src/gromacs/simd/impl_x86_mic/impl_x86_mic_util_double.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_MIC_UTIL_DOUBLE_H
  38 #define GMX_SIMD_IMPL_X86_MIC_UTIL_DOUBLE_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_mic_simd_double.h"
  50
  51 namespace gmx
  52 {
  53
  54 // On MIC it is better to use scatter operations, so we define the load routines
  55 // that use a SIMD offset variable first.
  56
  57 template<int align>
  58 static inline void gmx_simdcall gatherLoadBySimdIntTranspose(const double* base,
  59                                                              SimdDInt32    simdoffset,
  60                                                              SimdDouble*   v0,
  61                                                              SimdDouble*   v1,
  62                                                              SimdDouble*   v2,
  63                                                              SimdDouble*   v3)
  64 {
  65     assert((size_t)base % 32 == 0);
  66     assert(align % 4 == 0);
  67
  68     // All instructions might be latency ~4 on MIC, so we use shifts where we
  69     // only need a single instruction (since the shift parameter is an immediate),
  70     // but multiplication otherwise.
  71     if (align == 4)
  72     {
  73         simdoffset.simdInternal_ = _mm512_slli_epi32(simdoffset.simdInternal_, 2);
  74     }
  75     else if (align == 8)
  76     {
  77         simdoffset.simdInternal_ = _mm512_slli_epi32(simdoffset.simdInternal_, 3);
  78     }
  79     else
  80     {
  81         simdoffset = simdoffset * SimdDInt32(align);
  82     }
  83
  84     v0->simdInternal_ = _mm512_i32logather_pd(simdoffset.simdInternal_, base, sizeof(double));
  85     v1->simdInternal_ = _mm512_i32logather_pd(simdoffset.simdInternal_, base + 1, sizeof(double));
  86     v2->simdInternal_ = _mm512_i32logather_pd(simdoffset.simdInternal_, base + 2, sizeof(double));
  87     v3->simdInternal_ = _mm512_i32logather_pd(simdoffset.simdInternal_, base + 3, sizeof(double));
  88 }
  89
  90 template<int align>
  91 static inline void gmx_simdcall gatherLoadUBySimdIntTranspose(const double* base,
  92                                                               SimdDInt32    simdoffset,
  93                                                               SimdDouble*   v0,
  94                                                               SimdDouble*   v1)
  95 {
  96     // All instructions might be latency ~4 on MIC, so we use shifts where we
  97     // only need a single instruction (since the shift parameter is an immediate),
  98     // but multiplication otherwise.
  99     if (align == 2)
 100     {
 101         simdoffset.simdInternal_ = _mm512_slli_epi32(simdoffset.simdInternal_, 1);
 102     }
 103     else if (align == 4)
 104     {
 105         simdoffset.simdInternal_ = _mm512_slli_epi32(simdoffset.simdInternal_, 2);
 106     }
 107     else if (align == 8)
 108     {
 109         simdoffset.simdInternal_ = _mm512_slli_epi32(simdoffset.simdInternal_, 3);
 110     }
 111     else
 112     {
 113         simdoffset = simdoffset * SimdDInt32(align);
 114     }
 115
 116     v0->simdInternal_ = _mm512_i32logather_pd(simdoffset.simdInternal_, base, sizeof(double));
 117     v1->simdInternal_ = _mm512_i32logather_pd(simdoffset.simdInternal_, base + 1, sizeof(double));
 118 }
 119
 120 template<int align>
 121 static inline void gmx_simdcall gatherLoadBySimdIntTranspose(const double* base,
 122                                                              SimdDInt32    simdoffset,
 123                                                              SimdDouble*   v0,
 124                                                              SimdDouble*   v1)
 125 {
 126     assert(std::size_t(base) % 16 == 0);
 127     assert(align % 2 == 0);
 128     gatherLoadUBySimdIntTranspose<align>(base, simdoffset, v0, v1);
 129 }
 130
 131
 132 template<int align>
 133 static inline void gmx_simdcall gatherLoadTranspose(const double*      base,
 134                                                     const std::int32_t offset[],
 135                                                     SimdDouble*        v0,
 136                                                     SimdDouble*        v1,
 137                                                     SimdDouble*        v2,
 138                                                     SimdDouble*        v3)
 139 {
 140     gatherLoadBySimdIntTranspose<align>(base, simdLoad(offset, SimdDInt32Tag()), v0, v1, v2, v3);
 141 }
 142
 143 template<int align>
 144 static inline void gmx_simdcall
 145                    gatherLoadTranspose(const double* base, const std::int32_t offset[], SimdDouble* v0, SimdDouble* v1)
 146 {
 147     gatherLoadBySimdIntTranspose<align>(base, simdLoad(offset, SimdDInt32Tag()), v0, v1);
 148 }
 149
 150 static const int c_simdBestPairAlignmentDouble = 2;
 151
 152 template<int align>
 153 static inline void gmx_simdcall gatherLoadUTranspose(const double*      base,
 154                                                      const std::int32_t offset[],
 155                                                      SimdDouble*        v0,
 156                                                      SimdDouble*        v1,
 157                                                      SimdDouble*        v2)
 158 {
 159     SimdDInt32 simdoffset;
 160
 161     assert(std::size_t(offset) % 32 == 0);
 162
 163     simdoffset = simdLoad(offset, SimdDInt32Tag());
 164
 165     // All instructions might be latency ~4 on MIC, so we use shifts where we
 166     // only need a single instruction (since the shift parameter is an immediate),
 167     // but multiplication otherwise.
