src/gromacs/simd/impl_x86_avx_256/impl_x86_avx_256_simd_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_256_SIMD_FLOAT_H
  38 #define GMX_SIMD_IMPL_X86_AVX_256_SIMD_FLOAT_H
  39
  40 #include "config.h"
  41
  42 #include <cassert>
  43 #include <cstddef>
  44 #include <cstdint>
  45
  46 #include <immintrin.h>
  47
  48 #include "gromacs/math/utilities.h"
  49
  50 namespace gmx
  51 {
  52
  53 class SimdFloat
  54 {
  55 public:
  56     MSVC_DIAGNOSTIC_IGNORE(26495) // simdInternal_ is not being initialized!
  57     SimdFloat() {}
  58     MSVC_DIAGNOSTIC_RESET
  59     SimdFloat(float f) : simdInternal_(_mm256_set1_ps(f)) {}
  60
  61     // Internal utility constructor to simplify return statements
  62     SimdFloat(__m256 simd) : simdInternal_(simd) {}
  63
  64     __m256 simdInternal_;
  65 };
  66
  67 class SimdFInt32
  68 {
  69 public:
  70     MSVC_DIAGNOSTIC_IGNORE(26495) // simdInternal_ is not being initialized!
  71     SimdFInt32() {}
  72     MSVC_DIAGNOSTIC_RESET
  73     SimdFInt32(std::int32_t i) : simdInternal_(_mm256_set1_epi32(i)) {}
  74
  75     // Internal utility constructor to simplify return statements
  76     SimdFInt32(__m256i simd) : simdInternal_(simd) {}
  77
  78     __m256i simdInternal_;
  79 };
  80
  81 class SimdFBool
  82 {
  83 public:
  84     MSVC_DIAGNOSTIC_IGNORE(26495) // simdInternal_ is not being initialized!
  85     SimdFBool() {}
  86     MSVC_DIAGNOSTIC_RESET
  87     SimdFBool(bool b) : simdInternal_(_mm256_castsi256_ps(_mm256_set1_epi32(b ? 0xFFFFFFFF : 0))) {}
  88
  89     // Internal utility constructor to simplify return statements
  90     SimdFBool(__m256 simd) : simdInternal_(simd) {}
  91
  92     __m256 simdInternal_;
  93 };
  94
  95 static inline SimdFloat gmx_simdcall simdLoad(const float* m, SimdFloatTag /*unused*/ = {})
  96 {
  97     assert(std::size_t(m) % 32 == 0);
  98     return { _mm256_load_ps(m) };
  99 }
 100
 101 static inline void gmx_simdcall store(float* m, SimdFloat a)
 102 {
 103     assert(std::size_t(m) % 32 == 0);
 104     _mm256_store_ps(m, a.simdInternal_);
 105 }
 106
 107 static inline SimdFloat gmx_simdcall simdLoadU(const float* m, SimdFloatTag /*unused*/ = {})
 108 {
 109     return { _mm256_loadu_ps(m) };
 110 }
 111
 112 static inline void gmx_simdcall storeU(float* m, SimdFloat a)
 113 {
 114     _mm256_storeu_ps(m, a.simdInternal_);
 115 }
 116
 117 static inline SimdFloat gmx_simdcall setZeroF()
 118 {
 119     return { _mm256_setzero_ps() };
 120 }
 121
 122 static inline SimdFInt32 gmx_simdcall simdLoad(const std::int32_t* m, SimdFInt32Tag /*unused*/)
 123 {
 124     assert(std::size_t(m) % 32 == 0);
 125     return { _mm256_load_si256(reinterpret_cast<const __m256i*>(m)) };
 126 }
 127
 128 static inline void gmx_simdcall store(std::int32_t* m, SimdFInt32 a)
 129 {
 130     assert(std::size_t(m) % 32 == 0);
 131     _mm256_store_si256(reinterpret_cast<__m256i*>(m), a.simdInternal_);
