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, 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
  84 gatherLoadBySimdIntTranspose(const float *, SimdFInt32)
  85 {
  86     //Nothing to do. Termination of recursion.
  87 }
  88 }
  89
  90 template <int align, typename ... Targs>
  91 static inline void gmx_simdcall
  92 gatherLoadBySimdIntTranspose(const float *base, SimdFInt32 offset, SimdFloat *v, Targs... Fargs)
  93 {
  94     // For align 1 or 2: No multiplication of offset is needed
  95     if (align > 2)
  96     {
  97         offset = fastMultiply<align>(offset);
  98     }
  99     // For align 2: Scale of 2*sizeof(float) is used (maximum supported scale)
 100     constexpr int align_ = (align > 2) ? 1 : align;
 101     v->simdInternal_ = _mm512_i32gather_ps(offset.simdInternal_, base, sizeof(float)*align_);
 102     // Gather remaining elements. Avoid extra multiplication (new align is 1 or 2).
 103     gatherLoadBySimdIntTranspose<align_>(base+1, offset, Fargs ...);
 104 }
 105
 106 template <int align, typename ... Targs>
 107 static inline void gmx_simdcall
 108 gatherLoadUBySimdIntTranspose(const float *base, SimdFInt32 offset, Targs... Fargs)
 109 {
 110     gatherLoadBySimdIntTranspose<align>(base, offset, Fargs ...);
 111 }
 112
 113 template <int align, typename ... Targs>
 114 static inline void gmx_simdcall
 115 gatherLoadTranspose(const float *base, const std::int32_t offset[], Targs... Fargs)
 116 {
 117     gatherLoadBySimdIntTranspose<align>(base, simdLoad(offset, SimdFInt32Tag()), Fargs ...);
 118 }
 119
 120 template <int align, typename ... Targs>
 121 static inline void gmx_simdcall
 122 gatherLoadUTranspose(const float *base, const std::int32_t offset[], Targs... Fargs)
 123 {
 124     gatherLoadTranspose<align>(base, offset, Fargs ...);
 125 }
 126
 127 template <int align>
 128 static inline void gmx_simdcall
 129 transposeScatterStoreU(float *              base,
 130                        const std::int32_t   offset[],
 131                        SimdFloat            v0,
 132                        SimdFloat            v1,
 133                        SimdFloat            v2)
 134 {
 135     SimdFInt32 simdoffset = simdLoad(offset, SimdFInt32Tag());
 136     if (align > 2)
 137     {
 138         simdoffset = fastMultiply<align>(simdoffset);
 139     }
 140     constexpr size_t scale = (align > 2) ? sizeof(float) : sizeof(float) * align;
 141
 142     _mm512_i32scatter_ps(base,       simdoffset.simdInternal_, v0.simdInternal_, scale);
 143     _mm512_i32scatter_ps(&(base[1]), simdoffset.simdInternal_, v1.simdInternal_, scale);
 144     _mm512_i32scatter_ps(&(base[2]), simdoffset.simdInternal_, v2.simdInternal_, scale);
 145 }
 146
 147 template <int align>
 148 static inline void gmx_simdcall
 149 transposeScatterIncrU(float *              base,
 150                       const std::int32_t   offset[],
 151                       SimdFloat            v0,
 152                       SimdFloat            v1,
 153                       SimdFloat            v2)
 154 {
 155     __m512 t[4], t5, t6, t7, t8;
 156     int    i;
 157     alignas(GMX_SIMD_ALIGNMENT) std::int32_t    o[16];
 158     store(o, fastMultiply<align>(simdLoad(offset, SimdFInt32Tag())));
 159     if (align < 4)
 160     {