 168     if (align == 4)
 169     {
 170         simdoffset.simdInternal_ = _mm512_slli_epi32(simdoffset.simdInternal_, 2);
 171     }
 172     else if (align == 8)
 173     {
 174         simdoffset.simdInternal_ = _mm512_slli_epi32(simdoffset.simdInternal_, 3);
 175     }
 176     else
 177     {
 178         simdoffset = simdoffset * SimdDInt32(align);
 179     }
 180
 181     v0->simdInternal_ = _mm512_i32logather_pd(simdoffset.simdInternal_, base, sizeof(double));
 182     v1->simdInternal_ = _mm512_i32logather_pd(simdoffset.simdInternal_, base + 1, sizeof(double));
 183     v2->simdInternal_ = _mm512_i32logather_pd(simdoffset.simdInternal_, base + 2, sizeof(double));
 184 }
 185
 186 template<int align>
 187 static inline void gmx_simdcall transposeScatterStoreU(double*            base,
 188                                                        const std::int32_t offset[],
 189                                                        SimdDouble         v0,
 190                                                        SimdDouble         v1,
 191                                                        SimdDouble         v2)
 192 {
 193     SimdDInt32 simdoffset;
 194
 195     assert(std::size_t(offset) % 32 == 0);
 196
 197     simdoffset = simdLoad(offset, SimdDInt32Tag());
 198
 199     // All instructions might be latency ~4 on MIC, so we use shifts where we
 200     // only need a single instruction (since the shift parameter is an immediate),
 201     // but multiplication otherwise.
 202     if (align == 4)
 203     {
 204         simdoffset.simdInternal_ = _mm512_slli_epi32(simdoffset.simdInternal_, 2);
 205     }
 206     else if (align == 8)
 207     {
 208         simdoffset.simdInternal_ = _mm512_slli_epi32(simdoffset.simdInternal_, 3);
 209     }
 210     else
 211     {
 212         simdoffset = simdoffset * SimdDInt32(align);
 213     }
 214
 215     _mm512_i32loscatter_pd(base, simdoffset.simdInternal_, v0.simdInternal_, sizeof(double));
 216     _mm512_i32loscatter_pd(base + 1, simdoffset.simdInternal_, v1.simdInternal_, sizeof(double));
 217     _mm512_i32loscatter_pd(base + 2, simdoffset.simdInternal_, v2.simdInternal_, sizeof(double));
 218 }
 219
 220 template<int align>
 221 static inline void gmx_simdcall
 222                    transposeScatterIncrU(double* base, const std::int32_t offset[], SimdDouble v0, SimdDouble v1, SimdDouble v2)
 223 {
 224     alignas(GMX_SIMD_ALIGNMENT) double rdata0[GMX_SIMD_DOUBLE_WIDTH];
 225     alignas(GMX_SIMD_ALIGNMENT) double rdata1[GMX_SIMD_DOUBLE_WIDTH];
 226     alignas(GMX_SIMD_ALIGNMENT) double rdata2[GMX_SIMD_DOUBLE_WIDTH];
 227
 228     store(rdata0, v0);
 229     store(rdata1, v1);
 230     store(rdata2, v2);
 231
 232     for (int i = 0; i < GMX_SIMD_DOUBLE_WIDTH; i++)
 233     {
 234         base[align * offset[i] + 0] += rdata0[i];
 235         base[align * offset[i] + 1] += rdata1[i];
 236         base[align * offset[i] + 2] += rdata2[i];
 237     }
 238 }
 239
 240 template<int align>
 241 static inline void gmx_simdcall
