src/gromacs/simd/impl_x86_avx_512/impl_x86_avx_512_util_float.h

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2014,2015,2016,2017,2018,2019, 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
  36 #ifndef GMX_SIMD_IMPL_X86_AVX_512_UTIL_FLOAT_H
  37 #define GMX_SIMD_IMPL_X86_AVX_512_UTIL_FLOAT_H
  38
  39 #include "config.h"
  40
  41 #include <cassert>
  42 #include <cstdint>
  43
  44 #include <immintrin.h>
  45
  46 #include "gromacs/utility/basedefinitions.h"
  47
  48 #include "impl_x86_avx_512_general.h"
  49 #include "impl_x86_avx_512_simd_float.h"
  50
  51 namespace gmx
  52 {
  53
  54 static const int c_simdBestPairAlignmentFloat = 2;
  55
  56 namespace
  57 {
  58 // Multiply function optimized for powers of 2, for which it is done by
  59 // shifting. Currently up to 8 is accelerated. Could be accelerated for any
  60 // number with a constexpr log2 function.
  61 template<int n>
  62 SimdFInt32 fastMultiply(SimdFInt32 x)
  63 {
  64     if (n == 2)
  65     {
  66         return _mm512_slli_epi32(x.simdInternal_, 1);
  67     }
  68     else if (n == 4)
  69     {
  70         return _mm512_slli_epi32(x.simdInternal_, 2);
  71     }
  72     else if (n == 8)
  73     {
  74         return _mm512_slli_epi32(x.simdInternal_, 3);
  75     }
  76     else
  77     {
  78         return x * n;
  79     }
  80 }
  81
  82 template<int align>
  83 static inline void gmx_simdcall gatherLoadBySimdIntTranspose(const float*, SimdFInt32)
  84 {
  85     // Nothing to do. Termination of recursion.
  86 }
  87 } // namespace
  88
  89 template<int align, typename... Targs>
  90 static inline void gmx_simdcall
  91                    gatherLoadBySimdIntTranspose(const float* base, SimdFInt32 offset, SimdFloat* v, Targs... Fargs)
  92 {
  93     // For align 1 or 2: No multiplication of offset is needed
  94     if (align > 2)
  95     {
  96         offset = fastMultiply<align>(offset);
  97     }
  98     // For align 2: Scale of 2*sizeof(float) is used (maximum supported scale)
  99     constexpr int align_ = (align > 2) ? 1 : align;
 100     v->simdInternal_     = _mm512_i32gather_ps(offset.simdInternal_, base, sizeof(float) * align_);
 101     // Gather remaining elements. Avoid extra multiplication (new align is 1 or 2).
 102     gatherLoadBySimdIntTranspose<align_>(base + 1, offset, Fargs...);
 103 }
 104
 105 template<int align, typename... Targs>
 106 static inline void gmx_simdcall
 107                    gatherLoadUBySimdIntTranspose(const float* base, SimdFInt32 offset, SimdFloat* v, Targs... Fargs)
 108 {
 109     gatherLoadBySimdIntTranspose<align>(base, offset, v, Fargs...);
 110 }
 111
 112 template<int align, typename... Targs>
 113 static inline void gmx_simdcall
 114                    gatherLoadTranspose(const float* base, const std::int32_t offset[], SimdFloat* v, Targs... Fargs)
 115 {
 116     gatherLoadBySimdIntTranspose<align>(base, simdLoad(offset, SimdFInt32Tag()), v, Fargs...);
 117 }
 118
 119 template<int align, typename... Targs>
 120 static inline void gmx_simdcall
 121                    gatherLoadUTranspose(const float* base, const std::int32_t offset[], SimdFloat* v, Targs... Fargs)
 122 {
 123     gatherLoadBySimdIntTranspose<align>(base, simdLoad(offset, SimdFInt32Tag()), v, Fargs...);
 124 }
 125
 126 template<int align>
 127 static inline void gmx_simdcall
 128                    transposeScatterStoreU(float* base, const std::int32_t offset[], SimdFloat v0, SimdFloat v1, SimdFloat v2)
 129 {
 130     SimdFInt32 simdoffset = simdLoad(offset, SimdFInt32Tag());
