src/gromacs/simd/impl_x86_mic/impl_x86_mic_simd4_float.h

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2014,2015,2019, 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_MIC_SIMD4_FLOAT_H
  37 #define GMX_SIMD_IMPL_X86_MIC_SIMD4_FLOAT_H
  38
  39 #include "config.h"
  40
  41 #include <cassert>
  42
  43 #include <immintrin.h>
  44
  45 #include "gromacs/utility/basedefinitions.h"
  46
  47 #include "impl_x86_mic_simd_float.h"
  48
  49 namespace gmx
  50 {
  51
  52 class Simd4Float
  53 {
  54 public:
  55     Simd4Float() {}
  56
  57     Simd4Float(float f) : simdInternal_(_mm512_set1_ps(f)) {}
  58
  59     // Internal utility constructor to simplify return statements
  60     Simd4Float(__m512 simd) : simdInternal_(simd) {}
  61
  62     __m512 simdInternal_;
  63 };
  64
  65 class Simd4FBool
  66 {
  67 public:
  68     Simd4FBool() {}
  69
  70     // Internal utility constructor to simplify return statements
  71     Simd4FBool(__mmask16 simd) : simdInternal_(simd) {}
  72
  73     __mmask16 simdInternal_;
  74 };
  75
  76 static inline Simd4Float gmx_simdcall load4(const float* m)
  77 {
  78     assert(size_t(m) % 16 == 0);
  79     return { _mm512_mask_extload_ps(_mm512_undefined_ps(), _mm512_int2mask(0xF), m,
  80                                     _MM_UPCONV_PS_NONE, _MM_BROADCAST_4X16, _MM_HINT_NONE) };
  81 }
  82
  83 static inline void gmx_simdcall store4(float* m, Simd4Float a)
  84 {
  85     assert(size_t(m) % 16 == 0);
  86     _mm512_mask_packstorelo_ps(m, _mm512_int2mask(0xF), a.simdInternal_);
  87 }
  88
  89 static inline Simd4Float gmx_simdcall load4U(const float* m)
  90 {
  91     return { _mm512_mask_loadunpackhi_ps(
  92             _mm512_mask_loadunpacklo_ps(_mm512_undefined_ps(), _mm512_int2mask(0xF), m),
  93             _mm512_int2mask(0xF), m + 16) };
  94 }
  95
  96 static inline void gmx_simdcall store4U(float* m, Simd4Float a)
  97 {
  98     _mm512_mask_packstorelo_ps(m, _mm512_int2mask(0xF), a.simdInternal_);
  99     _mm512_mask_packstorehi_ps(m + 16, _mm512_int2mask(0xF), a.simdInternal_);
 100 }
 101
 102 static inline Simd4Float gmx_simdcall simd4SetZeroF()
 103 {
 104     return { _mm512_setzero_ps() };
 105 }
 106
 107 static inline Simd4Float gmx_simdcall operator&(Simd4Float a, Simd4Float b)
 108 {
 109     return { _mm512_castsi512_ps(_mm512_mask_and_epi32(_mm512_undefined_epi32(), _mm512_int2mask(0xF),
 110                                                        _mm512_castps_si512(a.simdInternal_),
 111                                                        _mm512_castps_si512(b.simdInternal_))) };
 112 }
 113
 114 static inline Simd4Float gmx_simdcall andNot(Simd4Float a, Simd4Float b)
 115 {
 116     return { _mm512_castsi512_ps(_mm512_mask_andnot_epi32(
 117             _mm512_undefined_epi32(), _mm512_int2mask(0xF), _mm512_castps_si512(a.simdInternal_),
 118             _mm512_castps_si512(b.simdInternal_))) };
 119 }
 120
 121 static inline Simd4Float gmx_simdcall operator|(Simd4Float a, Simd4Float b)
 122 {
 123     return { _mm512_castsi512_ps(_mm512_mask_or_epi32(_mm512_undefined_epi32(), _mm512_int2mask(0xF),
 124                                                       _mm512_castps_si512(a.simdInternal_),
 125                                                       _mm512_castps_si512(b.simdInternal_))) };
 126 }
 127
 128 static inline Simd4Float gmx_simdcall operator^(Simd4Float a, Simd4Float b)
 129 {
 130     return { _mm512_castsi512_ps(_mm512_mask_xor_epi32(_mm512_undefined_epi32(), _mm512_int2mask(0xF),
 131                                                        _mm512_castps_si512(a.simdInternal_),
 132                                                        _mm512_castps_si512(b.simdInternal_))) };
 133 }
 134
 135 static inline Simd4Float gmx_simdcall operator+(Simd4Float a, Simd4Float b)
 136 {
 137     return { _mm512_mask_add_ps(_mm512_undefined_ps(), _mm512_int2mask(0xF), a.simdInternal_,