 242                    transposeScatterDecrU(double* base, const std::int32_t offset[], SimdDouble v0, SimdDouble v1, SimdDouble v2)
 243 {
 244     alignas(GMX_SIMD_ALIGNMENT) double rdata0[GMX_SIMD_DOUBLE_WIDTH];
 245     alignas(GMX_SIMD_ALIGNMENT) double rdata1[GMX_SIMD_DOUBLE_WIDTH];
 246     alignas(GMX_SIMD_ALIGNMENT) double rdata2[GMX_SIMD_DOUBLE_WIDTH];
 247
 248     store(rdata0, v0);
 249     store(rdata1, v1);
 250     store(rdata2, v2);
 251
 252     for (int i = 0; i < GMX_SIMD_DOUBLE_WIDTH; i++)
 253     {
 254         base[align * offset[i] + 0] -= rdata0[i];
 255         base[align * offset[i] + 1] -= rdata1[i];
 256         base[align * offset[i] + 2] -= rdata2[i];
 257     }
 258 }
 259
 260 static inline void gmx_simdcall expandScalarsToTriplets(SimdDouble  scalar,
 261                                                         SimdDouble* triplets0,
 262                                                         SimdDouble* triplets1,
 263                                                         SimdDouble* triplets2)
 264 {
 265     triplets0->simdInternal_ = _mm512_castsi512_pd(
 266             _mm512_permutevar_epi32(_mm512_set_epi32(5, 4, 5, 4, 3, 2, 3, 2, 3, 2, 1, 0, 1, 0, 1, 0),
 267                                     _mm512_castpd_si512(scalar.simdInternal_)));
 268     triplets1->simdInternal_ = _mm512_castsi512_pd(_mm512_permutevar_epi32(
 269             _mm512_set_epi32(11, 10, 9, 8, 9, 8, 9, 8, 7, 6, 7, 6, 7, 6, 5, 4),
 270             _mm512_castpd_si512(scalar.simdInternal_)));
 271     triplets2->simdInternal_ = _mm512_castsi512_pd(_mm512_permutevar_epi32(
 272             _mm512_set_epi32(15, 14, 15, 14, 15, 14, 13, 12, 13, 12, 13, 12, 11, 10, 11, 10),
 273             _mm512_castpd_si512(scalar.simdInternal_)));
 274 }
 275
 276
 277 static inline double gmx_simdcall
 278                      reduceIncr4ReturnSum(double* m, SimdDouble v0, SimdDouble v1, SimdDouble v2, SimdDouble v3)
 279 {
 280     double  d;
 281     __m512d t0, t1, t2, t3;
 282
 283     assert(std::size_t(m) % 32 == 0);
 284
 285     t0 = _mm512_swizzle_pd(_mm512_mask_blend_pd(_mm512_int2mask(0x33), v0.simdInternal_, v2.simdInternal_),
 286                            _MM_SWIZ_REG_BADC);
 287     t2 = _mm512_mask_blend_pd(_mm512_int2mask(0x33), v2.simdInternal_, v0.simdInternal_);
 288     t1 = _mm512_swizzle_pd(_mm512_mask_blend_pd(_mm512_int2mask(0x33), v1.simdInternal_, v3.simdInternal_),
 289                            _MM_SWIZ_REG_BADC);
 290     t3 = _mm512_mask_blend_pd(_mm512_int2mask(0x33), v3.simdInternal_, v1.simdInternal_);
 291     t0 = _mm512_add_pd(t0, t2);
 292     t1 = _mm512_add_pd(t1, t3);
 293
 294     t2 = _mm512_swizzle_pd(_mm512_mask_blend_pd(_mm512_int2mask(0b01010101), t0, t1), _MM_SWIZ_REG_CDAB);
 295     t3 = _mm512_mask_blend_pd(_mm512_int2mask(0b01010101), t1, t0);
 296     t2 = _mm512_add_pd(t2, t3);
 297
 298     t2 = _mm512_add_pd(t2, _mm512_castps_pd(_mm512_permute4f128_ps(_mm512_castpd_ps(t2), _MM_PERM_BADC)));
 299
 300     t0 = _mm512_mask_extload_pd(_mm512_undefined_pd(), _mm512_int2mask(0xF), m, _MM_UPCONV_PD_NONE,