 161         t5   = _mm512_unpacklo_ps(v0.simdInternal_, v1.simdInternal_);
 162         t6   = _mm512_unpackhi_ps(v0.simdInternal_, v1.simdInternal_);
 163         t[0] = _mm512_shuffle_ps(t5, v2.simdInternal_, _MM_SHUFFLE(0, 0, 1, 0));
 164         t[1] = _mm512_shuffle_ps(t5, v2.simdInternal_, _MM_SHUFFLE(1, 1, 3, 2));
 165         t[2] = _mm512_shuffle_ps(t6, v2.simdInternal_, _MM_SHUFFLE(2, 2, 1, 0));
 166         t[3] = _mm512_shuffle_ps(t6, v2.simdInternal_, _MM_SHUFFLE(3, 3, 3, 2));
 167         for (i = 0; i < 4; i++)
 168         {
 169             _mm512_mask_storeu_ps(base + o[i], avx512Int2Mask(7), _mm512_castps128_ps512(
 170                                           _mm_add_ps(_mm_loadu_ps(base + o[i]), _mm512_castps512_ps128(t[i]))));
 171             _mm512_mask_storeu_ps(base + o[ 4 + i], avx512Int2Mask(7), _mm512_castps128_ps512(
 172                                           _mm_add_ps(_mm_loadu_ps(base + o[ 4 + i]), _mm512_extractf32x4_ps(t[i], 1))));
 173             _mm512_mask_storeu_ps(base + o[ 8 + i], avx512Int2Mask(7), _mm512_castps128_ps512(
 174                                           _mm_add_ps(_mm_loadu_ps(base + o[ 8 + i]), _mm512_extractf32x4_ps(t[i], 2))));
 175             _mm512_mask_storeu_ps(base + o[12 + i], avx512Int2Mask(7), _mm512_castps128_ps512(
 176                                           _mm_add_ps(_mm_loadu_ps(base + o[12 + i]), _mm512_extractf32x4_ps(t[i], 3))));
 177         }
 178     }
 179     else
 180     {
 181         //One could use shuffle here too if it is OK to overwrite the padded elements for alignment
 182         t5    = _mm512_unpacklo_ps(v0.simdInternal_, v2.simdInternal_);
 183         t6    = _mm512_unpackhi_ps(v0.simdInternal_, v2.simdInternal_);
 184         t7    = _mm512_unpacklo_ps(v1.simdInternal_, _mm512_setzero_ps());
 185         t8    = _mm512_unpackhi_ps(v1.simdInternal_, _mm512_setzero_ps());
 186         t[0]  = _mm512_unpacklo_ps(t5, t7);                             // x0 y0 z0  0 | x4 y4 z4 0
 187         t[1]  = _mm512_unpackhi_ps(t5, t7);                             // x1 y1 z1  0 | x5 y5 z5 0
 188         t[2]  = _mm512_unpacklo_ps(t6, t8);                             // x2 y2 z2  0 | x6 y6 z6 0
 189         t[3]  = _mm512_unpackhi_ps(t6, t8);                             // x3 y3 z3  0 | x7 y7 z7 0
 190         if (align % 4 == 0)
 191         {
 192             for (i = 0; i < 4; i++)
 193             {
 194                 _mm_store_ps(base + o[i], _mm_add_ps(_mm_load_ps(base + o[i]), _mm512_castps512_ps128(t[i])));
 195                 _mm_store_ps(base + o[ 4 + i],
 196                              _mm_add_ps(_mm_load_ps(base + o[ 4 + i]), _mm512_extractf32x4_ps(t[i], 1)));
 197                 _mm_store_ps(base + o[ 8 + i],
 198                              _mm_add_ps(_mm_load_ps(base + o[ 8 + i]), _mm512_extractf32x4_ps(t[i], 2)));
 199                 _mm_store_ps(base + o[12 + i],
 200                              _mm_add_ps(_mm_load_ps(base + o[12 + i]), _mm512_extractf32x4_ps(t[i], 3)));
 201             }
 202         }
 203         else
 204         {
 205             for (i = 0; i < 4; i++)
 206             {
 207                 _mm_storeu_ps(base + o[i], _mm_add_ps(_mm_loadu_ps(base + o[i]), _mm512_castps512_ps128(t[i])));