 138                                 b.simdInternal_) };
 139 }
 140
 141 static inline Simd4Float gmx_simdcall operator-(Simd4Float a, Simd4Float b)
 142 {
 143     return { _mm512_mask_sub_ps(_mm512_undefined_ps(), _mm512_int2mask(0xF), a.simdInternal_,
 144                                 b.simdInternal_) };
 145 }
 146
 147 static inline Simd4Float gmx_simdcall operator-(Simd4Float x)
 148 {
 149     return { _mm512_mask_addn_ps(_mm512_undefined_ps(), _mm512_int2mask(0xF), x.simdInternal_,
 150                                  _mm512_setzero_ps()) };
 151 }
 152
 153 static inline Simd4Float gmx_simdcall operator*(Simd4Float a, Simd4Float b)
 154 {
 155     return { _mm512_mask_mul_ps(_mm512_undefined_ps(), _mm512_int2mask(0xF), a.simdInternal_,
 156                                 b.simdInternal_) };
 157 }
 158
 159 static inline Simd4Float gmx_simdcall fma(Simd4Float a, Simd4Float b, Simd4Float c)
 160 {
 161     return { _mm512_mask_fmadd_ps(a.simdInternal_, _mm512_int2mask(0xF), b.simdInternal_, c.simdInternal_) };
 162 }
 163
 164 static inline Simd4Float gmx_simdcall fms(Simd4Float a, Simd4Float b, Simd4Float c)
 165 {
 166     return { _mm512_mask_fmsub_ps(a.simdInternal_, _mm512_int2mask(0xF), b.simdInternal_, c.simdInternal_) };
 167 }
 168
 169 static inline Simd4Float gmx_simdcall fnma(Simd4Float a, Simd4Float b, Simd4Float c)
 170 {
 171     return { _mm512_mask_fnmadd_ps(a.simdInternal_, _mm512_int2mask(0xF), b.simdInternal_, c.simdInternal_) };
 172 }
 173
 174 static inline Simd4Float gmx_simdcall fnms(Simd4Float a, Simd4Float b, Simd4Float c)
 175 {
 176     return { _mm512_mask_fnmsub_ps(a.simdInternal_, _mm512_int2mask(0xF), b.simdInternal_, c.simdInternal_) };
 177 }
 178
 179 static inline Simd4Float gmx_simdcall rsqrt(Simd4Float x)
 180 {
 181     return { _mm512_mask_rsqrt23_ps(_mm512_undefined_ps(), _mm512_int2mask(0xF), x.simdInternal_) };
 182 }
 183
 184 static inline Simd4Float gmx_simdcall abs(Simd4Float x)
 185 {
 186     return { _mm512_castsi512_ps(_mm512_mask_andnot_epi32(
 187             _mm512_undefined_epi32(), _mm512_int2mask(0xF),
 188             _mm512_castps_si512(_mm512_set1_ps(GMX_FLOAT_NEGZERO)), _mm512_castps_si512(x.simdInternal_))) };
 189 }
 190
 191 static inline Simd4Float gmx_simdcall max(Simd4Float a, Simd4Float b)
 192 {
 193     return { _mm512_mask_gmax_ps(_mm512_undefined_ps(), _mm512_int2mask(0xF), a.simdInternal_,
 194                                  b.simdInternal_) };
 195 }
 196
 197 static inline Simd4Float gmx_simdcall min(Simd4Float a, Simd4Float b)
 198 {
 199     return { _mm512_mask_gmin_ps(_mm512_undefined_ps(), _mm512_int2mask(0xF), a.simdInternal_,
 200                                  b.simdInternal_) };
 201 }
 202
 203 static inline Simd4Float gmx_simdcall round(Simd4Float x)
 204 {
 205     return { _mm512_mask_round_ps(_mm512_undefined_ps(), _mm512_int2mask(0xF), x.simdInternal_,
 206                                   _MM_FROUND_TO_NEAREST_INT, _MM_EXPADJ_NONE) };
 207 }
 208
 209 static inline Simd4Float gmx_simdcall trunc(Simd4Float x)
 210 {
 211     return { _mm512_mask_round_ps(_mm512_undefined_ps(), _mm512_int2mask(0xF), x.simdInternal_,
 212                                   _MM_FROUND_TO_ZERO, _MM_EXPADJ_NONE) };
 213 }
 214
 215 static inline float gmx_simdcall dotProduct(Simd4Float a, Simd4Float b)
 216 {
 217     __m512 x = _mm512_mask_mul_ps(_mm512_setzero_ps(), _mm512_int2mask(0x7), a.simdInternal_,
 218                                   b.simdInternal_);
 219     x        = _mm512_add_ps(x, _mm512_swizzle_ps(x, _MM_SWIZ_REG_BADC));
 220     x        = _mm512_add_ps(x, _mm512_swizzle_ps(x, _MM_SWIZ_REG_CDAB));
 221     float f;
 222     _mm512_mask_packstorelo_ps(&f, _mm512_mask2int(0x1), x);
 223     return f;
 224 }
 225
 226 static inline void gmx_simdcall transpose(Simd4Float* v0, Simd4Float* v1, Simd4Float* v2, Simd4Float* v3)
 227 {
 228     v0->simdInternal_ = _mm512_mask_permute4f128_ps(v0->simdInternal_, _mm512_int2mask(0x00F0),
