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 by the GROMACS development team.
   5  * Copyright (c) 2019,2020, by the GROMACS development team, led by
   6  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   7  * and including many others, as listed in the AUTHORS file in the
   8  * top-level source directory and at http://www.gromacs.org.
   9  *
  10  * GROMACS is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public License
  12  * as published by the Free Software Foundation; either version 2.1
  13  * of the License, or (at your option) any later version.
  14  *
  15  * GROMACS is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * Lesser General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU Lesser General Public
  21  * License along with GROMACS; if not, see
  22  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  23  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  24  *
  25  * If you want to redistribute modifications to GROMACS, please
  26  * consider that scientific software is very special. Version
  27  * control is crucial - bugs must be traceable. We will be happy to
  28  * consider code for inclusion in the official distribution, but
  29  * derived work must not be called official GROMACS. Details are found
  30  * in the README & COPYING files - if they are missing, get the
  31  * official version at http://www.gromacs.org.
  32  *
  33  * To help us fund GROMACS development, we humbly ask that you cite
  34  * the research papers on the package. Check out http://www.gromacs.org.
  35  */
  36
  37 #ifndef GMX_SIMD_IMPL_X86_AVX_512_UTIL_FLOAT_H
  38 #define GMX_SIMD_IMPL_X86_AVX_512_UTIL_FLOAT_H
  39
  40 #include "config.h"
  41
  42 #include <cassert>
  43 #include <cstdint>
  44
  45 #include <immintrin.h>
  46
  47 #include "gromacs/utility/basedefinitions.h"
  48
  49 #include "impl_x86_avx_512_general.h"
  50 #include "impl_x86_avx_512_simd_float.h"
  51
  52 namespace gmx
  53 {
  54
  55 static const int c_simdBestPairAlignmentFloat = 2;
  56
  57 namespace
  58 {
  59 // Multiply function optimized for powers of 2, for which it is done by
  60 // shifting. Currently up to 8 is accelerated. Could be accelerated for any
  61 // number with a constexpr log2 function.
  62 template<int n>
  63 static inline SimdFInt32 fastMultiply(SimdFInt32 x)
  64 {
  65     if (n == 2)
  66     {
  67         return _mm512_slli_epi32(x.simdInternal_, 1);
  68     }
  69     else if (n == 4)
  70     {
  71         return _mm512_slli_epi32(x.simdInternal_, 2);
  72     }
  73     else if (n == 8)
  74     {
  75         return _mm512_slli_epi32(x.simdInternal_, 3);
  76     }
  77     else
  78     {
  79         return x * n;
  80     }
  81 }
  82
  83 template<int align>
  84 static inline void gmx_simdcall gatherLoadBySimdIntTranspose(const float*, SimdFInt32)
  85 {
  86     // Nothing to do. Termination of recursion.
  87 }
  88
  89 /* This is an internal helper function used by decr3Hsimd(...).
  90  */
  91 inline void gmx_simdcall decrHsimd(float* m, SimdFloat a)
  92 {
  93     __m256 t;
  94
  95     assert(std::size_t(m) % 32 == 0);
  96
  97     a.simdInternal_ = _mm512_add_ps(a.simdInternal_,
  98                                     _mm512_shuffle_f32x4(a.simdInternal_, a.simdInternal_, 0xEE));
  99     t               = _mm256_load_ps(m);
 100     t               = _mm256_sub_ps(t, _mm512_castps512_ps256(a.simdInternal_));
 101     _mm256_store_ps(m, t);
 102 }
 103 } // namespace
 104
 105 template<int align, typename... Targs>
 106 static inline void gmx_simdcall
 107                    gatherLoadBySimdIntTranspose(const float* base, SimdFInt32 offset, SimdFloat* v, Targs... Fargs)
 108 {
 109     // For align 1 or 2: No multiplication of offset is needed
 110     if (align > 2)
 111     {
 112         offset = fastMultiply<align>(offset);
 113     }
 114     // For align 2: Scale of 2*sizeof(float) is used (maximum supported scale)
 115     constexpr int align_ = (align > 2) ? 1 : align;
 116     v->simdInternal_     = _mm512_i32gather_ps(offset.simdInternal_, base, sizeof(float) * align_);
 117     // Gather remaining elements. Avoid extra multiplication (new align is 1 or 2).