 208                 _mm_storeu_ps(base + o[ 4 + i],
 209                               _mm_add_ps(_mm_loadu_ps(base + o[ 4 + i]), _mm512_extractf32x4_ps(t[i], 1)));
 210                 _mm_storeu_ps(base + o[ 8 + i],
 211                               _mm_add_ps(_mm_loadu_ps(base + o[ 8 + i]), _mm512_extractf32x4_ps(t[i], 2)));
 212                 _mm_storeu_ps(base + o[12 + i],
 213                               _mm_add_ps(_mm_loadu_ps(base + o[12 + i]), _mm512_extractf32x4_ps(t[i], 3)));
 214             }
 215         }
 216     }
 217 }
 218
 219 template <int align>
 220 static inline void gmx_simdcall
 221 transposeScatterDecrU(float *              base,
 222                       const std::int32_t   offset[],
 223                       SimdFloat            v0,
 224                       SimdFloat            v1,
 225                       SimdFloat            v2)
 226 {
 227     __m512 t[4], t5, t6, t7, t8;
 228     int    i;
 229     alignas(GMX_SIMD_ALIGNMENT) std::int32_t    o[16];
 230     store(o, fastMultiply<align>(simdLoad(offset, SimdFInt32Tag())));
 231     if (align < 4)
 232     {
 233         t5   = _mm512_unpacklo_ps(v0.simdInternal_, v1.simdInternal_);
 234         t6   = _mm512_unpackhi_ps(v0.simdInternal_, v1.simdInternal_);
 235         t[0] = _mm512_shuffle_ps(t5, v2.simdInternal_, _MM_SHUFFLE(0, 0, 1, 0));
 236         t[1] = _mm512_shuffle_ps(t5, v2.simdInternal_, _MM_SHUFFLE(1, 1, 3, 2));
 237         t[2] = _mm512_shuffle_ps(t6, v2.simdInternal_, _MM_SHUFFLE(2, 2, 1, 0));
 238         t[3] = _mm512_shuffle_ps(t6, v2.simdInternal_, _MM_SHUFFLE(3, 3, 3, 2));
 239         for (i = 0; i < 4; i++)
 240         {
 241             _mm512_mask_storeu_ps(base + o[i], avx512Int2Mask(7), _mm512_castps128_ps512(
 242                                           _mm_sub_ps(_mm_loadu_ps(base + o[i]), _mm512_castps512_ps128(t[i]))));
 243             _mm512_mask_storeu_ps(base + o[ 4 + i], avx512Int2Mask(7), _mm512_castps128_ps512(
 244                                           _mm_sub_ps(_mm_loadu_ps(base + o[ 4 + i]), _mm512_extractf32x4_ps(t[i], 1))));
 245             _mm512_mask_storeu_ps(base + o[ 8 + i], avx512Int2Mask(7), _mm512_castps128_ps512(
 246                                           _mm_sub_ps(_mm_loadu_ps(base + o[ 8 + i]), _mm512_extractf32x4_ps(t[i], 2))));
 247             _mm512_mask_storeu_ps(base + o[12 + i], avx512Int2Mask(7), _mm512_castps128_ps512(
 248                                           _mm_sub_ps(_mm_loadu_ps(base + o[12 + i]), _mm512_extractf32x4_ps(t[i], 3))));
 249         }
 250     }
 251     else
 252     {
 253         //One could use shuffle here too if it is OK to overwrite the padded elements for alignment
 254         t5    = _mm512_unpacklo_ps(v0.simdInternal_, v2.simdInternal_);
 255         t6    = _mm512_unpackhi_ps(v0.simdInternal_, v2.simdInternal_);
 256         t7    = _mm512_unpacklo_ps(v1.simdInternal_, _mm512_setzero_ps());
 257         t8    = _mm512_unpackhi_ps(v1.simdInternal_, _mm512_setzero_ps());
 258         t[0]  = _mm512_unpacklo_ps(t5, t7);                             // x0 y0 z0  0 | x4 y4 z4 0
 259         t[1]  = _mm512_unpackhi_ps(t5, t7);                             // x1 y1 z1  0 | x5 y5 z5 0
 260         t[2]  = _mm512_unpacklo_ps(t6, t8);                             // x2 y2 z2  0 | x6 y6 z6 0
