem_sxy, mem_sz;
 170
 171     mem_xy  = _mm_loadu_pd(xyz);
 172     mem_z   = _mm_load_sd(xyz+2);
 173     mem_sxy = _mm_loadu_pd(xyz_shift);
 174     mem_sz  = _mm_load_sd(xyz_shift+2);
 175
 176     mem_xy  = _mm_add_pd(mem_xy, mem_sxy);
 177     mem_z   = _mm_add_pd(mem_z, mem_sz);
 178
 179     *x1  = _mm_shuffle_pd(mem_xy, mem_xy, _MM_SHUFFLE2(0, 0));
 180     *y1  = _mm_shuffle_pd(mem_xy, mem_xy, _MM_SHUFFLE2(1, 1));
 181     *z1  = _mm_shuffle_pd(mem_z, mem_z, _MM_SHUFFLE2(0, 0));
 182 }
 183
 184
 185 static gmx_inline void gmx_simdcall
 186 gmx_mm_load_shift_and_3rvec_broadcast_pd(const double * gmx_restrict xyz_shift,
 187                                          const double * gmx_restrict xyz,
 188                                          __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 189                                          __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 190                                          __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3)
 191 {
 192     __m128d t1, t2, t3, t4, t5, sxy, sz, szx, syz;
 193
 194     t1  = _mm_loadu_pd(xyz);
 195     t2  = _mm_loadu_pd(xyz+2);
 196     t3  = _mm_loadu_pd(xyz+4);
 197     t4  = _mm_loadu_pd(xyz+6);
 198     t5  = _mm_load_sd(xyz+8);
 199
 200     sxy = _mm_loadu_pd(xyz_shift);
 201     sz  = _mm_load_sd(xyz_shift+2);
 202     szx = _mm_shuffle_pd(sz, sxy, _MM_SHUFFLE2(0, 0));
 203     syz = _mm_shuffle_pd(sxy, sz, _MM_SHUFFLE2(0, 1));
 204
 205     t1  = _mm_add_pd(t1, sxy);
 206     t2  = _mm_add_pd(t2, szx);
 207     t3  = _mm_add_pd(t3, syz);
 208     t4  = _mm_add_pd(t4, sxy);
 209     t5  = _mm_add_sd(t5, sz);
 210
 211     *x1  = _mm_shuffle_pd(t1, t1, _MM_SHUFFLE2(0, 0));
 212     *y1  = _mm_shuffle_pd(t1, t1, _MM_SHUFFLE2(1, 1));
 213     *z1  = _mm_shuffle_pd(t2, t2, _MM_SHUFFLE2(0, 0));
 214     *x2  = _mm_shuffle_pd(t2, t2, _MM_SHUFFLE2(1, 1));
 215     *y2  = _mm_shuffle_pd(t3, t3, _MM_SHUFFLE2(0, 0));
 216     *z2  = _mm_shuffle_pd(t3, t3, _MM_SHUFFLE2(1, 1));
 217     *x3  = _mm_shuffle_pd(t4, t4, _MM_SHUFFLE2(0, 0));
 218     *y3  = _mm_shuffle_pd(t4, t4, _MM_SHUFFLE2(1, 1));
 219     *z3  = _mm_shuffle_pd(t5, t5, _MM_SHUFFLE2(0, 0));
 220 }
 221
 222
 223 static gmx_inline void gmx_simdcall
 224 gmx_mm_load_shift_and_4rvec_broadcast_pd(const double * gmx_restrict xyz_shift,
 225                                          const double * gmx_restrict xyz,
 226                                          __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 227                                          __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 228                                          __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3,
 229                                          __m128d * gmx_restrict x4, __m128d * gmx_restrict y4, __m128d * gmx_restrict z4)
 230 {
 231     __m128d t1, t2, t3, t4, t5, t6, sxy, sz, szx, syz;
 232
 233     t1  = _mm_loadu_pd(xyz);
 234     t2  = _mm_loadu_pd(xyz+2);
 235     t3  = _mm_loadu_pd(xyz+4);