 132 }
 133
 134 static inline SimdFInt32 gmx_simdcall simdLoadU(const std::int32_t* m, SimdFInt32Tag /*unused*/)
 135 {
 136     return { _mm256_loadu_si256(reinterpret_cast<const __m256i*>(m)) };
 137 }
 138
 139 static inline void gmx_simdcall storeU(std::int32_t* m, SimdFInt32 a)
 140 {
 141     _mm256_storeu_si256(reinterpret_cast<__m256i*>(m), a.simdInternal_);
 142 }
 143
 144 static inline SimdFInt32 gmx_simdcall setZeroFI()
 145 {
 146     return { _mm256_setzero_si256() };
 147 }
 148
 149 template<int index>
 150 static inline std::int32_t gmx_simdcall extract(SimdFInt32 a)
 151 {
 152     return _mm_extract_epi32(_mm256_extractf128_si256(a.simdInternal_, index >> 2), index & 0x3);
 153 }
 154
 155 static inline SimdFloat gmx_simdcall operator&(SimdFloat a, SimdFloat b)
 156 {
 157     return { _mm256_and_ps(a.simdInternal_, b.simdInternal_) };
 158 }
 159
 160 static inline SimdFloat gmx_simdcall andNot(SimdFloat a, SimdFloat b)
 161 {
 162     return { _mm256_andnot_ps(a.simdInternal_, b.simdInternal_) };
 163 }
 164
 165 static inline SimdFloat gmx_simdcall operator|(SimdFloat a, SimdFloat b)
 166 {
 167     return { _mm256_or_ps(a.simdInternal_, b.simdInternal_) };
 168 }
 169
 170 static inline SimdFloat gmx_simdcall operator^(SimdFloat a, SimdFloat b)
 171 {
 172     return { _mm256_xor_ps(a.simdInternal_, b.simdInternal_) };
 173 }
 174
 175 static inline SimdFloat gmx_simdcall operator+(SimdFloat a, SimdFloat b)
 176 {
 177     return { _mm256_add_ps(a.simdInternal_, b.simdInternal_) };
 178 }
 179
 180 static inline SimdFloat gmx_simdcall operator-(SimdFloat a, SimdFloat b)
 181 {
 182     return { _mm256_sub_ps(a.simdInternal_, b.simdInternal_) };
 183 }
 184
 185 static inline SimdFloat gmx_simdcall operator-(SimdFloat x)
 186 {
 187     return { _mm256_xor_ps(x.simdInternal_, _mm256_set1_ps(GMX_FLOAT_NEGZERO)) };
 188 }
 189
 190 static inline SimdFloat gmx_simdcall operator*(SimdFloat a, SimdFloat b)
 191 {
 192     return { _mm256_mul_ps(a.simdInternal_, b.simdInternal_) };
 193 }
 194
 195 // Override for AVX2 and higher
 196 #if GMX_SIMD_X86_AVX_256
 197 static inline SimdFloat gmx_simdcall fma(SimdFloat a, SimdFloat b, SimdFloat c)
 198 {
 199     return { _mm256_add_ps(_mm256_mul_ps(a.simdInternal_, b.simdInternal_), c.simdInternal_) };
 200 }
 201
 202 static inline SimdFloat gmx_simdcall fms(SimdFloat a, SimdFloat b, SimdFloat c)
 203 {
 204     return { _mm256_sub_ps(_mm256_mul_ps(a.simdInternal_, b.simdInternal_), c.simdInternal_) };
 205 }
 206
 207 static inline SimdFloat gmx_simdcall fnma(SimdFloat a, SimdFloat b, SimdFloat c)
 208 {
 209     return { _mm256_sub_ps(c.simdInternal_, _mm256_mul_ps(a.simdInternal_, b.simdInternal_)) };
 210 }
 211
 212 static inline SimdFloat gmx_simdcall fnms(SimdFloat a, SimdFloat b, SimdFloat c)
 213 {
 214     return { _mm256_sub_ps(_mm256_setzero_ps(),
