include/gmx_x86_simd_macros.h

   1 /*
   2  *                This source code is part of
   3  *
   4  *                 G   R   O   M   A   C   S
   5  *
   6  * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
   7  * Copyright (c) 2001-2012, The GROMACS Development Team
   8  *
   9  * Gromacs is a library for molecular simulation and trajectory analysis,
  10  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  11  * a full list of developers and information, check out http://www.gromacs.org
  12  *
  13  * This program is free software; you can redistribute it and/or modify it under
  14  * the terms of the GNU Lesser General Public License as published by the Free
  15  * Software Foundation; either version 2 of the License, or (at your option) any
  16  * later version.
  17  * As a special exception, you may use this file as part of a free software
  18  * library without restriction.  Specifically, if other files instantiate
  19  * templates or use macros or inline functions from this file, or you compile
  20  * this file and link it with other files to produce an executable, this
  21  * file does not by itself cause the resulting executable to be covered by
  22  * the GNU Lesser General Public License.
  23  *
  24  * In plain-speak: do not worry about classes/macros/templates either - only
  25  * changes to the library have to be LGPL, not an application linking with it.
  26  *
  27  * To help fund GROMACS development, we humbly ask that you cite
  28  * the papers people have written on it - you can find them on the website!
  29  */
  30
  31 /* Undefine all defines used below so we can include this file multiple times
  32  * with different settings from the same source file.
  33  */
  34
  35 /* NOTE: floor and blend are NOT available with SSE2 only acceleration */
  36
  37 #undef GMX_X86_SIMD_WIDTH_HERE
  38
  39 #undef gmx_epi32
  40
  41 #undef gmx_mm_pr
  42
  43 #undef gmx_load_pr
  44 #undef gmx_load1_pr
  45 #undef gmx_set1_pr
  46 #undef gmx_setzero_pr
  47 #undef gmx_store_pr
  48 #undef gmx_storeu_pr
  49
  50 #undef gmx_add_pr
  51 #undef gmx_sub_pr
  52 #undef gmx_mul_pr
  53 #undef gmx_max_pr
  54 #undef gmx_cmplt_pr
  55 #undef gmx_and_pr
  56 #undef gmx_or_pr
  57 #undef gmx_andnot_pr
  58
  59 #undef gmx_floor_pr
  60 #undef gmx_blendv_pr
  61
  62 #undef gmx_movemask_pr
  63
  64 #undef gmx_mm_castsi128_pr
  65
  66 #undef gmx_cvttpr_epi32
  67 #undef gmx_cvtepi32_pr
  68
  69 #undef gmx_invsqrt_pr
  70 #undef gmx_calc_rsq_pr
  71 #undef gmx_sum4_pr
  72
  73
  74 /* By defining GMX_MM128_HERE or GMX_MM256_HERE before including this file
  75  * the same intrinsics, with defines, can be compiled for either 128 or 256
  76  * bit wide SSE or AVX instructions.
  77  * The gmx_ prefix is replaced by _mm_ or _mm256_ (SSE or AVX).
  78  * The _pr suffix is replaced by _ps or _pd (single or double precision).
  79  * Note that compiler settings will decide if 128-bit intrinsics will
  80  * be translated into SSE or AVX instructions.