 236     t4  = _mm_loadu_pd(xyz+6);
 237     t5  = _mm_loadu_pd(xyz+8);
 238     t6  = _mm_loadu_pd(xyz+10);
 239
 240     sxy = _mm_loadu_pd(xyz_shift);
 241     sz  = _mm_load_sd(xyz_shift+2);
 242     szx = _mm_shuffle_pd(sz, sxy, _MM_SHUFFLE2(0, 0));
 243     syz = _mm_shuffle_pd(sxy, sz, _MM_SHUFFLE2(0, 1));
 244
 245     t1  = _mm_add_pd(t1, sxy);
 246     t2  = _mm_add_pd(t2, szx);
 247     t3  = _mm_add_pd(t3, syz);
 248     t4  = _mm_add_pd(t4, sxy);
 249     t5  = _mm_add_pd(t5, szx);
 250     t6  = _mm_add_pd(t6, syz);
 251
 252     *x1  = _mm_shuffle_pd(t1, t1, _MM_SHUFFLE2(0, 0));
 253     *y1  = _mm_shuffle_pd(t1, t1, _MM_SHUFFLE2(1, 1));
 254     *z1  = _mm_shuffle_pd(t2, t2, _MM_SHUFFLE2(0, 0));
 255     *x2  = _mm_shuffle_pd(t2, t2, _MM_SHUFFLE2(1, 1));
 256     *y2  = _mm_shuffle_pd(t3, t3, _MM_SHUFFLE2(0, 0));
 257     *z2  = _mm_shuffle_pd(t3, t3, _MM_SHUFFLE2(1, 1));
 258     *x3  = _mm_shuffle_pd(t4, t4, _MM_SHUFFLE2(0, 0));
 259     *y3  = _mm_shuffle_pd(t4, t4, _MM_SHUFFLE2(1, 1));
 260     *z3  = _mm_shuffle_pd(t5, t5, _MM_SHUFFLE2(0, 0));
 261     *x4  = _mm_shuffle_pd(t5, t5, _MM_SHUFFLE2(1, 1));
 262     *y4  = _mm_shuffle_pd(t6, t6, _MM_SHUFFLE2(0, 0));
 263     *z4  = _mm_shuffle_pd(t6, t6, _MM_SHUFFLE2(1, 1));
 264 }
 265
 266
 267
 268 static gmx_inline void gmx_simdcall
 269 gmx_mm_load_1rvec_1ptr_swizzle_pd(const double * gmx_restrict p1,
 270                                   __m128d * gmx_restrict x, __m128d * gmx_restrict y, __m128d * gmx_restrict z)
 271 {
 272     *x            = _mm_load_sd(p1);
 273     *y            = _mm_load_sd(p1+1);
 274     *z            = _mm_load_sd(p1+2);
 275 }
 276
 277 static gmx_inline void gmx_simdcall
 278 gmx_mm_load_3rvec_1ptr_swizzle_pd(const double * gmx_restrict p1,
 279                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 280                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 281                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3)
 282 {
 283     *x1            = _mm_load_sd(p1);
 284     *y1            = _mm_load_sd(p1+1);
 285     *z1            = _mm_load_sd(p1+2);
 286     *x2            = _mm_load_sd(p1+3);
 287     *y2            = _mm_load_sd(p1+4);
 288     *z2            = _mm_load_sd(p1+5);
 289     *x3            = _mm_load_sd(p1+6);
 290     *y3            = _mm_load_sd(p1+7);
 291     *z3            = _mm_load_sd(p1+8);
 292 }
 293
 294 static gmx_inline void gmx_simdcall
 295 gmx_mm_load_4rvec_1ptr_swizzle_pd(const double * gmx_restrict p1,
 296                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 297                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 298                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3,
 299                                   __m128d * gmx_restrict x4, __m128d * gmx_restrict y4, __m128d * gmx_restrict z4)
 300 {
 301     *x1            = _mm_load_sd(p1);
 302     *y1            = _mm_load_sd(p1+1);
 303     *z1            = _mm_load_sd(p1+2);
 304     *x2            = _mm_load_sd(p1+3);