 301                                 _MM_BROADCAST_4X8, _MM_HINT_NONE);
 302     t0 = _mm512_add_pd(t0, t2);
 303     _mm512_mask_packstorelo_pd(m, _mm512_int2mask(0xF), t0);
 304
 305     t2 = _mm512_add_pd(t2, _mm512_swizzle_pd(t2, _MM_SWIZ_REG_BADC));
 306     t2 = _mm512_add_pd(t2, _mm512_swizzle_pd(t2, _MM_SWIZ_REG_CDAB));
 307
 308     _mm512_mask_packstorelo_pd(&d, _mm512_mask2int(0x01), t2);
 309     return d;
 310 }
 311
 312 static inline SimdDouble gmx_simdcall loadDualHsimd(const double* m0, const double* m1)
 313 {
 314     assert(std::size_t(m0) % 32 == 0);
 315     assert(std::size_t(m1) % 32 == 0);
 316
 317     return _mm512_mask_extload_pd(
 318             _mm512_extload_pd(m0, _MM_UPCONV_PD_NONE, _MM_BROADCAST_4X8, _MM_HINT_NONE),
 319             _mm512_int2mask(0xF0), m1, _MM_UPCONV_PD_NONE, _MM_BROADCAST_4X8, _MM_HINT_NONE);
 320 }
 321
 322 static inline SimdDouble gmx_simdcall loadDuplicateHsimd(const double* m)
 323 {
 324     assert(std::size_t(m) % 32 == 0);
 325
 326     return _mm512_extload_pd(m, _MM_UPCONV_PD_NONE, _MM_BROADCAST_4X8, _MM_HINT_NONE);
 327 }
 328
 329 static inline SimdDouble gmx_simdcall loadU1DualHsimd(const double* m)
 330 {
 331     return _mm512_mask_extload_pd(
 332             _mm512_extload_pd(m, _MM_UPCONV_PD_NONE, _MM_BROADCAST_1X8, _MM_HINT_NONE),
 333             _mm512_int2mask(0xF0), m + 1, _MM_UPCONV_PD_NONE, _MM_BROADCAST_1X8, _MM_HINT_NONE);
 334 }
 335
 336
 337 static inline void gmx_simdcall storeDualHsimd(double* m0, double* m1, SimdDouble a)
 338 {
 339     assert(std::size_t(m0) % 32 == 0);
 340     assert(std::size_t(m1) % 32 == 0);
 341
 342     _mm512_mask_packstorelo_pd(m0, _mm512_int2mask(0x0F), a.simdInternal_);
 343     _mm512_mask_packstorelo_pd(m1, _mm512_int2mask(0xF0), a.simdInternal_);
 344 }
 345
 346 static inline void gmx_simdcall incrDualHsimd(double* m0, double* m1, SimdDouble a)
 347 {
 348     assert(std::size_t(m0) % 32 == 0);
 349     assert(std::size_t(m1) % 32 == 0);
 350
 351     __m512d x;
 352
 353     // Update lower half
 354     x = _mm512_extload_pd(m0, _MM_UPCONV_PD_NONE, _MM_BROADCAST_4X8, _MM_HINT_NONE);
 355     x = _mm512_add_pd(x, a.simdInternal_);
 356     _mm512_mask_packstorelo_pd(m0, _mm512_int2mask(0x0F), x);
 357
 358     // Update upper half
 359     x = _mm512_extload_pd(m1, _MM_UPCONV_PD_NONE, _MM_BROADCAST_4X8, _MM_HINT_NONE);
 360     x = _mm512_add_pd(x, a.simdInternal_);
 361     _mm512_mask_packstorelo_pd(m1, _mm512_int2mask(0xF0), x);
 362 }
 363
 364 static inline void gmx_simdcall decrHsimd(double* m, SimdDouble a)
 365 {
 366     __m512d t;
 367
 368     assert(std::size_t(m) % 32 == 0);
 369
 370     t               = _mm512_extload_pd(m, _MM_UPCONV_PD_NONE, _MM_BROADCAST_4X8, _MM_HINT_NONE);
 371     a.simdInternal_ = _mm512_add_pd(
 372             a.simdInternal_,
 373             _mm512_castps_pd(_mm512_permute4f128_ps(_mm512_castpd_ps(a.simdInternal_), _MM_PERM_BADC)));
 374     t = _mm512_sub_pd(t, a.simdInternal_);