 131     if (align > 2)
 132     {
 133         simdoffset = fastMultiply<align>(simdoffset);
 134     }
 135     constexpr size_t scale = (align > 2) ? sizeof(float) : sizeof(float) * align;
 136
 137     _mm512_i32scatter_ps(base, simdoffset.simdInternal_, v0.simdInternal_, scale);
 138     _mm512_i32scatter_ps(&(base[1]), simdoffset.simdInternal_, v1.simdInternal_, scale);
 139     _mm512_i32scatter_ps(&(base[2]), simdoffset.simdInternal_, v2.simdInternal_, scale);
 140 }
 141
 142 template<int align>
 143 static inline void gmx_simdcall
 144                    transposeScatterIncrU(float* base, const std::int32_t offset[], SimdFloat v0, SimdFloat v1, SimdFloat v2)
 145 {
 146     __m512                                   t[4], t5, t6, t7, t8;
 147     int                                      i;
 148     alignas(GMX_SIMD_ALIGNMENT) std::int32_t o[16];
 149     store(o, fastMultiply<align>(simdLoad(offset, SimdFInt32Tag())));
 150     if (align < 4)
 151     {
 152         t5   = _mm512_unpacklo_ps(v0.simdInternal_, v1.simdInternal_);
 153         t6   = _mm512_unpackhi_ps(v0.simdInternal_, v1.simdInternal_);
 154         t[0] = _mm512_shuffle_ps(t5, v2.simdInternal_, _MM_SHUFFLE(0, 0, 1, 0));
 155         t[1] = _mm512_shuffle_ps(t5, v2.simdInternal_, _MM_SHUFFLE(1, 1, 3, 2));
 156         t[2] = _mm512_shuffle_ps(t6, v2.simdInternal_, _MM_SHUFFLE(2, 2, 1, 0));
 157         t[3] = _mm512_shuffle_ps(t6, v2.simdInternal_, _MM_SHUFFLE(3, 3, 3, 2));
 158         for (i = 0; i < 4; i++)
 159         {
 160             _mm512_mask_storeu_ps(base + o[i], avx512Int2Mask(7),
 161                                   _mm512_castps128_ps512(_mm_add_ps(_mm_loadu_ps(base + o[i]),
 162                                                                     _mm512_castps512_ps128(t[i]))));
 163             _mm512_mask_storeu_ps(base + o[4 + i], avx512Int2Mask(7),
 164                                   _mm512_castps128_ps512(_mm_add_ps(_mm_loadu_ps(base + o[4 + i]),
 165                                                                     _mm512_extractf32x4_ps(t[i], 1))));
 166             _mm512_mask_storeu_ps(base + o[8 + i], avx512Int2Mask(7),
 167                                   _mm512_castps128_ps512(_mm_add_ps(_mm_loadu_ps(base + o[8 + i]),
 168                                                                     _mm512_extractf32x4_ps(t[i], 2))));
 169             _mm512_mask_storeu_ps(base + o[12 + i], avx512Int2Mask(7),
 170                                   _mm512_castps128_ps512(_mm_add_ps(_mm_loadu_ps(base + o[12 + i]),
 171                                                                     _mm512_extractf32x4_ps(t[i], 3))));
 172         }
 173     }
 174     else
 175     {
 176         // One could use shuffle here too if it is OK to overwrite the padded elements for alignment
 177         t5   = _mm512_unpacklo_ps(v0.simdInternal_, v2.simdInternal_);
 178         t6   = _mm512_unpackhi_ps(v0.simdInternal_, v2.simdInternal_);
 179         t7   = _mm512_unpacklo_ps(v1.simdInternal_, _mm512_setzero_ps());
 180         t8   = _mm512_unpackhi_ps(v1.simdInternal_, _mm512_setzero_ps());
 181         t[0] = _mm512_unpacklo_ps(t5, t7); // x0 y0 z0  0 | x4 y4 z4 0
 182         t[1] = _mm512_unpackhi_ps(t5, t7); // x1 y1 z1  0 | x5 y5 z5 0
 183         t[2] = _mm512_unpacklo_ps(t6, t8); // x2 y2 z2  0 | x6 y6 z6 0
 184         t[3] = _mm512_unpackhi_ps(t6, t8); // x3 y3 z3  0 | x7 y7 z7 0
 185         if (align % 4 == 0)
 186         {
 187             for (i = 0; i < 4; i++)
 188             {
 189                 _mm_store_ps(base + o[i],