 305     *y2            = _mm_load_sd(p1+4);
 306     *z2            = _mm_load_sd(p1+5);
 307     *x3            = _mm_load_sd(p1+6);
 308     *y3            = _mm_load_sd(p1+7);
 309     *z3            = _mm_load_sd(p1+8);
 310     *x4            = _mm_load_sd(p1+9);
 311     *y4            = _mm_load_sd(p1+10);
 312     *z4            = _mm_load_sd(p1+11);
 313 }
 314
 315
 316 static gmx_inline void gmx_simdcall
 317 gmx_mm_load_1rvec_2ptr_swizzle_pd(const double * gmx_restrict ptrA,
 318                                   const double * gmx_restrict ptrB,
 319                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1)
 320 {
 321     __m128d t1, t2, t3, t4;
 322     t1           = _mm_loadu_pd(ptrA);
 323     t2           = _mm_loadu_pd(ptrB);
 324     t3           = _mm_load_sd(ptrA+2);
 325     t4           = _mm_load_sd(ptrB+2);
 326     GMX_MM_TRANSPOSE2_PD(t1, t2);
 327     *x1          = t1;
 328     *y1          = t2;
 329     *z1          = _mm_unpacklo_pd(t3, t4);
 330 }
 331
 332 static gmx_inline void gmx_simdcall
 333 gmx_mm_load_3rvec_2ptr_swizzle_pd(const double * gmx_restrict ptrA, const double * gmx_restrict ptrB,
 334                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 335                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 336                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3)
 337 {
 338     __m128d t1, t2, t3, t4, t5, t6, t7, t8, t9, t10;
 339     t1           = _mm_loadu_pd(ptrA);
 340     t2           = _mm_loadu_pd(ptrB);
 341     t3           = _mm_loadu_pd(ptrA+2);
 342     t4           = _mm_loadu_pd(ptrB+2);
 343     t5           = _mm_loadu_pd(ptrA+4);
 344     t6           = _mm_loadu_pd(ptrB+4);
 345     t7           = _mm_loadu_pd(ptrA+6);
 346     t8           = _mm_loadu_pd(ptrB+6);
 347     t9           = _mm_load_sd(ptrA+8);
 348     t10          = _mm_load_sd(ptrB+8);
 349     GMX_MM_TRANSPOSE2_PD(t1, t2);
 350     GMX_MM_TRANSPOSE2_PD(t3, t4);
 351     GMX_MM_TRANSPOSE2_PD(t5, t6);
 352     GMX_MM_TRANSPOSE2_PD(t7, t8);
 353     *x1          = t1;
 354     *y1          = t2;
 355     *z1          = t3;
 356     *x2          = t4;
 357     *y2          = t5;
 358     *z2          = t6;
 359     *x3          = t7;
 360     *y3          = t8;
 361     *z3          = _mm_unpacklo_pd(t9, t10);
 362 }
 363
 364
 365 static gmx_inline void gmx_simdcall
 366 gmx_mm_load_4rvec_2ptr_swizzle_pd(const double * gmx_restrict ptrA, const double * gmx_restrict ptrB,
 367                                   __m128d * gmx_restrict x1, __m128d * gmx_restrict y1, __m128d * gmx_restrict z1,
 368                                   __m128d * gmx_restrict x2, __m128d * gmx_restrict y2, __m128d * gmx_restrict z2,
 369                                   __m128d * gmx_restrict x3, __m128d * gmx_restrict y3, __m128d * gmx_restrict z3,
 370                                   __m128d * gmx_restrict x4, __m128d * gmx_restrict y4, __m128d * gmx_restrict z4)
 371 {
 372     __m128d t1, t2, t3, t4, t5, t6;
 373     t1           = _mm_loadu_pd(ptrA);