 215                            _mm256_add_ps(_mm256_mul_ps(a.simdInternal_, b.simdInternal_), c.simdInternal_)) };
 216 }
 217 #endif
 218
 219 static inline SimdFloat gmx_simdcall rsqrt(SimdFloat x)
 220 {
 221     return { _mm256_rsqrt_ps(x.simdInternal_) };
 222 }
 223
 224 static inline SimdFloat gmx_simdcall rcp(SimdFloat x)
 225 {
 226     return { _mm256_rcp_ps(x.simdInternal_) };
 227 }
 228
 229 static inline SimdFloat gmx_simdcall maskAdd(SimdFloat a, SimdFloat b, SimdFBool m)
 230 {
 231     return { _mm256_add_ps(a.simdInternal_, _mm256_and_ps(b.simdInternal_, m.simdInternal_)) };
 232 }
 233
 234 static inline SimdFloat gmx_simdcall maskzMul(SimdFloat a, SimdFloat b, SimdFBool m)
 235 {
 236     return { _mm256_and_ps(_mm256_mul_ps(a.simdInternal_, b.simdInternal_), m.simdInternal_) };
 237 }
 238
 239 static inline SimdFloat maskzFma(SimdFloat a, SimdFloat b, SimdFloat c, SimdFBool m)
 240 {
 241     return { _mm256_and_ps(_mm256_add_ps(_mm256_mul_ps(a.simdInternal_, b.simdInternal_), c.simdInternal_),
 242                            m.simdInternal_) };
 243 }
 244
 245 static inline SimdFloat maskzRsqrt(SimdFloat x, SimdFBool m)
 246 {
 247 #ifndef NDEBUG
 248     x.simdInternal_ = _mm256_blendv_ps(_mm256_set1_ps(1.0F), x.simdInternal_, m.simdInternal_);
 249 #endif
 250     return { _mm256_and_ps(_mm256_rsqrt_ps(x.simdInternal_), m.simdInternal_) };
 251 }
 252
 253 static inline SimdFloat maskzRcp(SimdFloat x, SimdFBool m)
 254 {
 255 #ifndef NDEBUG
 256     x.simdInternal_ = _mm256_blendv_ps(_mm256_set1_ps(1.0F), x.simdInternal_, m.simdInternal_);
 257 #endif
 258     return { _mm256_and_ps(_mm256_rcp_ps(x.simdInternal_), m.simdInternal_) };
 259 }
 260
 261 static inline SimdFloat gmx_simdcall abs(SimdFloat x)
 262 {
 263     return { _mm256_andnot_ps(_mm256_set1_ps(GMX_FLOAT_NEGZERO), x.simdInternal_) };
 264 }
 265
 266 static inline SimdFloat gmx_simdcall max(SimdFloat a, SimdFloat b)
 267 {
 268     return { _mm256_max_ps(a.simdInternal_, b.simdInternal_) };
 269 }
 270
 271 static inline SimdFloat gmx_simdcall min(SimdFloat a, SimdFloat b)
 272 {
 273     return { _mm256_min_ps(a.simdInternal_, b.simdInternal_) };
 274 }
 275
 276 static inline SimdFloat gmx_simdcall round(SimdFloat x)
 277 {
 278     return { _mm256_round_ps(x.simdInternal_, _MM_FROUND_NINT) };
 279 }
 280
 281 static inline SimdFloat gmx_simdcall trunc(SimdFloat x)
 282 {
 283     return { _mm256_round_ps(x.simdInternal_, _MM_FROUND_TRUNC) };
 284 }
 285
 286 // Override for AVX2 and higher
 287 #if GMX_SIMD_X86_AVX_256
 288 template<MathOptimization opt = MathOptimization::Safe>
 289 static inline SimdFloat gmx_simdcall frexp(SimdFloat value, SimdFInt32* exponent)
 290 {
 291     const __m256 exponentMask = _mm256_castsi256_ps(_mm256_set1_epi32(0x7F800000));
 292     const __m256 mantissaMask = _mm256_castsi256_ps(_mm256_set1_epi32(0x807FFFFF));
 293     const __m256 half         = _mm256_set1_ps(0.5);