 118     gatherLoadBySimdIntTranspose<align_>(base + 1, offset, Fargs...);
 119 }
 120
 121 template<int align, typename... Targs>
 122 static inline void gmx_simdcall
 123                    gatherLoadUBySimdIntTranspose(const float* base, SimdFInt32 offset, SimdFloat* v, Targs... Fargs)
 124 {
 125     gatherLoadBySimdIntTranspose<align>(base, offset, v, Fargs...);
 126 }
 127
 128 template<int align, typename... Targs>
 129 static inline void gmx_simdcall
 130                    gatherLoadTranspose(const float* base, const std::int32_t offset[], SimdFloat* v, Targs... Fargs)
 131 {
 132     gatherLoadBySimdIntTranspose<align>(base, simdLoad(offset, SimdFInt32Tag()), v, Fargs...);
 133 }
 134
 135 template<int align, typename... Targs>
 136 static inline void gmx_simdcall
 137                    gatherLoadUTranspose(const float* base, const std::int32_t offset[], SimdFloat* v, Targs... Fargs)
 138 {
 139     gatherLoadBySimdIntTranspose<align>(base, simdLoad(offset, SimdFInt32Tag()), v, Fargs...);
 140 }
 141
 142 template<int align>
 143 static inline void gmx_simdcall
 144                    transposeScatterStoreU(float* base, const std::int32_t offset[], SimdFloat v0, SimdFloat v1, SimdFloat v2)
 145 {
 146     SimdFInt32 simdoffset = simdLoad(offset, SimdFInt32Tag());
 147     if (align > 2)
 148     {
 149         simdoffset = fastMultiply<align>(simdoffset);
 150     }
 151     constexpr size_t scale = (align > 2) ? sizeof(float) : sizeof(float) * align;
 152
 153     _mm512_i32scatter_ps(base, simdoffset.simdInternal_, v0.simdInternal_, scale);
 154     _mm512_i32scatter_ps(&(base[1]), simdoffset.simdInternal_, v1.simdInternal_, scale);
 155     _mm512_i32scatter_ps(&(base[2]), simdoffset.simdInternal_, v2.simdInternal_, scale);
 156 }
 157
 158 template<int align>
 159 static inline void gmx_simdcall
 160                    transposeScatterIncrU(float* base, const std::int32_t offset[], SimdFloat v0, SimdFloat v1, SimdFloat v2)
 161 {
 162     __m512                                   t[4], t5, t6, t7, t8;
 163     int                                      i;
 164     alignas(GMX_SIMD_ALIGNMENT) std::int32_t o[16];
 165     store(o, fastMultiply<align>(simdLoad(offset, SimdFInt32Tag())));
 166     if (align < 4)
 167     {
 168         t5   = _mm512_unpacklo_ps(v0.simdInternal_, v1.simdInternal_);
 169         t6   = _mm512_unpackhi_ps(v0.simdInternal_, v1.simdInternal_);
 170         t[0] = _mm512_shuffle_ps(t5, v2.simdInternal_, _MM_SHUFFLE(0, 0, 1, 0));
 171         t[1] = _mm512_shuffle_ps(t5, v2.simdInternal_, _MM_SHUFFLE(1, 1, 3, 2));
 172         t[2] = _mm512_shuffle_ps(t6, v2.simdInternal_, _MM_SHUFFLE(2, 2, 1, 0));
 173         t[3] = _mm512_shuffle_ps(t6, v2.simdInternal_, _MM_SHUFFLE(3, 3, 3, 2));
 174         for (i = 0; i < 4; i++)
 175         {
 176             _mm512_mask_storeu_ps(base + o[i],
 177                                   avx512Int2Mask(7),
 178                                   _mm512_castps128_ps512(_mm_add_ps(_mm_loadu_ps(base + o[i]),
 179                                                                     _mm512_castps512_ps128(t[i]))));
 180             _mm512_mask_storeu_ps(base + o[4 + i],