 374     t2           = _mm_loadu_pd(ptrB);
 375     t3           = _mm_loadu_pd(ptrA+2);
 376     t4           = _mm_loadu_pd(ptrB+2);
 377     t5           = _mm_loadu_pd(ptrA+4);
 378     t6           = _mm_loadu_pd(ptrB+4);
 379     GMX_MM_TRANSPOSE2_PD(t1, t2);
 380     GMX_MM_TRANSPOSE2_PD(t3, t4);
 381     GMX_MM_TRANSPOSE2_PD(t5, t6);
 382     *x1          = t1;
 383     *y1          = t2;
 384     *z1          = t3;
 385     *x2          = t4;
 386     *y2          = t5;
 387     *z2          = t6;
 388     t1           = _mm_loadu_pd(ptrA+6);
 389     t2           = _mm_loadu_pd(ptrB+6);
 390     t3           = _mm_loadu_pd(ptrA+8);
 391     t4           = _mm_loadu_pd(ptrB+8);
 392     t5           = _mm_loadu_pd(ptrA+10);
 393     t6           = _mm_loadu_pd(ptrB+10);
 394     GMX_MM_TRANSPOSE2_PD(t1, t2);
 395     GMX_MM_TRANSPOSE2_PD(t3, t4);
 396     GMX_MM_TRANSPOSE2_PD(t5, t6);
 397     *x3          = t1;
 398     *y3          = t2;
 399     *z3          = t3;
 400     *x4          = t4;
 401     *y4          = t5;
 402     *z4          = t6;
 403 }
 404
 405
 406 /* Routines to decrement rvec in memory, typically use for j particle force updates */
 407 static gmx_inline void gmx_simdcall
 408 gmx_mm_decrement_1rvec_1ptr_swizzle_pd(double * gmx_restrict ptrA,
 409                                        __m128d x1, __m128d y1, __m128d z1)
 410 {
 411     __m128d t1, t2, t3;
 412
 413     t1           = _mm_load_sd(ptrA);
 414     t2           = _mm_load_sd(ptrA+1);
 415     t3           = _mm_load_sd(ptrA+2);
 416
 417     t1           = _mm_sub_sd(t1, x1);
 418     t2           = _mm_sub_sd(t2, y1);
 419     t3           = _mm_sub_sd(t3, z1);
 420     _mm_store_sd(ptrA, t1);
 421     _mm_store_sd(ptrA+1, t2);
 422     _mm_store_sd(ptrA+2, t3);
 423 }
 424
 425
 426
 427 static gmx_inline void gmx_simdcall
 428 gmx_mm_decrement_3rvec_1ptr_swizzle_pd(double * gmx_restrict ptrA,
 429                                        __m128d x1, __m128d y1, __m128d z1,
 430                                        __m128d x2, __m128d y2, __m128d z2,
 431                                        __m128d x3, __m128d y3, __m128d z3)
 432 {
 433     __m128d t1, t2, t3, t4, t5;
 434
 435     t1          = _mm_loadu_pd(ptrA);
 436     t2          = _mm_loadu_pd(ptrA+2);
 437     t3          = _mm_loadu_pd(ptrA+4);
 438     t4          = _mm_loadu_pd(ptrA+6);
 439     t5          = _mm_load_sd(ptrA+8);
 440
 441     x1          = _mm_unpacklo_pd(x1, y1);
 442     z1          = _mm_unpacklo_pd(z1, x2);
 443     y2          = _mm_unpacklo_pd(y2, z2);
 444     x3          = _mm_unpacklo_pd(x3, y3);
 445     /* nothing to be done for z3 */
 446
 447     t1          = _mm_sub_pd(t1, x1);
 448     t2          = _mm_sub_pd(t2, z1);
 449     t3          = _mm_sub_pd(t3, y2);
 450     t4          = _mm_sub_pd(t4, x3);
 451     t5          = _mm_sub_sd(t5, z3);
 452     _mm_storeu_pd(ptrA, t1);
 453     _mm_storeu_pd(ptrA+2, t2);
 454     _mm_storeu_pd(ptrA+4, t3);
 455     _mm_storeu_pd(ptrA+6, t4);