 294     const __m128i exponentBias = _mm_set1_epi32(126); // add 1 to make our definition identical to frexp()
 295
 296     __m256i iExponent     = _mm256_castps_si256(_mm256_and_ps(value.simdInternal_, exponentMask));
 297     __m128i iExponentHigh = _mm256_extractf128_si256(iExponent, 0x1);
 298     __m128i iExponentLow  = _mm256_castsi256_si128(iExponent);
 299     iExponentLow          = _mm_srli_epi32(iExponentLow, 23);
 300     iExponentHigh         = _mm_srli_epi32(iExponentHigh, 23);
 301     iExponentLow          = _mm_sub_epi32(iExponentLow, exponentBias);
 302     iExponentHigh         = _mm_sub_epi32(iExponentHigh, exponentBias);
 303     iExponent             = _mm256_castsi128_si256(iExponentLow);
 304     iExponent             = _mm256_insertf128_si256(iExponent, iExponentHigh, 0x1);
 305
 306     __m256 result = _mm256_or_ps(_mm256_and_ps(value.simdInternal_, mantissaMask), half);
 307
 308     if (opt == MathOptimization::Safe)
 309     {
 310         __m256 valueIsZero = _mm256_cmp_ps(_mm256_setzero_ps(), value.simdInternal_, _CMP_EQ_OQ);
 311         // Set the upper/lower 64-bit-fields of "iExponent" to 0-bits if the corresponding input value was +-0.0
 312         // ... but we need to do the actual andnot operation as float, since AVX1 does not support integer ops.
 313         iExponent = _mm256_castps_si256(_mm256_andnot_ps(valueIsZero, _mm256_castsi256_ps(iExponent)));
 314
 315         // Set result to +-0 if the corresponding input value was +-0
 316         result = _mm256_blendv_ps(result, value.simdInternal_, valueIsZero);
 317     }
 318
 319     exponent->simdInternal_ = iExponent;
 320
 321     return { result };
 322 }
 323
 324 template<MathOptimization opt = MathOptimization::Safe>
 325 static inline SimdFloat gmx_simdcall ldexp(SimdFloat value, SimdFInt32 exponent)
 326 {
 327     const __m128i exponentBias = _mm_set1_epi32(127);
 328     __m256i       iExponent;
 329     __m128i       iExponentLow, iExponentHigh;
 330
 331     iExponentHigh = _mm256_extractf128_si256(exponent.simdInternal_, 0x1);
 332     iExponentLow  = _mm256_castsi256_si128(exponent.simdInternal_);
 333
 334     iExponentLow  = _mm_add_epi32(iExponentLow, exponentBias);
 335     iExponentHigh = _mm_add_epi32(iExponentHigh, exponentBias);
 336
 337     if (opt == MathOptimization::Safe)
 338     {
 339         // Make sure biased argument is not negative
 340         iExponentLow  = _mm_max_epi32(iExponentLow, _mm_setzero_si128());
 341         iExponentHigh = _mm_max_epi32(iExponentHigh, _mm_setzero_si128());
 342     }
 343
 344     iExponentLow  = _mm_slli_epi32(iExponentLow, 23);
 345     iExponentHigh = _mm_slli_epi32(iExponentHigh, 23);
 346     iExponent     = _mm256_castsi128_si256(iExponentLow);
 347     iExponent     = _mm256_insertf128_si256(iExponent, iExponentHigh, 0x1);
 348     return { _mm256_mul_ps(value.simdInternal_, _mm256_castsi256_ps(iExponent)) };
 349 }
 350 #endif
 351