 190                              _mm_add_ps(_mm_load_ps(base + o[i]), _mm512_castps512_ps128(t[i])));
 191                 _mm_store_ps(base + o[4 + i],
 192                              _mm_add_ps(_mm_load_ps(base + o[4 + i]), _mm512_extractf32x4_ps(t[i], 1)));
 193                 _mm_store_ps(base + o[8 + i],
 194                              _mm_add_ps(_mm_load_ps(base + o[8 + i]), _mm512_extractf32x4_ps(t[i], 2)));
 195                 _mm_store_ps(base + o[12 + i], _mm_add_ps(_mm_load_ps(base + o[12 + i]),
 196                                                           _mm512_extractf32x4_ps(t[i], 3)));
 197             }
 198         }
 199         else
 200         {
 201             for (i = 0; i < 4; i++)
 202             {
 203                 _mm_storeu_ps(base + o[i],
 204                               _mm_add_ps(_mm_loadu_ps(base + o[i]), _mm512_castps512_ps128(t[i])));
 205                 _mm_storeu_ps(base + o[4 + i], _mm_add_ps(_mm_loadu_ps(base + o[4 + i]),
 206                                                           _mm512_extractf32x4_ps(t[i], 1)));
 207                 _mm_storeu_ps(base + o[8 + i], _mm_add_ps(_mm_loadu_ps(base + o[8 + i]),
 208                                                           _mm512_extractf32x4_ps(t[i], 2)));
 209                 _mm_storeu_ps(base + o[12 + i], _mm_add_ps(_mm_loadu_ps(base + o[12 + i]),
 210                                                            _mm512_extractf32x4_ps(t[i], 3)));
 211             }
 212         }
 213     }
 214 }
 215
 216 template<int align>
 217 static inline void gmx_simdcall
 218                    transposeScatterDecrU(float* base, const std::int32_t offset[], SimdFloat v0, SimdFloat v1, SimdFloat v2)
 219 {
 220     __m512                                   t[4], t5, t6, t7, t8;
 221     int                                      i;
 222     alignas(GMX_SIMD_ALIGNMENT) std::int32_t o[16];
 223     store(o, fastMultiply<align>(simdLoad(offset, SimdFInt32Tag())));
 224     if (align < 4)
 225     {
 226         t5   = _mm512_unpacklo_ps(v0.simdInternal_, v1.simdInternal_);
 227         t6   = _mm512_unpackhi_ps(v0.simdInternal_, v1.simdInternal_);
 228         t[0] = _mm512_shuffle_ps(t5, v2.simdInternal_, _MM_SHUFFLE(0, 0, 1, 0));
 229         t[1] = _mm512_shuffle_ps(t5, v2.simdInternal_, _MM_SHUFFLE(1, 1, 3, 2));
 230         t[2] = _mm512_shuffle_ps(t6, v2.simdInternal_, _MM_SHUFFLE(2, 2, 1, 0));
 231         t[3] = _mm512_shuffle_ps(t6, v2.simdInternal_, _MM_SHUFFLE(3, 3, 3, 2));
 232         for (i = 0; i < 4; i++)
 233         {
 234             _mm512_mask_storeu_ps(base + o[i], avx512Int2Mask(7),
 235                                   _mm512_castps128_ps512(_mm_sub_ps(_mm_loadu_ps(base + o[i]),
 236                                                                     _mm512_castps512_ps128(t[i]))));
 237             _mm512_mask_storeu_ps(base + o[4 + i], avx512Int2Mask(7),
 238                                   _mm512_castps128_ps512(_mm_sub_ps(_mm_loadu_ps(base + o[4 + i]),
 239                                                                     _mm512_extractf32x4_ps(t[i], 1))));
 240             _mm512_mask_storeu_ps(base + o[8 + i], avx512Int2Mask(7),
 241                                   _mm512_castps128_ps512(_mm_sub_ps(_mm_loadu_ps(base + o[8 + i]),
 242                                                                     _mm512_extractf32x4_ps(t[i], 2))));
 243             _mm512_mask_storeu_ps(base + o[12 + i], avx512Int2Mask(7),
 244                                   _mm512_castps128_ps512(_mm_sub_ps(_mm_loadu_ps(base + o[12 + i]),