 456     _mm_store_sd(ptrA+8, t5);
 457 }
 458
 459
 460 static gmx_inline void gmx_simdcall
 461 gmx_mm_decrement_4rvec_1ptr_swizzle_pd(double * gmx_restrict ptrA,
 462                                        __m128d x1, __m128d y1, __m128d z1,
 463                                        __m128d x2, __m128d y2, __m128d z2,
 464                                        __m128d x3, __m128d y3, __m128d z3,
 465                                        __m128d x4, __m128d y4, __m128d z4)
 466 {
 467     __m128d t1, t2, t3, t4, t5, t6;
 468
 469     t1          = _mm_loadu_pd(ptrA);
 470     t2          = _mm_loadu_pd(ptrA+2);
 471     t3          = _mm_loadu_pd(ptrA+4);
 472     t4          = _mm_loadu_pd(ptrA+6);
 473     t5          = _mm_loadu_pd(ptrA+8);
 474     t6          = _mm_loadu_pd(ptrA+10);
 475
 476     x1          = _mm_unpacklo_pd(x1, y1);
 477     z1          = _mm_unpacklo_pd(z1, x2);
 478     y2          = _mm_unpacklo_pd(y2, z2);
 479     x3          = _mm_unpacklo_pd(x3, y3);
 480     z3          = _mm_unpacklo_pd(z3, x4);
 481     y4          = _mm_unpacklo_pd(y4, z4);
 482
 483     _mm_storeu_pd(ptrA,    _mm_sub_pd( t1, x1 ));
 484     _mm_storeu_pd(ptrA+2,  _mm_sub_pd( t2, z1 ));
 485     _mm_storeu_pd(ptrA+4,  _mm_sub_pd( t3, y2 ));
 486     _mm_storeu_pd(ptrA+6,  _mm_sub_pd( t4, x3 ));
 487     _mm_storeu_pd(ptrA+8,  _mm_sub_pd( t5, z3 ));
 488     _mm_storeu_pd(ptrA+10, _mm_sub_pd( t6, y4 ));
 489 }
 490
 491
 492 static gmx_inline void gmx_simdcall
 493 gmx_mm_decrement_1rvec_2ptr_swizzle_pd(double * gmx_restrict ptrA, double * gmx_restrict ptrB,
 494                                        __m128d x1, __m128d y1, __m128d z1)
 495 {
 496     __m128d t1, t2, t3, t4, t5, t6, t7;
 497
 498     t1          = _mm_loadu_pd(ptrA);
 499     t2          = _mm_load_sd(ptrA+2);
 500     t3          = _mm_loadu_pd(ptrB);
 501     t4          = _mm_load_sd(ptrB+2);
 502
 503     t5          = _mm_unpacklo_pd(x1, y1);
 504     t6          = _mm_unpackhi_pd(x1, y1);
 505     t7          = _mm_unpackhi_pd(z1, z1);
 506
 507     t1          = _mm_sub_pd(t1, t5);
 508     t2          = _mm_sub_sd(t2, z1);
 509
 510     t3          = _mm_sub_pd(t3, t6);
 511     t4          = _mm_sub_sd(t4, t7);
 512
 513     _mm_storeu_pd(ptrA, t1);
 514     _mm_store_sd(ptrA+2, t2);
 515     _mm_storeu_pd(ptrB, t3);
 516     _mm_store_sd(ptrB+2, t4);
 517 }
 518
 519
 520
 521 static gmx_inline void gmx_simdcall
 522 gmx_mm_decrement_3rvec_2ptr_swizzle_pd(double * gmx_restrict ptrA, double * gmx_restrict ptrB,
 523                                        __m128d x1, __m128d y1, __m128d z1,
 524                                        __m128d x2, __m128d y2, __m128d z2,
 525                                        __m128d x3, __m128d y3, __m128d z3)
 526 {
 527     __m128d t1, t2, t3, t4, t5, t6, t7, t8, t9, t10;
 528     __m128d tA, tB, tC, tD, tE, tF, tG, tH, tI;
 529
 530     t1          = _mm_loadu_pd(ptrA);
 531     t2          = _mm_loadu_pd(ptrA+2);
 532     t3          = _mm_loadu_pd(ptrA+4);