 261         t[3]  = _mm512_unpackhi_ps(t6, t8);                             // x3 y3 z3  0 | x7 y7 z7 0
 262         if (align % 4 == 0)
 263         {
 264             for (i = 0; i < 4; i++)
 265             {
 266                 _mm_store_ps(base + o[i], _mm_sub_ps(_mm_load_ps(base + o[i]), _mm512_castps512_ps128(t[i])));
 267                 _mm_store_ps(base + o[ 4 + i],
 268                              _mm_sub_ps(_mm_load_ps(base + o[ 4 + i]), _mm512_extractf32x4_ps(t[i], 1)));
 269                 _mm_store_ps(base + o[ 8 + i],
 270                              _mm_sub_ps(_mm_load_ps(base + o[ 8 + i]), _mm512_extractf32x4_ps(t[i], 2)));
 271                 _mm_store_ps(base + o[12 + i],
 272                              _mm_sub_ps(_mm_load_ps(base + o[12 + i]), _mm512_extractf32x4_ps(t[i], 3)));
 273             }
 274         }
 275         else
 276         {
 277             for (i = 0; i < 4; i++)
 278             {
 279                 _mm_storeu_ps(base + o[i], _mm_sub_ps(_mm_loadu_ps(base + o[i]), _mm512_castps512_ps128(t[i])));
 280                 _mm_storeu_ps(base + o[ 4 + i],
 281                               _mm_sub_ps(_mm_loadu_ps(base + o[ 4 + i]), _mm512_extractf32x4_ps(t[i], 1)));
 282                 _mm_storeu_ps(base + o[ 8 + i],
 283                               _mm_sub_ps(_mm_loadu_ps(base + o[ 8 + i]), _mm512_extractf32x4_ps(t[i], 2)));
 284                 _mm_storeu_ps(base + o[12 + i],
 285                               _mm_sub_ps(_mm_loadu_ps(base + o[12 + i]), _mm512_extractf32x4_ps(t[i], 3)));
 286             }
 287         }
 288     }
 289 }
 290
 291 static inline void gmx_simdcall
 292 expandScalarsToTriplets(SimdFloat    scalar,
 293                         SimdFloat *  triplets0,
 294                         SimdFloat *  triplets1,
 295                         SimdFloat *  triplets2)
 296 {
 297     triplets0->simdInternal_ = _mm512_permutexvar_ps(_mm512_set_epi32(5, 4, 4, 4, 3, 3, 3, 2, 2, 2, 1, 1, 1, 0, 0, 0),
 298                                                      scalar.simdInternal_);
 299     triplets1->simdInternal_ = _mm512_permutexvar_ps(_mm512_set_epi32(10, 10, 9, 9, 9, 8, 8, 8, 7, 7, 7, 6, 6, 6, 5, 5),
 300                                                      scalar.simdInternal_);
 301     triplets2->simdInternal_ = _mm512_permutexvar_ps(_mm512_set_epi32(15, 15, 15, 14, 14, 14, 13, 13, 13, 12, 12, 12, 11, 11, 11, 10),
 302                                                      scalar.simdInternal_);
 303 }
 304
 305
 306 static inline float gmx_simdcall
 307 reduceIncr4ReturnSum(float *    m,
 308                      SimdFloat  v0,
 309                      SimdFloat  v1,
 310                      SimdFloat  v2,
 311                      SimdFloat  v3)
 312 {
 313     __m512 t0, t1, t2;
 314     __m128 t3, t4;
 315
 316     assert(std::size_t(m) % 16 == 0);
 317
 318     t0 = _mm512_add_ps(v0.simdInternal_, _mm512_permute_ps(v0.simdInternal_, 0x4E));
 319     t0 = _mm512_mask_add_ps(t0, avx512Int2Mask(0xCCCC), v2.simdInternal_, _mm512_permute_ps(v2.simdInternal_, 0x4E));
 320     t1 = _mm512_add_ps(v1.simdInternal_, _mm512_permute_ps(v1.simdInternal_, 0x4E));
 321     t1 = _mm512_mask_add_ps(t1, avx512Int2Mask(0xCCCC), v3.simdInternal_, _mm512_permute_ps(v3.simdInternal_, 0x4E));