 245                                                                     _mm512_extractf32x4_ps(t[i], 3))));
 246         }
 247     }
 248     else
 249     {
 250         // One could use shuffle here too if it is OK to overwrite the padded elements for alignment
 251         t5   = _mm512_unpacklo_ps(v0.simdInternal_, v2.simdInternal_);
 252         t6   = _mm512_unpackhi_ps(v0.simdInternal_, v2.simdInternal_);
 253         t7   = _mm512_unpacklo_ps(v1.simdInternal_, _mm512_setzero_ps());
 254         t8   = _mm512_unpackhi_ps(v1.simdInternal_, _mm512_setzero_ps());
 255         t[0] = _mm512_unpacklo_ps(t5, t7); // x0 y0 z0  0 | x4 y4 z4 0
 256         t[1] = _mm512_unpackhi_ps(t5, t7); // x1 y1 z1  0 | x5 y5 z5 0
 257         t[2] = _mm512_unpacklo_ps(t6, t8); // x2 y2 z2  0 | x6 y6 z6 0
 258         t[3] = _mm512_unpackhi_ps(t6, t8); // x3 y3 z3  0 | x7 y7 z7 0
 259         if (align % 4 == 0)
 260         {
 261             for (i = 0; i < 4; i++)
 262             {
 263                 _mm_store_ps(base + o[i],
 264                              _mm_sub_ps(_mm_load_ps(base + o[i]), _mm512_castps512_ps128(t[i])));
 265                 _mm_store_ps(base + o[4 + i],
 266                              _mm_sub_ps(_mm_load_ps(base + o[4 + i]), _mm512_extractf32x4_ps(t[i], 1)));
 267                 _mm_store_ps(base + o[8 + i],
 268                              _mm_sub_ps(_mm_load_ps(base + o[8 + i]), _mm512_extractf32x4_ps(t[i], 2)));
 269                 _mm_store_ps(base + o[12 + i], _mm_sub_ps(_mm_load_ps(base + o[12 + i]),
 270                                                           _mm512_extractf32x4_ps(t[i], 3)));
 271             }
 272         }
 273         else
 274         {
 275             for (i = 0; i < 4; i++)
 276             {
 277                 _mm_storeu_ps(base + o[i],
 278                               _mm_sub_ps(_mm_loadu_ps(base + o[i]), _mm512_castps512_ps128(t[i])));
 279                 _mm_storeu_ps(base + o[4 + i], _mm_sub_ps(_mm_loadu_ps(base + o[4 + i]),
 280                                                           _mm512_extractf32x4_ps(t[i], 1)));
 281                 _mm_storeu_ps(base + o[8 + i], _mm_sub_ps(_mm_loadu_ps(base + o[8 + i]),
 282                                                           _mm512_extractf32x4_ps(t[i], 2)));
 283                 _mm_storeu_ps(base + o[12 + i], _mm_sub_ps(_mm_loadu_ps(base + o[12 + i]),
 284                                                            _mm512_extractf32x4_ps(t[i], 3)));
 285             }
 286         }
 287     }
 288 }
 289
 290 static inline void gmx_simdcall expandScalarsToTriplets(SimdFloat  scalar,
 291                                                         SimdFloat* triplets0,
 292                                                         SimdFloat* triplets1,
 293                                                         SimdFloat* triplets2)
 294 {
 295     triplets0->simdInternal_ = _mm512_permutexvar_ps(
 296             _mm512_set_epi32(5, 4, 4, 4, 3, 3, 3, 2, 2, 2, 1, 1, 1, 0, 0, 0), scalar.simdInternal_);
 297     triplets1->simdInternal_ = _mm512_permutexvar_ps(
 298             _mm512_set_epi32(10, 10, 9, 9, 9, 8, 8, 8, 7, 7, 7, 6, 6, 6, 5, 5), scalar.simdInternal_);
 299     triplets2->simdInternal_ = _mm512_permutexvar_ps(
 300             _mm512_set_epi32(15, 15, 15, 14, 14, 14, 13, 13, 13, 12, 12, 12, 11, 11, 11, 10),
 301             scalar.simdInternal_);
 302 }
 303
 304
 305 static inline float gmx_simdcall reduceIncr4ReturnSum(float* m, SimdFloat v0, SimdFloat v1, SimdFloat v2, SimdFloat v3)