 533     t4          = _mm_loadu_pd(ptrA+6);
 534     t5          = _mm_load_sd(ptrA+8);
 535     t6          = _mm_loadu_pd(ptrB);
 536     t7          = _mm_loadu_pd(ptrB+2);
 537     t8          = _mm_loadu_pd(ptrB+4);
 538     t9          = _mm_loadu_pd(ptrB+6);
 539     t10         = _mm_load_sd(ptrB+8);
 540
 541     tA          = _mm_unpacklo_pd(x1, y1);
 542     tB          = _mm_unpackhi_pd(x1, y1);
 543     tC          = _mm_unpacklo_pd(z1, x2);
 544     tD          = _mm_unpackhi_pd(z1, x2);
 545     tE          = _mm_unpacklo_pd(y2, z2);
 546     tF          = _mm_unpackhi_pd(y2, z2);
 547     tG          = _mm_unpacklo_pd(x3, y3);
 548     tH          = _mm_unpackhi_pd(x3, y3);
 549     tI          = _mm_unpackhi_pd(z3, z3);
 550
 551     t1          = _mm_sub_pd(t1, tA);
 552     t2          = _mm_sub_pd(t2, tC);
 553     t3          = _mm_sub_pd(t3, tE);
 554     t4          = _mm_sub_pd(t4, tG);
 555     t5          = _mm_sub_sd(t5, z3);
 556
 557     t6          = _mm_sub_pd(t6, tB);
 558     t7          = _mm_sub_pd(t7, tD);
 559     t8          = _mm_sub_pd(t8, tF);
 560     t9          = _mm_sub_pd(t9, tH);
 561     t10         = _mm_sub_sd(t10, tI);
 562
 563     _mm_storeu_pd(ptrA, t1);
 564     _mm_storeu_pd(ptrA+2, t2);
 565     _mm_storeu_pd(ptrA+4, t3);
 566     _mm_storeu_pd(ptrA+6, t4);
 567     _mm_store_sd(ptrA+8, t5);
 568     _mm_storeu_pd(ptrB, t6);
 569     _mm_storeu_pd(ptrB+2, t7);
 570     _mm_storeu_pd(ptrB+4, t8);
 571     _mm_storeu_pd(ptrB+6, t9);
 572     _mm_store_sd(ptrB+8, t10);
 573 }
 574
 575
 576 static gmx_inline void gmx_simdcall
 577 gmx_mm_decrement_4rvec_2ptr_swizzle_pd(double * gmx_restrict ptrA, double * gmx_restrict ptrB,
 578                                        __m128d x1, __m128d y1, __m128d z1,
 579                                        __m128d x2, __m128d y2, __m128d z2,
 580                                        __m128d x3, __m128d y3, __m128d z3,
 581                                        __m128d x4, __m128d y4, __m128d z4)
 582 {
 583     __m128d t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12;
 584     __m128d tA, tB, tC, tD, tE, tF, tG, tH, tI, tJ, tK, tL;
 585
 586     t1          = _mm_loadu_pd(ptrA);
 587     t2          = _mm_loadu_pd(ptrA+2);
 588     t3          = _mm_loadu_pd(ptrA+4);
 589     t4          = _mm_loadu_pd(ptrA+6);
 590     t5          = _mm_loadu_pd(ptrA+8);
 591     t6          = _mm_loadu_pd(ptrA+10);
 592     t7          = _mm_loadu_pd(ptrB);
 593     t8          = _mm_loadu_pd(ptrB+2);
 594     t9          = _mm_loadu_pd(ptrB+4);
 595     t10         = _mm_loadu_pd(ptrB+6);
 596     t11         = _mm_loadu_pd(ptrB+8);
 597     t12         = _mm_loadu_pd(ptrB+10);
 598
 599     tA          = _mm_unpacklo_pd(x1, y1);
 600     tB          = _mm_unpackhi_pd(x1, y1);
 601     tC          = _mm_unpacklo_pd(z1, x2);
 602     tD          = _mm_unpackhi_pd(z1, x2);
 603     tE          = _mm_unpacklo_pd(y2, z2);
 604     tF          = _mm_unpackhi_pd(y2, z2);
 605     tG          = _mm_unpacklo_pd(x3, y3);