 181                                   avx512Int2Mask(7),
 182                                   _mm512_castps128_ps512(_mm_add_ps(_mm_loadu_ps(base + o[4 + i]),
 183                                                                     _mm512_extractf32x4_ps(t[i], 1))));
 184             _mm512_mask_storeu_ps(base + o[8 + i],
 185                                   avx512Int2Mask(7),
 186                                   _mm512_castps128_ps512(_mm_add_ps(_mm_loadu_ps(base + o[8 + i]),
 187                                                                     _mm512_extractf32x4_ps(t[i], 2))));
 188             _mm512_mask_storeu_ps(base + o[12 + i],
 189                                   avx512Int2Mask(7),
 190                                   _mm512_castps128_ps512(_mm_add_ps(_mm_loadu_ps(base + o[12 + i]),
 191                                                                     _mm512_extractf32x4_ps(t[i], 3))));
 192         }
 193     }
 194     else
 195     {
 196         // One could use shuffle here too if it is OK to overwrite the padded elements for alignment
 197         t5   = _mm512_unpacklo_ps(v0.simdInternal_, v2.simdInternal_);
 198         t6   = _mm512_unpackhi_ps(v0.simdInternal_, v2.simdInternal_);
 199         t7   = _mm512_unpacklo_ps(v1.simdInternal_, _mm512_setzero_ps());
 200         t8   = _mm512_unpackhi_ps(v1.simdInternal_, _mm512_setzero_ps());
 201         t[0] = _mm512_unpacklo_ps(t5, t7); // x0 y0 z0  0 | x4 y4 z4 0
 202         t[1] = _mm512_unpackhi_ps(t5, t7); // x1 y1 z1  0 | x5 y5 z5 0
 203         t[2] = _mm512_unpacklo_ps(t6, t8); // x2 y2 z2  0 | x6 y6 z6 0
 204         t[3] = _mm512_unpackhi_ps(t6, t8); // x3 y3 z3  0 | x7 y7 z7 0
 205         if (align % 4 == 0)
 206         {
 207             for (i = 0; i < 4; i++)
 208             {
 209                 _mm_store_ps(base + o[i],
 210                              _mm_add_ps(_mm_load_ps(base + o[i]), _mm512_castps512_ps128(t[i])));
 211                 _mm_store_ps(base + o[4 + i],
 212                              _mm_add_ps(_mm_load_ps(base + o[4 + i]), _mm512_extractf32x4_ps(t[i], 1)));
 213                 _mm_store_ps(base + o[8 + i],
 214                              _mm_add_ps(_mm_load_ps(base + o[8 + i]), _mm512_extractf32x4_ps(t[i], 2)));
 215                 _mm_store_ps(base + o[12 + i],
 216                              _mm_add_ps(_mm_load_ps(base + o[12 + i]), _mm512_extractf32x4_ps(t[i], 3)));
 217             }
 218         }
 219         else
 220         {
 221             for (i = 0; i < 4; i++)
 222             {
 223                 _mm_storeu_ps(base + o[i],
 224                               _mm_add_ps(_mm_loadu_ps(base + o[i]), _mm512_castps512_ps128(t[i])));
 225                 _mm_storeu_ps(base + o[4 + i],
 226                               _mm_add_ps(_mm_loadu_ps(base + o[4 + i]), _mm512_extractf32x4_ps(t[i], 1)));
 227                 _mm_storeu_ps(base + o[8 + i],
 228                               _mm_add_ps(_mm_loadu_ps(base + o[8 + i]), _mm512_extractf32x4_ps(t[i], 2)));
 229                 _mm_storeu_ps(base + o[12 + i],
