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