 606     tH          = _mm_unpackhi_pd(x3, y3);
 607     tI          = _mm_unpacklo_pd(z3, x4);
 608     tJ          = _mm_unpackhi_pd(z3, x4);
 609     tK          = _mm_unpacklo_pd(y4, z4);
 610     tL          = _mm_unpackhi_pd(y4, z4);
 611
 612     t1          = _mm_sub_pd(t1, tA);
 613     t2          = _mm_sub_pd(t2, tC);
 614     t3          = _mm_sub_pd(t3, tE);
 615     t4          = _mm_sub_pd(t4, tG);
 616     t5          = _mm_sub_pd(t5, tI);
 617     t6          = _mm_sub_pd(t6, tK);
 618
 619     t7          = _mm_sub_pd(t7, tB);
 620     t8          = _mm_sub_pd(t8, tD);
 621     t9          = _mm_sub_pd(t9, tF);
 622     t10         = _mm_sub_pd(t10, tH);
 623     t11         = _mm_sub_pd(t11, tJ);
 624     t12         = _mm_sub_pd(t12, tL);
 625
 626     _mm_storeu_pd(ptrA,  t1);
 627     _mm_storeu_pd(ptrA+2, t2);
 628     _mm_storeu_pd(ptrA+4, t3);
 629     _mm_storeu_pd(ptrA+6, t4);
 630     _mm_storeu_pd(ptrA+8, t5);
 631     _mm_storeu_pd(ptrA+10, t6);
 632     _mm_storeu_pd(ptrB,  t7);
 633     _mm_storeu_pd(ptrB+2, t8);
 634     _mm_storeu_pd(ptrB+4, t9);
 635     _mm_storeu_pd(ptrB+6, t10);
 636     _mm_storeu_pd(ptrB+8, t11);
 637     _mm_storeu_pd(ptrB+10, t12);
 638 }
 639
 640
 641 static gmx_inline void gmx_simdcall
 642 gmx_mm_update_iforce_1atom_swizzle_pd(__m128d fix1, __m128d fiy1, __m128d fiz1,
 643                                       double * gmx_restrict fptr,
 644                                       double * gmx_restrict fshiftptr)
 645 {
 646     fix1 = _mm_hadd_pd(fix1, fiy1);
 647     fiz1 = _mm_hadd_pd(fiz1, fiz1);
 648
 649     _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr), fix1 ));
 650     _mm_store_sd( fptr+2, _mm_add_sd( _mm_load_sd(fptr+2), fiz1 ));
 651
 652     _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 ));
 653     _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 ));
 654 }
 655
 656
 657 static gmx_inline void gmx_simdcall
 658 gmx_mm_update_iforce_3atom_swizzle_pd(__m128d fix1, __m128d fiy1, __m128d fiz1,
 659                                       __m128d fix2, __m128d fiy2, __m128d fiz2,
 660                                       __m128d fix3, __m128d fiy3, __m128d fiz3,
 661                                       double * gmx_restrict fptr,
 662                                       double * gmx_restrict fshiftptr)
 663 {
 664     __m128d t1, t2;
 665
 666     fix1 = _mm_hadd_pd(fix1, fiy1);
 667     fiz1 = _mm_hadd_pd(fiz1, fix2);
 668     fiy2 = _mm_hadd_pd(fiy2, fiz2);
 669     fix3 = _mm_hadd_pd(fix3, fiy3);
 670     fiz3 = _mm_hadd_pd(fiz3, fiz3);
 671
 672     _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr), fix1 ));
 673     _mm_storeu_pd( fptr+2, _mm_add_pd( _mm_loadu_pd(fptr+2), fiz1 ));
 674     _mm_storeu_pd( fptr+4, _mm_add_pd( _mm_loadu_pd(fptr+4), fiy2 ));
 675     _mm_storeu_pd( fptr+6, _mm_add_pd( _mm_loadu_pd(fptr+6), fix3 ));
 676     _mm_store_sd( fptr+8, _mm_add_sd( _mm_load_sd(fptr+8), fiz3 ));
 677
 678     fix1 = _mm_add_pd(fix1, fix3);