 230                               _mm_add_ps(_mm_loadu_ps(base + o[12 + i]), _mm512_extractf32x4_ps(t[i], 3)));
 231             }
 232         }
 233     }
 234 }
 235
 236 template<int align>
 237 static inline void gmx_simdcall
 238                    transposeScatterDecrU(float* base, const std::int32_t offset[], SimdFloat v0, SimdFloat v1, SimdFloat v2)
 239 {
 240     __m512                                   t[4], t5, t6, t7, t8;
 241     int                                      i;
 242     alignas(GMX_SIMD_ALIGNMENT) std::int32_t o[16];
 243     store(o, fastMultiply<align>(simdLoad(offset, SimdFInt32Tag())));
 244     if (align < 4)
 245     {
 246         t5   = _mm512_unpacklo_ps(v0.simdInternal_, v1.simdInternal_);
 247         t6   = _mm512_unpackhi_ps(v0.simdInternal_, v1.simdInternal_);
 248         t[0] = _mm512_shuffle_ps(t5, v2.simdInternal_, _MM_SHUFFLE(0, 0, 1, 0));
 249         t[1] = _mm512_shuffle_ps(t5, v2.simdInternal_, _MM_SHUFFLE(1, 1, 3, 2));
 250         t[2] = _mm512_shuffle_ps(t6, v2.simdInternal_, _MM_SHUFFLE(2, 2, 1, 0));
 251         t[3] = _mm512_shuffle_ps(t6, v2.simdInternal_, _MM_SHUFFLE(3, 3, 3, 2));
 252         for (i = 0; i < 4; i++)
 253         {
 254             _mm512_mask_storeu_ps(base + o[i],
 255                                   avx512Int2Mask(7),
 256                                   _mm512_castps128_ps512(_mm_sub_ps(_mm_loadu_ps(base + o[i]),
 257                                                                     _mm512_castps512_ps128(t[i]))));
 258             _mm512_mask_storeu_ps(base + o[4 + i],
 259                                   avx512Int2Mask(7),
 260                                   _mm512_castps128_ps512(_mm_sub_ps(_mm_loadu_ps(base + o[4 + i]),
 261                                                                     _mm512_extractf32x4_ps(t[i], 1))));
 262             _mm512_mask_storeu_ps(base + o[8 + i],
 263                                   avx512Int2Mask(7),
 264                                   _mm512_castps128_ps512(_mm_sub_ps(_mm_loadu_ps(base + o[8 + i]),
 265                                                                     _mm512_extractf32x4_ps(t[i], 2))));
 266             _mm512_mask_storeu_ps(base + o[12 + i],
 267                                   avx512Int2Mask(7),
 268                                   _mm512_castps128_ps512(_mm_sub_ps(_mm_loadu_ps(base + o[12 + i]),
 269                                                                     _mm512_extractf32x4_ps(t[i], 3))));
 270         }
 271     }
 272     else
 273     {
 274         // One could use shuffle here too if it is OK to overwrite the padded elements for alignment
 275         t5   = _mm512_unpacklo_ps(v0.simdInternal_, v2.simdInternal_);
 276         t6   = _mm512_unpackhi_ps(v0.simdInternal_, v2.simdInternal_);
 277         t7   = _mm512_unpacklo_ps(v1.simdInternal_, _mm512_setzero_ps());
 278         t8   = _mm512_unpackhi_ps(v1.simdInternal_, _mm512_setzero_ps());
 279         t[0] = _mm512_unpacklo_ps(t5, t7); // x0 y0 z0  0 | x4 y4 z4 0
 280         t[1] = _mm512_unpackhi_ps(t5, t7); // x1 y1 z1  0 | x5 y5 z5 0
 281         t[2] = _mm512_unpacklo_ps(t6, t8); // x2 y2 z2  0 | x6 y6 z6 0