 229                                                     v1->simdInternal_, _MM_PERM_AAAA);
 230     v2->simdInternal_ = _mm512_mask_permute4f128_ps(v2->simdInternal_, _mm512_int2mask(0x00F0),
 231                                                     v3->simdInternal_, _MM_PERM_AAAA);
 232     v0->simdInternal_ = _mm512_mask_permute4f128_ps(v0->simdInternal_, _mm512_int2mask(0xFF00),
 233                                                     v2->simdInternal_, _MM_PERM_BABA);
 234     v0->simdInternal_ = _mm512_castsi512_ps(_mm512_permutevar_epi32(
 235             _mm512_set_epi32(15, 11, 7, 3, 14, 10, 6, 2, 13, 9, 5, 1, 12, 8, 4, 0),
 236             _mm512_castps_si512(v0->simdInternal_)));
 237     v1->simdInternal_ = _mm512_mask_permute4f128_ps(_mm512_setzero_ps(), _mm512_int2mask(0x000F),
 238                                                     v0->simdInternal_, _MM_PERM_BBBB);
 239     v2->simdInternal_ = _mm512_mask_permute4f128_ps(_mm512_setzero_ps(), _mm512_int2mask(0x000F),
 240                                                     v0->simdInternal_, _MM_PERM_CCCC);
 241     v3->simdInternal_ = _mm512_mask_permute4f128_ps(_mm512_setzero_ps(), _mm512_int2mask(0x000F),
 242                                                     v0->simdInternal_, _MM_PERM_DDDD);
 243 }
 244
 245 // Picky, picky, picky:
 246 // icc-16 complains about "Illegal value of immediate argument to intrinsic"
 247 // unless we use
 248 // 1) Ordered-quiet for ==
 249 // 2) Unordered-quiet for !=
 250 // 3) Ordered-signaling for < and <=
 251
 252 static inline Simd4FBool gmx_simdcall operator==(Simd4Float a, Simd4Float b)
 253 {
 254     return { _mm512_mask_cmp_ps_mask(_mm512_int2mask(0xF), a.simdInternal_, b.simdInternal_, _CMP_EQ_OQ) };
 255 }
 256
 257 static inline Simd4FBool gmx_simdcall operator!=(Simd4Float a, Simd4Float b)
 258 {
 259     return { _mm512_mask_cmp_ps_mask(_mm512_int2mask(0xF), a.simdInternal_, b.simdInternal_, _CMP_NEQ_UQ) };
 260 }
 261
 262 static inline Simd4FBool gmx_simdcall operator<(Simd4Float a, Simd4Float b)
 263 {
 264     return { _mm512_mask_cmp_ps_mask(_mm512_int2mask(0xF), a.simdInternal_, b.simdInternal_, _CMP_LT_OS) };
 265 }
 266
 267 static inline Simd4FBool gmx_simdcall operator<=(Simd4Float a, Simd4Float b)
 268 {
 269     return { _mm512_mask_cmp_ps_mask(_mm512_int2mask(0xF), a.simdInternal_, b.simdInternal_, _CMP_LE_OS) };
 270 }
 271
 272 static inline Simd4FBool gmx_simdcall operator&&(Simd4FBool a, Simd4FBool b)
 273 {
 274     return { _mm512_kand(a.simdInternal_, b.simdInternal_) };
 275 }
 276
 277 static inline Simd4FBool gmx_simdcall operator||(Simd4FBool a, Simd4FBool b)
 278 {
 279     return { _mm512_kor(a.simdInternal_, b.simdInternal_) };
 280 }
 281
 282 static inline bool gmx_simdcall anyTrue(Simd4FBool a)
 283 {
 284     return (_mm512_mask2int(a.simdInternal_) & 0xF) != 0;
 285 }
 286
 287 static inline Simd4Float gmx_simdcall selectByMask(Simd4Float a, Simd4FBool m)
 288 {
 289     return { _mm512_mask_mov_ps(_mm512_setzero_ps(), m.simdInternal_, a.simdInternal_) };
 290 }
 291
 292 static inline Simd4Float gmx_simdcall selectByNotMask(Simd4Float a, Simd4FBool m)
 293 {
 294     return { _mm512_mask_mov_ps(_mm512_setzero_ps(), _mm512_knot(m.simdInternal_), a.simdInternal_) };
 295 }
 296
 297 static inline Simd4Float gmx_simdcall blend(Simd4Float a, Simd4Float b, Simd4FBool sel)
 298 {
 299     return { _mm512_mask_blend_ps(sel.simdInternal_, a.simdInternal_, b.simdInternal_) };
 300 }
 301
 302 static inline float gmx_simdcall reduce(Simd4Float a)
 303 {
 304     __m512 x = a.simdInternal_;
 305     x        = _mm512_add_ps(x, _mm512_swizzle_ps(x, _MM_SWIZ_REG_BADC));
 306     x        = _mm512_add_ps(x, _mm512_swizzle_ps(x, _MM_SWIZ_REG_CDAB));
 307     float f;
 308     _mm512_mask_packstorelo_ps(&f, _mm512_mask2int(0x1), x);
 309     return f;
 310 }
 311
 312 } // namespace gmx
 313
 314 #endif // GMX_SIMD_IMPL_X86_MIC_SIMD4_FLOAT_H