 352 static inline float gmx_simdcall reduce(SimdFloat a)
 353 {
 354     __m128 t0;
 355     t0 = _mm_add_ps(_mm256_castps256_ps128(a.simdInternal_), _mm256_extractf128_ps(a.simdInternal_, 0x1));
 356     t0 = _mm_add_ps(t0, _mm_permute_ps(t0, _MM_SHUFFLE(1, 0, 3, 2)));
 357     t0 = _mm_add_ss(t0, _mm_permute_ps(t0, _MM_SHUFFLE(0, 3, 2, 1)));
 358     return *reinterpret_cast<float*>(&t0);
 359 }
 360
 361 static inline SimdFBool gmx_simdcall operator==(SimdFloat a, SimdFloat b)
 362 {
 363     return { _mm256_cmp_ps(a.simdInternal_, b.simdInternal_, _CMP_EQ_OQ) };
 364 }
 365
 366 static inline SimdFBool gmx_simdcall operator!=(SimdFloat a, SimdFloat b)
 367 {
 368     return { _mm256_cmp_ps(a.simdInternal_, b.simdInternal_, _CMP_NEQ_OQ) };
 369 }
 370
 371 static inline SimdFBool gmx_simdcall operator<(SimdFloat a, SimdFloat b)
 372 {
 373     return { _mm256_cmp_ps(a.simdInternal_, b.simdInternal_, _CMP_LT_OQ) };
 374 }
 375
 376 static inline SimdFBool gmx_simdcall operator<=(SimdFloat a, SimdFloat b)
 377 {
 378     return { _mm256_cmp_ps(a.simdInternal_, b.simdInternal_, _CMP_LE_OQ) };
 379 }
 380
 381 // Override for AVX2 and higher
 382 #if GMX_SIMD_X86_AVX_256
 383 static inline SimdFBool gmx_simdcall testBits(SimdFloat a)
 384 {
 385     __m256 tst = _mm256_cvtepi32_ps(_mm256_castps_si256(a.simdInternal_));
 386
 387     return { _mm256_cmp_ps(tst, _mm256_setzero_ps(), _CMP_NEQ_OQ) };
 388 }
 389 #endif
 390
 391 static inline SimdFBool gmx_simdcall operator&&(SimdFBool a, SimdFBool b)
 392 {
 393     return { _mm256_and_ps(a.simdInternal_, b.simdInternal_) };
 394 }
 395
 396 static inline SimdFBool gmx_simdcall operator||(SimdFBool a, SimdFBool b)
 397 {
 398     return { _mm256_or_ps(a.simdInternal_, b.simdInternal_) };
 399 }
 400
 401 static inline bool gmx_simdcall anyTrue(SimdFBool a)
 402 {
 403     return _mm256_movemask_ps(a.simdInternal_) != 0;
 404 }
 405
 406 static inline SimdFloat gmx_simdcall selectByMask(SimdFloat a, SimdFBool mask)
 407 {
 408     return { _mm256_and_ps(a.simdInternal_, mask.simdInternal_) };
 409 }
 410
 411 static inline SimdFloat gmx_simdcall selectByNotMask(SimdFloat a, SimdFBool mask)
 412 {
 413     return { _mm256_andnot_ps(mask.simdInternal_, a.simdInternal_) };
 414 }
 415
 416 static inline SimdFloat gmx_simdcall blend(SimdFloat a, SimdFloat b, SimdFBool sel)
 417 {
 418     return { _mm256_blendv_ps(a.simdInternal_, b.simdInternal_, sel.simdInternal_) };
 419 }
 420
 421 static inline SimdFInt32 gmx_simdcall cvtR2I(SimdFloat a)
 422 {
 423     return { _mm256_cvtps_epi32(a.simdInternal_) };
 424 }
 425
 426 static inline SimdFInt32 gmx_simdcall cvttR2I(SimdFloat a)
 427 {
 428     return { _mm256_cvttps_epi32(a.simdInternal_) };
 429 }
 430
 431 static inline SimdFloat gmx_simdcall cvtI2R(SimdFInt32 a)
 432 {
 433     return { _mm256_cvtepi32_ps(a.simdInternal_) };
 434 }
 435
 436 } // namespace gmx
 437
 438 #endif // GMX_SIMD_IMPL_X86_AVX_256_SIMD_FLOAT_H