 282         t[3] = _mm512_unpackhi_ps(t6, t8); // x3 y3 z3  0 | x7 y7 z7 0
 283         if (align % 4 == 0)
 284         {
 285             for (i = 0; i < 4; i++)
 286             {
 287                 _mm_store_ps(base + o[i],
 288                              _mm_sub_ps(_mm_load_ps(base + o[i]), _mm512_castps512_ps128(t[i])));
 289                 _mm_store_ps(base + o[4 + i],
 290                              _mm_sub_ps(_mm_load_ps(base + o[4 + i]), _mm512_extractf32x4_ps(t[i], 1)));
 291                 _mm_store_ps(base + o[8 + i],
 292                              _mm_sub_ps(_mm_load_ps(base + o[8 + i]), _mm512_extractf32x4_ps(t[i], 2)));
 293                 _mm_store_ps(base + o[12 + i],
 294                              _mm_sub_ps(_mm_load_ps(base + o[12 + i]), _mm512_extractf32x4_ps(t[i], 3)));
 295             }
 296         }
 297         else
 298         {
 299             for (i = 0; i < 4; i++)
 300             {
 301                 _mm_storeu_ps(base + o[i],
 302                               _mm_sub_ps(_mm_loadu_ps(base + o[i]), _mm512_castps512_ps128(t[i])));
 303                 _mm_storeu_ps(base + o[4 + i],
 304                               _mm_sub_ps(_mm_loadu_ps(base + o[4 + i]), _mm512_extractf32x4_ps(t[i], 1)));
 305                 _mm_storeu_ps(base + o[8 + i],
 306                               _mm_sub_ps(_mm_loadu_ps(base + o[8 + i]), _mm512_extractf32x4_ps(t[i], 2)));
 307                 _mm_storeu_ps(base + o[12 + i],
 308                               _mm_sub_ps(_mm_loadu_ps(base + o[12 + i]), _mm512_extractf32x4_ps(t[i], 3)));
 309             }
 310         }
 311     }
 312 }
 313
 314 static inline void gmx_simdcall expandScalarsToTriplets(SimdFloat  scalar,
 315                                                         SimdFloat* triplets0,
 316                                                         SimdFloat* triplets1,
 317                                                         SimdFloat* triplets2)
 318 {
 319     triplets0->simdInternal_ = _mm512_permutexvar_ps(
 320             _mm512_set_epi32(5, 4, 4, 4, 3, 3, 3, 2, 2, 2, 1, 1, 1, 0, 0, 0), scalar.simdInternal_);
 321     triplets1->simdInternal_ = _mm512_permutexvar_ps(
 322             _mm512_set_epi32(10, 10, 9, 9, 9, 8, 8, 8, 7, 7, 7, 6, 6, 6, 5, 5), scalar.simdInternal_);
 323     triplets2->simdInternal_ = _mm512_permutexvar_ps(
 324             _mm512_set_epi32(15, 15, 15, 14, 14, 14, 13, 13, 13, 12, 12, 12, 11, 11, 11, 10),
 325             scalar.simdInternal_);
 326 }
 327
 328
 329 static inline float gmx_simdcall reduceIncr4ReturnSum(float* m, SimdFloat v0, SimdFloat v1, SimdFloat v2, SimdFloat v3)
 330 {
 331     __m512 t0, t1, t2;
 332     __m128 t3, t4;
 333
 334     assert(std::size_t(m) % 16 == 0);
 335
 336     t0 = _mm512_add_ps(v0.simdInternal_, _mm512_permute_ps(v0.simdInternal_, 0x4E));
 337     t0 = _mm512_mask_add_ps(
 338             t0, avx512Int2Mask(0xCCCC), v2.simdInternal_, _mm512_permute_ps(v2.simdInternal_, 0x4E));
 339     t1 = _mm512_add_ps(v1.simdInternal_, _mm512_permute_ps(v1.simdInternal_, 0x4E));
 340     t1 = _mm512_mask_add_ps(
 341             t1, avx512Int2Mask(0xCCCC), v3.simdInternal_, _mm512_permute_ps(v3.simdInternal_, 0x4E));