 322     t2 = _mm512_add_ps(t0, _mm512_permute_ps(t0, 0xB1));
 323     t2 = _mm512_mask_add_ps(t2, avx512Int2Mask(0xAAAA), t1, _mm512_permute_ps(t1, 0xB1));
 324
 325     t2 = _mm512_add_ps(t2, _mm512_shuffle_f32x4(t2, t2, 0x4E));
 326     t2 = _mm512_add_ps(t2, _mm512_shuffle_f32x4(t2, t2, 0xB1));
 327
 328     t3 = _mm512_castps512_ps128(t2);
 329     t4 = _mm_load_ps(m);
 330     t4 = _mm_add_ps(t4, t3);
 331     _mm_store_ps(m, t4);
 332
 333     t3 = _mm_add_ps(t3, _mm_permute_ps(t3, 0x4E));
 334     t3 = _mm_add_ps(t3, _mm_permute_ps(t3, 0xB1));
 335
 336     return _mm_cvtss_f32(t3);
 337
 338 }
 339
 340 static inline SimdFloat gmx_simdcall
 341 loadDualHsimd(const float * m0,
 342               const float * m1)
 343 {
 344     assert(std::size_t(m0) % 32 == 0);
 345     assert(std::size_t(m1) % 32 == 0);
 346
 347     return {
 348                _mm512_castpd_ps(_mm512_insertf64x4(_mm512_castpd256_pd512(_mm256_load_pd(reinterpret_cast<const double*>(m0))),
 349                                                    _mm256_load_pd(reinterpret_cast<const double*>(m1)), 1))
 350     };
 351 }
 352
 353 static inline SimdFloat gmx_simdcall
 354 loadDuplicateHsimd(const float * m)
 355 {
 356     assert(std::size_t(m) % 32 == 0);
 357     return {
 358                _mm512_castpd_ps(_mm512_broadcast_f64x4(_mm256_load_pd(reinterpret_cast<const double*>(m))))
 359     };
 360 }
 361
 362 static inline SimdFloat gmx_simdcall
 363 loadU1DualHsimd(const float * m)
 364 {
 365     return {
 366                _mm512_shuffle_f32x4(_mm512_broadcastss_ps(_mm_load_ss(m)),
 367                                     _mm512_broadcastss_ps(_mm_load_ss(m+1)), 0x44)
 368     };
 369 }
 370
 371
 372 static inline void gmx_simdcall
 373 storeDualHsimd(float *     m0,
 374                float *     m1,
 375                SimdFloat   a)
 376 {
 377     assert(std::size_t(m0) % 32 == 0);
 378     assert(std::size_t(m1) % 32 == 0);
 379
 380     _mm256_store_ps(m0, _mm512_castps512_ps256(a.simdInternal_));
 381     _mm256_store_pd(reinterpret_cast<double*>(m1), _mm512_extractf64x4_pd(_mm512_castps_pd(a.simdInternal_), 1));
 382 }
 383
 384 static inline void gmx_simdcall
 385 incrDualHsimd(float *     m0,
 386               float *     m1,
 387               SimdFloat   a)
 388 {
 389     assert(std::size_t(m0) % 32 == 0);
 390     assert(std::size_t(m1) % 32 == 0);
 391
 392     __m256 x;
 393
 394     // Lower half
 395     x = _mm256_load_ps(m0);
 396     x = _mm256_add_ps(x, _mm512_castps512_ps256(a.simdInternal_));
 397     _mm256_store_ps(m0, x);
 398
 399     // Upper half
 400     x = _mm256_load_ps(m1);
 401     x = _mm256_add_ps(x, _mm256_castpd_ps(_mm512_extractf64x4_pd(_mm512_castps_pd(a.simdInternal_), 1)));
 402     _mm256_store_ps(m1, x);
 403 }
 404
 405 static inline void gmx_simdcall
 406 decrHsimd(float *    m,
 407           SimdFloat  a)
 408 {
 409     __m256 t;
 410
 411     assert(std::size_t(m) % 32 == 0);
 412
 413     a.simdInternal_ = _mm512_add_ps(a.simdInternal_, _mm512_shuffle_f32x4(a.simdInternal_, a.simdInternal_, 0xEE));
 414     t               = _mm256_load_ps(m);
 415     t               = _mm256_sub_ps(t, _mm512_castps512_ps256(a.simdInternal_));