 306 {
 307     __m512 t0, t1, t2;
 308     __m128 t3, t4;
 309
 310     assert(std::size_t(m) % 16 == 0);
 311
 312     t0 = _mm512_add_ps(v0.simdInternal_, _mm512_permute_ps(v0.simdInternal_, 0x4E));
 313     t0 = _mm512_mask_add_ps(t0, avx512Int2Mask(0xCCCC), v2.simdInternal_,
 314                             _mm512_permute_ps(v2.simdInternal_, 0x4E));
 315     t1 = _mm512_add_ps(v1.simdInternal_, _mm512_permute_ps(v1.simdInternal_, 0x4E));
 316     t1 = _mm512_mask_add_ps(t1, avx512Int2Mask(0xCCCC), v3.simdInternal_,
 317                             _mm512_permute_ps(v3.simdInternal_, 0x4E));
 318     t2 = _mm512_add_ps(t0, _mm512_permute_ps(t0, 0xB1));
 319     t2 = _mm512_mask_add_ps(t2, avx512Int2Mask(0xAAAA), t1, _mm512_permute_ps(t1, 0xB1));
 320
 321     t2 = _mm512_add_ps(t2, _mm512_shuffle_f32x4(t2, t2, 0x4E));
 322     t2 = _mm512_add_ps(t2, _mm512_shuffle_f32x4(t2, t2, 0xB1));
 323
 324     t3 = _mm512_castps512_ps128(t2);
 325     t4 = _mm_load_ps(m);
 326     t4 = _mm_add_ps(t4, t3);
 327     _mm_store_ps(m, t4);
 328
 329     t3 = _mm_add_ps(t3, _mm_permute_ps(t3, 0x4E));
 330     t3 = _mm_add_ps(t3, _mm_permute_ps(t3, 0xB1));
 331
 332     return _mm_cvtss_f32(t3);
 333 }
 334
 335 static inline SimdFloat gmx_simdcall loadDualHsimd(const float* m0, const float* m1)
 336 {
 337     assert(std::size_t(m0) % 32 == 0);
 338     assert(std::size_t(m1) % 32 == 0);
 339
 340     return { _mm512_castpd_ps(_mm512_insertf64x4(
 341             _mm512_castpd256_pd512(_mm256_load_pd(reinterpret_cast<const double*>(m0))),
 342             _mm256_load_pd(reinterpret_cast<const double*>(m1)), 1)) };
 343 }
 344
 345 static inline SimdFloat gmx_simdcall loadDuplicateHsimd(const float* m)
 346 {
 347     assert(std::size_t(m) % 32 == 0);
 348     return { _mm512_castpd_ps(_mm512_broadcast_f64x4(_mm256_load_pd(reinterpret_cast<const double*>(m)))) };
 349 }
 350
 351 static inline SimdFloat gmx_simdcall loadU1DualHsimd(const float* m)
 352 {
 353     return { _mm512_shuffle_f32x4(_mm512_broadcastss_ps(_mm_load_ss(m)),
 354                                   _mm512_broadcastss_ps(_mm_load_ss(m + 1)), 0x44) };
 355 }
 356
 357
 358 static inline void gmx_simdcall storeDualHsimd(float* m0, float* m1, SimdFloat a)
 359 {
 360     assert(std::size_t(m0) % 32 == 0);
 361     assert(std::size_t(m1) % 32 == 0);
 362
 363     _mm256_store_ps(m0, _mm512_castps512_ps256(a.simdInternal_));
 364     _mm256_store_pd(reinterpret_cast<double*>(m1),
 365                     _mm512_extractf64x4_pd(_mm512_castps_pd(a.simdInternal_), 1));
 366 }
 367
 368 static inline void gmx_simdcall incrDualHsimd(float* m0, float* m1, SimdFloat a)
 369 {
 370     assert(std::size_t(m0) % 32 == 0);
 371     assert(std::size_t(m1) % 32 == 0);
 372
 373     __m256 x;
 374
 375     // Lower half
 376     x = _mm256_load_ps(m0);
 377     x = _mm256_add_ps(x, _mm512_castps512_ps256(a.simdInternal_));
 378     _mm256_store_ps(m0, x);
 379
 380     // Upper half
 381     x = _mm256_load_ps(m1);
 382     x = _mm256_add_ps(x, _mm256_castpd_ps(_mm512_extractf64x4_pd(_mm512_castps_pd(a.simdInternal_), 1)));
 383     _mm256_store_ps(m1, x);
 384 }
 385
 386 static inline void gmx_simdcall decrHsimd(float* m, SimdFloat a)
 387 {
 388     __m256 t;
 389
 390     assert(std::size_t(m) % 32 == 0);
 391
 392     a.simdInternal_ = _mm512_add_ps(a.simdInternal_,
 393                                     _mm512_shuffle_f32x4(a.simdInternal_, a.simdInternal_, 0xEE));
 394     t               = _mm256_load_ps(m);
 395     t               = _mm256_sub_ps(t, _mm512_castps512_ps256(a.simdInternal_));