 375     _mm512_mask_packstorelo_pd(m, _mm512_int2mask(0x0F), t);
 376 }
 377
 378
 379 template<int align>
 380 static inline void gmx_simdcall gatherLoadTransposeHsimd(const double*      base0,
 381                                                          const double*      base1,
 382                                                          const std::int32_t offset[],
 383                                                          SimdDouble*        v0,
 384                                                          SimdDouble*        v1)
 385 {
 386     __m512i idx0, idx1, idx;
 387     __m512d tmp1, tmp2;
 388
 389     assert(std::size_t(offset) % 16 == 0);
 390     assert(std::size_t(base0) % 16 == 0);
 391     assert(std::size_t(base1) % 16 == 0);
 392     assert(std::size_t(align) % 2 == 0);
 393
 394     idx0 = _mm512_extload_epi32(offset, _MM_UPCONV_EPI32_NONE, _MM_BROADCAST_4X16, _MM_HINT_NONE);
 395
 396     idx0 = _mm512_mullo_epi32(idx0, _mm512_set1_epi32(align));
 397     idx1 = _mm512_add_epi32(idx0, _mm512_set1_epi32(1));
 398
 399     idx = _mm512_mask_permute4f128_epi32(idx0, _mm512_int2mask(0x00F0), idx1, _MM_PERM_AAAA);
 400
 401     tmp1 = _mm512_i32logather_pd(idx, base0, sizeof(double));
 402     tmp2 = _mm512_i32logather_pd(idx, base1, sizeof(double));
 403
 404     v0->simdInternal_ = _mm512_castps_pd(_mm512_mask_permute4f128_ps(
 405             _mm512_castpd_ps(tmp1), _mm512_int2mask(0xFF00), _mm512_castpd_ps(tmp2), _MM_PERM_BABA));
 406     v1->simdInternal_ = _mm512_castps_pd(_mm512_mask_permute4f128_ps(
 407             _mm512_castpd_ps(tmp2), _mm512_int2mask(0x00FF), _mm512_castpd_ps(tmp1), _MM_PERM_DCDC));
 408 }
 409
 410 static inline double gmx_simdcall reduceIncr4ReturnSumHsimd(double* m, SimdDouble v0, SimdDouble v1)
 411 {
 412     double  d;
 413     __m512d t0, t1;
 414
 415     assert(std::size_t(m) % 32 == 0);
 416
 417     t0 = _mm512_add_pd(v0.simdInternal_, _mm512_swizzle_pd(v0.simdInternal_, _MM_SWIZ_REG_BADC));
 418     t0 = _mm512_mask_add_pd(t0, _mm512_int2mask(0xCC), v1.simdInternal_,
 419                             _mm512_swizzle_pd(v1.simdInternal_, _MM_SWIZ_REG_BADC));
 420     t0 = _mm512_add_pd(t0, _mm512_swizzle_pd(t0, _MM_SWIZ_REG_CDAB));
 421     t0 = _mm512_castps_pd(_mm512_mask_permute4f128_ps(_mm512_castpd_ps(t0), _mm512_int2mask(0xCCCC),
 422                                                       _mm512_castpd_ps(t0), _MM_PERM_DCDC));
 423
 424     t1 = _mm512_mask_extload_pd(_mm512_undefined_pd(), _mm512_int2mask(0xF), m, _MM_UPCONV_PD_NONE,
 425                                 _MM_BROADCAST_4X8, _MM_HINT_NONE);
 426     t1 = _mm512_add_pd(t1, t0);
 427     _mm512_mask_packstorelo_pd(m, _mm512_int2mask(0xF), t1);
 428
 429     t0 = _mm512_add_pd(t0, _mm512_swizzle_pd(t0, _MM_SWIZ_REG_BADC));
 430     t0 = _mm512_add_pd(t0, _mm512_swizzle_pd(t0, _MM_SWIZ_REG_CDAB));
 431
 432     _mm512_mask_packstorelo_pd(&d, _mm512_mask2int(0x03), t0);
 433     return d;
 434 }
 435
 436 } // namespace gmx
 437
 438 #endif // GMX_SIMD_IMPL_X86_MIC_UTIL_DOUBLE_H