 679     t1   = _mm_shuffle_pd(fiz1, fiy2, _MM_SHUFFLE2(0, 1));
 680     fix1 = _mm_add_pd(fix1, t1); /* x and y sums */
 681
 682     t2   = _mm_shuffle_pd(fiy2, fiy2, _MM_SHUFFLE2(1, 1));
 683     fiz1 = _mm_add_sd(fiz1, fiz3);
 684     fiz1 = _mm_add_sd(fiz1, t2); /* z sum */
 685
 686     _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 ));
 687     _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 ));
 688 }
 689
 690
 691 static gmx_inline void gmx_simdcall
 692 gmx_mm_update_iforce_4atom_swizzle_pd(__m128d fix1, __m128d fiy1, __m128d fiz1,
 693                                       __m128d fix2, __m128d fiy2, __m128d fiz2,
 694                                       __m128d fix3, __m128d fiy3, __m128d fiz3,
 695                                       __m128d fix4, __m128d fiy4, __m128d fiz4,
 696                                       double * gmx_restrict fptr,
 697                                       double * gmx_restrict fshiftptr)
 698 {
 699     __m128d t1, t2;
 700
 701     fix1 = _mm_hadd_pd(fix1, fiy1);
 702     fiz1 = _mm_hadd_pd(fiz1, fix2);
 703     fiy2 = _mm_hadd_pd(fiy2, fiz2);
 704     fix3 = _mm_hadd_pd(fix3, fiy3);
 705     fiz3 = _mm_hadd_pd(fiz3, fix4);
 706     fiy4 = _mm_hadd_pd(fiy4, fiz4);
 707
 708     _mm_storeu_pd( fptr, _mm_add_pd( _mm_loadu_pd(fptr),       fix1 ));
 709     _mm_storeu_pd( fptr+2, _mm_add_pd( _mm_loadu_pd(fptr+2),   fiz1 ));
 710     _mm_storeu_pd( fptr+4, _mm_add_pd( _mm_loadu_pd(fptr+4),   fiy2 ));
 711     _mm_storeu_pd( fptr+6, _mm_add_pd( _mm_loadu_pd(fptr+6),   fix3 ));
 712     _mm_storeu_pd( fptr+8, _mm_add_pd( _mm_loadu_pd(fptr+8),   fiz3 ));
 713     _mm_storeu_pd( fptr+10, _mm_add_pd( _mm_loadu_pd(fptr+10), fiy4 ));
 714
 715     t1   = _mm_shuffle_pd(fiz1, fiy2, _MM_SHUFFLE2(0, 1));
 716     fix1 = _mm_add_pd(fix1, t1);
 717     t2   = _mm_shuffle_pd(fiz3, fiy4, _MM_SHUFFLE2(0, 1));
 718     fix3 = _mm_add_pd(fix3, t2);
 719     fix1 = _mm_add_pd(fix1, fix3); /* x and y sums */
 720
 721     fiz1 = _mm_add_sd(fiz1, _mm_unpackhi_pd(fiy2, fiy2));
 722     fiz3 = _mm_add_sd(fiz3, _mm_unpackhi_pd(fiy4, fiy4));
 723     fiz1 = _mm_add_sd(fiz1, fiz3); /* z sum */
 724
 725     _mm_storeu_pd( fshiftptr, _mm_add_pd( _mm_loadu_pd(fshiftptr), fix1 ));
 726     _mm_store_sd( fshiftptr+2, _mm_add_sd( _mm_load_sd(fshiftptr+2), fiz1 ));
 727 }
 728
 729
 730 static gmx_inline void gmx_simdcall
 731 gmx_mm_update_1pot_pd(__m128d pot1, double * gmx_restrict ptrA)
 732 {
 733     pot1 = _mm_hadd_pd(pot1, pot1);
 734     _mm_store_sd(ptrA, _mm_add_sd(pot1, _mm_load_sd(ptrA)));
 735 }
 736
 737 static gmx_inline void gmx_simdcall
 738 gmx_mm_update_2pot_pd(__m128d pot1, double * gmx_restrict ptrA,
 739                       __m128d pot2, double * gmx_restrict ptrB)
 740 {
 741     pot1 = _mm_hadd_pd(pot1, pot2);
 742     pot2 = _mm_unpackhi_pd(pot1, pot1);
 743
 744     _mm_store_sd(ptrA, _mm_add_sd(pot1, _mm_load_sd(ptrA)));
 745     _mm_store_sd(ptrB, _mm_add_sd(pot2, _mm_load_sd(ptrB)));
 746 }
 747
 748
 749 #endif /* _kernelutil_x86_avx_128_fma_double_h_ */