 342     t2 = _mm512_add_ps(t0, _mm512_permute_ps(t0, 0xB1));
 343     t2 = _mm512_mask_add_ps(t2, avx512Int2Mask(0xAAAA), t1, _mm512_permute_ps(t1, 0xB1));
 344
 345     t2 = _mm512_add_ps(t2, _mm512_shuffle_f32x4(t2, t2, 0x4E));
 346     t2 = _mm512_add_ps(t2, _mm512_shuffle_f32x4(t2, t2, 0xB1));
 347
 348     t3 = _mm512_castps512_ps128(t2);
 349     t4 = _mm_load_ps(m);
 350     t4 = _mm_add_ps(t4, t3);
 351     _mm_store_ps(m, t4);
 352
 353     t3 = _mm_add_ps(t3, _mm_permute_ps(t3, 0x4E));
 354     t3 = _mm_add_ps(t3, _mm_permute_ps(t3, 0xB1));
 355
 356     return _mm_cvtss_f32(t3);
 357 }
 358
 359 static inline SimdFloat gmx_simdcall loadDualHsimd(const float* m0, const float* m1)
 360 {
 361     assert(std::size_t(m0) % 32 == 0);
 362     assert(std::size_t(m1) % 32 == 0);
 363
 364     return { _mm512_castpd_ps(_mm512_insertf64x4(
 365             _mm512_castpd256_pd512(_mm256_load_pd(reinterpret_cast<const double*>(m0))),
 366             _mm256_load_pd(reinterpret_cast<const double*>(m1)),
 367             1)) };
 368 }
 369
 370 static inline SimdFloat gmx_simdcall loadDuplicateHsimd(const float* m)
 371 {
 372     assert(std::size_t(m) % 32 == 0);
 373     return { _mm512_castpd_ps(_mm512_broadcast_f64x4(_mm256_load_pd(reinterpret_cast<const double*>(m)))) };
 374 }
 375
 376 static inline SimdFloat gmx_simdcall loadU1DualHsimd(const float* m)
 377 {
 378     return { _mm512_shuffle_f32x4(
 379             _mm512_broadcastss_ps(_mm_load_ss(m)), _mm512_broadcastss_ps(_mm_load_ss(m + 1)), 0x44) };
 380 }
 381
 382
 383 static inline void gmx_simdcall storeDualHsimd(float* m0, float* m1, SimdFloat a)
 384 {
 385     assert(std::size_t(m0) % 32 == 0);
 386     assert(std::size_t(m1) % 32 == 0);
 387
 388     _mm256_store_ps(m0, _mm512_castps512_ps256(a.simdInternal_));
 389     _mm256_store_pd(reinterpret_cast<double*>(m1),
 390                     _mm512_extractf64x4_pd(_mm512_castps_pd(a.simdInternal_), 1));
 391 }
 392
 393 static inline void gmx_simdcall incrDualHsimd(float* m0, float* m1, SimdFloat a)
 394 {
 395     assert(std::size_t(m0) % 32 == 0);
 396     assert(std::size_t(m1) % 32 == 0);
 397
 398     __m256 x;
 399
 400     // Lower half
 401     x = _mm256_load_ps(m0);
 402     x = _mm256_add_ps(x, _mm512_castps512_ps256(a.simdInternal_));
 403     _mm256_store_ps(m0, x);
 404
 405     // Upper half
 406     x = _mm256_load_ps(m1);
 407     x = _mm256_add_ps(x, _mm256_castpd_ps(_mm512_extractf64x4_pd(_mm512_castps_pd(a.simdInternal_), 1)));
 408     _mm256_store_ps(m1, x);
 409 }
 410
 411 static inline void gmx_simdcall decr3Hsimd(float* m, SimdFloat a0, SimdFloat a1, SimdFloat a2)
 412 {
 413     decrHsimd(m, a0);
 414     decrHsimd(m + GMX_SIMD_FLOAT_WIDTH / 2, a1);
 415     decrHsimd(m + GMX_SIMD_FLOAT_WIDTH, a2);
 416 }
 417
 418
 419 template<int align>
 420 static inline void gmx_simdcall gatherLoadTransposeHsimd(const float*       base0,
 421                                                          const float*       base1,
 422                                                          const std::int32_t offset[],