 416     _mm256_store_ps(m, t);
 417 }
 418
 419
 420 template <int align>
 421 static inline void gmx_simdcall
 422 gatherLoadTransposeHsimd(const float *        base0,
 423                          const float *        base1,
 424                          const std::int32_t   offset[],
 425                          SimdFloat *          v0,
 426                          SimdFloat *          v1)
 427 {
 428     __m256i idx;
 429     __m512  tmp1, tmp2;
 430
 431     assert(std::size_t(offset) % 32 == 0);
 432     assert(std::size_t(base0) % 8 == 0);
 433     assert(std::size_t(base1) % 8 == 0);
 434
 435     idx = _mm256_load_si256(reinterpret_cast<const __m256i*>(offset));
 436
 437     static_assert(align == 2 || align == 4, "If more are needed use fastMultiply");
 438     if (align == 4)
 439     {
 440         idx = _mm256_slli_epi32(idx, 1);
 441     }
 442
 443     tmp1 = _mm512_castpd_ps(_mm512_i32gather_pd(idx, reinterpret_cast<const double *>(base0), sizeof(double)));
 444     tmp2 = _mm512_castpd_ps(_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(_mm512_set_epi32(15, 13, 11, 9, 7, 5, 3, 1, 14, 12, 10, 8, 6, 4, 2, 0), v0->simdInternal_);
 450     v1->simdInternal_ = _mm512_permutexvar_ps(_mm512_set_epi32(15, 13, 11, 9, 7, 5, 3, 1, 14, 12, 10, 8, 6, 4, 2, 0), v1->simdInternal_);
 451 }
 452
 453 static inline float gmx_simdcall
 454 reduceIncr4ReturnSumHsimd(float *     m,
 455                           SimdFloat   v0,
 456                           SimdFloat   v1)
 457 {
 458     __m512 t0, t1;
 459     __m128 t2, t3;
 460
 461     assert(std::size_t(m) % 16 == 0);
 462
 463     t0 = _mm512_shuffle_f32x4(v0.simdInternal_, v1.simdInternal_, 0x88);
 464     t1 = _mm512_shuffle_f32x4(v0.simdInternal_, v1.simdInternal_, 0xDD);
 465     t0 = _mm512_add_ps(t0, t1);
 466     t0 = _mm512_add_ps(t0, _mm512_permute_ps(t0, 0x4E));
 467     t0 = _mm512_add_ps(t0, _mm512_permute_ps(t0, 0xB1));
 468     t0 = _mm512_maskz_compress_ps(avx512Int2Mask(0x1111), t0);
 469
 470     t3 = _mm512_castps512_ps128(t0);
 471     t2 = _mm_load_ps(m);
 472     t2 = _mm_add_ps(t2, t3);
 473     _mm_store_ps(m, t2);
 474
 475     t3 = _mm_add_ps(t3, _mm_permute_ps(t3, 0x4E));
 476     t3 = _mm_add_ps(t3, _mm_permute_ps(t3, 0xB1));
 477
 478     return _mm_cvtss_f32(t3);
 479 }
 480
 481 static inline SimdFloat gmx_simdcall
 482 loadUNDuplicate4(const float* f)
 483 {
 484     return {
 485                _mm512_permute_ps(_mm512_maskz_expandloadu_ps(0x1111, f), 0)
 486     };
 487 }
 488
 489 static inline SimdFloat gmx_simdcall
 490 load4DuplicateN(const float* f)
 491 {
 492     return {
 493                _mm512_broadcast_f32x4(_mm_load_ps(f))
 494     };
 495 }
 496
 497 static inline SimdFloat gmx_simdcall
 498 loadU4NOffset(const float* f, int offset)
 499 {
 500     const __m256i idx = _mm256_setr_epi32(0, 0, 1, 1, 2, 2, 3, 3);
 501     const __m256i gdx = _mm256_add_epi32(_mm256_setr_epi32(0, 2, 0, 2, 0, 2, 0, 2),
 502                                          _mm256_mullo_epi32(idx, _mm256_set1_epi32(offset)));
 503     return {
 504                _mm512_castpd_ps(_mm512_i32gather_pd(gdx, reinterpret_cast<const double*>(f), sizeof(float)))
 505     };
 506 }
 507
 508 }      // namespace gmx
 509
 510 #endif // GMX_SIMD_IMPL_X86_AVX_512_UTIL_FLOAT_H