 396     _mm256_store_ps(m, t);
 397 }
 398
 399
 400 template<int align>
 401 static inline void gmx_simdcall gatherLoadTransposeHsimd(const float*       base0,
 402                                                          const float*       base1,
 403                                                          const std::int32_t offset[],
 404                                                          SimdFloat*         v0,
 405                                                          SimdFloat*         v1)
 406 {
 407     __m256i idx;
 408     __m512  tmp1, tmp2;
 409
 410     assert(std::size_t(offset) % 32 == 0);
 411     assert(std::size_t(base0) % 8 == 0);
 412     assert(std::size_t(base1) % 8 == 0);
 413
 414     idx = _mm256_load_si256(reinterpret_cast<const __m256i*>(offset));
 415
 416     static_assert(align == 2 || align == 4, "If more are needed use fastMultiply");
 417     if (align == 4)
 418     {
 419         idx = _mm256_slli_epi32(idx, 1);
 420     }
 421
 422     tmp1 = _mm512_castpd_ps(
 423             _mm512_i32gather_pd(idx, reinterpret_cast<const double*>(base0), sizeof(double)));
 424     tmp2 = _mm512_castpd_ps(
 425             _mm512_i32gather_pd(idx, reinterpret_cast<const double*>(base1), sizeof(double)));
 426
 427     v0->simdInternal_ = _mm512_mask_moveldup_ps(tmp1, 0xAAAA, tmp2);
 428     v1->simdInternal_ = _mm512_mask_movehdup_ps(tmp2, 0x5555, tmp1);
 429
 430     v0->simdInternal_ = _mm512_permutexvar_ps(
 431             _mm512_set_epi32(15, 13, 11, 9, 7, 5, 3, 1, 14, 12, 10, 8, 6, 4, 2, 0), v0->simdInternal_);
 432     v1->simdInternal_ = _mm512_permutexvar_ps(
 433             _mm512_set_epi32(15, 13, 11, 9, 7, 5, 3, 1, 14, 12, 10, 8, 6, 4, 2, 0), v1->simdInternal_);
 434 }
 435
 436 static inline float gmx_simdcall reduceIncr4ReturnSumHsimd(float* m, SimdFloat v0, SimdFloat v1)
 437 {
 438     __m512 t0, t1;
 439     __m128 t2, t3;
 440
 441     assert(std::size_t(m) % 16 == 0);
 442
 443     t0 = _mm512_shuffle_f32x4(v0.simdInternal_, v1.simdInternal_, 0x88);
 444     t1 = _mm512_shuffle_f32x4(v0.simdInternal_, v1.simdInternal_, 0xDD);
 445     t0 = _mm512_add_ps(t0, t1);
 446     t0 = _mm512_add_ps(t0, _mm512_permute_ps(t0, 0x4E));
 447     t0 = _mm512_add_ps(t0, _mm512_permute_ps(t0, 0xB1));
 448     t0 = _mm512_maskz_compress_ps(avx512Int2Mask(0x1111), t0);
 449
 450     t3 = _mm512_castps512_ps128(t0);
 451     t2 = _mm_load_ps(m);
 452     t2 = _mm_add_ps(t2, t3);
 453     _mm_store_ps(m, t2);
 454
 455     t3 = _mm_add_ps(t3, _mm_permute_ps(t3, 0x4E));
 456     t3 = _mm_add_ps(t3, _mm_permute_ps(t3, 0xB1));
 457
 458     return _mm_cvtss_f32(t3);
 459 }
 460
 461 static inline SimdFloat gmx_simdcall loadUNDuplicate4(const float* f)
 462 {
 463     return { _mm512_permute_ps(_mm512_maskz_expandloadu_ps(0x1111, f), 0) };
 464 }
 465
 466 static inline SimdFloat gmx_simdcall load4DuplicateN(const float* f)
 467 {
 468     return { _mm512_broadcast_f32x4(_mm_load_ps(f)) };
 469 }
 470
 471 static inline SimdFloat gmx_simdcall loadU4NOffset(const float* f, int offset)
 472 {
 473     const __m256i idx = _mm256_setr_epi32(0, 0, 1, 1, 2, 2, 3, 3);
 474     const __m256i gdx = _mm256_add_epi32(_mm256_setr_epi32(0, 2, 0, 2, 0, 2, 0, 2),
 475                                          _mm256_mullo_epi32(idx, _mm256_set1_epi32(offset)));
 476     return { _mm512_castpd_ps(_mm512_i32gather_pd(gdx, reinterpret_cast<const double*>(f), sizeof(float))) };
 477 }
 478
 479 } // namespace gmx
 480
 481 #endif // GMX_SIMD_IMPL_X86_AVX_512_UTIL_FLOAT_H