 423                                                          SimdFloat*         v0,
 424                                                          SimdFloat*         v1)
 425 {
 426     __m256i idx;
 427     __m512  tmp1, tmp2;
 428
 429     assert(std::size_t(offset) % 32 == 0);
 430     assert(std::size_t(base0) % 8 == 0);
 431     assert(std::size_t(base1) % 8 == 0);
 432
 433     idx = _mm256_load_si256(reinterpret_cast<const __m256i*>(offset));
 434
 435     static_assert(align == 2 || align == 4, "If more are needed use fastMultiply");
 436     if (align == 4)
 437     {
 438         idx = _mm256_slli_epi32(idx, 1);
 439     }
 440
 441     tmp1 = _mm512_castpd_ps(
 442             _mm512_i32gather_pd(idx, reinterpret_cast<const double*>(base0), sizeof(double)));
 443     tmp2 = _mm512_castpd_ps(
 444             _mm512_i32gather_pd(idx, reinterpret_cast<const double*>(base1), sizeof(double)));
 445
 446     v0->simdInternal_ = _mm512_mask_moveldup_ps(tmp1, 0xAAAA, tmp2);
 447     v1->simdInternal_ = _mm512_mask_movehdup_ps(tmp2, 0x5555, tmp1);
 448
 449     v0->simdInternal_ = _mm512_permutexvar_ps(
 450             _mm512_set_epi32(15, 13, 11, 9, 7, 5, 3, 1, 14, 12, 10, 8, 6, 4, 2, 0), v0->simdInternal_);
 451     v1->simdInternal_ = _mm512_permutexvar_ps(
 452             _mm512_set_epi32(15, 13, 11, 9, 7, 5, 3, 1, 14, 12, 10, 8, 6, 4, 2, 0), v1->simdInternal_);
 453 }
 454
 455 static inline float gmx_simdcall reduceIncr4ReturnSumHsimd(float* m, SimdFloat v0, SimdFloat v1)
 456 {
 457     __m512 t0, t1;
 458     __m128 t2, t3;
 459
 460     assert(std::size_t(m) % 16 == 0);
 461
 462     t0 = _mm512_shuffle_f32x4(v0.simdInternal_, v1.simdInternal_, 0x88);
 463     t1 = _mm512_shuffle_f32x4(v0.simdInternal_, v1.simdInternal_, 0xDD);
 464     t0 = _mm512_add_ps(t0, t1);
 465     t0 = _mm512_add_ps(t0, _mm512_permute_ps(t0, 0x4E));
 466     t0 = _mm512_add_ps(t0, _mm512_permute_ps(t0, 0xB1));
 467     t0 = _mm512_maskz_compress_ps(avx512Int2Mask(0x1111), t0);
 468
 469     t3 = _mm512_castps512_ps128(t0);
 470     t2 = _mm_load_ps(m);
 471     t2 = _mm_add_ps(t2, t3);
 472     _mm_store_ps(m, t2);
 473
 474     t3 = _mm_add_ps(t3, _mm_permute_ps(t3, 0x4E));
 475     t3 = _mm_add_ps(t3, _mm_permute_ps(t3, 0xB1));
 476
 477     return _mm_cvtss_f32(t3);
 478 }
 479
 480 static inline SimdFloat gmx_simdcall loadUNDuplicate4(const float* f)
 481 {
 482     return { _mm512_permute_ps(_mm512_maskz_expandloadu_ps(0x1111, f), 0) };
 483 }
 484
 485 static inline SimdFloat gmx_simdcall load4DuplicateN(const float* f)
 486 {
 487     return { _mm512_broadcast_f32x4(_mm_load_ps(f)) };
 488 }
 489
 490 static inline SimdFloat gmx_simdcall loadU4NOffset(const float* f, int offset)
 491 {
 492     const __m256i idx = _mm256_setr_epi32(0, 0, 1, 1, 2, 2, 3, 3);
 493     const __m256i gdx = _mm256_add_epi32(_mm256_setr_epi32(0, 2, 0, 2, 0, 2, 0, 2),
 494                                          _mm256_mullo_epi32(idx, _mm256_set1_epi32(offset)));
 495     return { _mm512_castpd_ps(_mm512_i32gather_pd(gdx, reinterpret_cast<const double*>(f), sizeof(float))) };
 496 }
 497
 498 } // namespace gmx
 499
 500 #endif // GMX_SIMD_IMPL_X86_AVX_512_UTIL_FLOAT_H