  81  */
  82
  83 #if !defined GMX_MM128_HERE && !defined GMX_MM256_HERE
  84 "You should define GMX_MM128_HERE or GMX_MM256_HERE"
  85 #endif
  86
  87 #if defined GMX_MM128_HERE && defined GMX_MM256_HERE
  88 "You should not define both GMX_MM128_HERE and GMX_MM256_HERE"
  89 #endif
  90
  91 #ifdef GMX_MM128_HERE
  92
  93 #define gmx_epi32  __m128i
  94
  95 #ifndef GMX_DOUBLE
  96
  97 #include "gmx_x86_simd_single.h"
  98
  99 #define GMX_X86_SIMD_WIDTH_HERE  4
 100
 101 #define gmx_mm_pr  __m128
 102
 103 #define gmx_load_pr       _mm_load_ps
 104 #define gmx_load1_pr      _mm_load1_ps
 105 #define gmx_set1_pr       _mm_set1_ps
 106 #define gmx_setzero_pr    _mm_setzero_ps
 107 #define gmx_store_pr      _mm_store_ps
 108 #define gmx_storeu_pr     _mm_storeu_ps
 109
 110 #define gmx_add_pr        _mm_add_ps
 111 #define gmx_sub_pr        _mm_sub_ps
 112 #define gmx_mul_pr        _mm_mul_ps
 113 #define gmx_max_pr        _mm_max_ps
 114 #define gmx_cmplt_pr      _mm_cmplt_ps
 115 #define gmx_and_pr        _mm_and_ps
 116 #define gmx_or_pr         _mm_or_ps
 117 #define gmx_andnot_pr     _mm_andnot_ps
 118
 119 #define gmx_floor_pr      _mm_floor_ps
 120 #define gmx_blendv_pr     _mm_blendv_ps
 121
 122 #define gmx_movemask_pr   _mm_movemask_ps
 123
 124 #define gmx_mm_castsi128_pr gmx_mm_castsi128_ps
 125
 126 #define gmx_cvttpr_epi32  _mm_cvttps_epi32
 127 #define gmx_cvtepi32_pr   _mm_cvtepi32_ps
 128
 129 #define gmx_invsqrt_pr    gmx_mm_invsqrt_ps
 130 #define gmx_calc_rsq_pr   gmx_mm_calc_rsq_ps
 131 #define gmx_sum4_pr       gmx_mm_sum4_ps
 132
 133 #else /* ifndef GMX_DOUBLE */
 134
 135 #include "gmx_x86_simd_double.h"
 136
 137 #define GMX_X86_SIMD_WIDTH_HERE  2
 138
 139 #define gmx_mm_pr  __m128d
 140
 141 #define gmx_load_pr       _mm_load_pd
 142 #define gmx_load1_pr      _mm_load1_pd
 143 #define gmx_set1_pr       _mm_set1_pd
 144 #define gmx_setzero_pr    _mm_setzero_pd
 145 #define gmx_store_pr      _mm_store_pd
 146 #define gmx_storeu_pr     _mm_storeu_pd
 147
 148 #define gmx_add_pr        _mm_add_pd
 149 #define gmx_sub_pr        _mm_sub_pd
 150 #define gmx_mul_pr        _mm_mul_pd
 151 #define gmx_max_pr        _mm_max_pd
 152 #define gmx_cmplt_pr      _mm_cmplt_pd
 153 #define gmx_and_pr        _mm_and_pd
 154 #define gmx_or_pr         _mm_or_pd
 155 #define gmx_andnot_pr     _mm_andnot_pd
 156
 157 #define gmx_floor_pr      _mm_floor_pd
 158 #define gmx_blendv_pr     _mm_blendv_pd
 159
 160 #define gmx_movemask_pr   _mm_movemask_pd
 161
 162 #define gmx_mm_castsi128_pr gmx_mm_castsi128_pd
 163
 164 #define gmx_cvttpr_epi32  _mm_cvttpd_epi32
 165 #define gmx_cvtepi32_pr   _mm_cvtepi32_pd
 166
 167 #define gmx_invsqrt_pr    gmx_mm_invsqrt_pd
 168 #define gmx_calc_rsq_pr   gmx_mm_calc_rsq_pd
 169 #define gmx_sum4_pr       gmx_mm_sum4_pd
 170
 171 #endif /* ifndef GMX_DOUBLE */
 172
 173 #endif /* GMX_MM128_HERE */
 174
 175 #ifdef GMX_MM256_HERE
 176
 177 #define gmx_epi32 __m256i
 178
 179 #ifndef GMX_DOUBLE
 180
 181 #include "gmx_x86_simd_single.h"
 182
 183 #define GMX_X86_SIMD_WIDTH_HERE  8
 184
 185 #define gmx_mm_pr  __m256
 186
 187 #define gmx_load_pr       _mm256_load_ps
 188 #define gmx_load1_pr(x)   _mm256_set1_ps((x)[0])
 189 #define gmx_set1_pr       _mm256_set1_ps
 190 #define gmx_setzero_pr    _mm256_setzero_ps
 191 #define gmx_store_pr      _mm256_store_ps
 192 #define gmx_storeu_pr     _mm256_storeu_ps
 193
 194 #define gmx_add_pr        _mm256_add_ps
 195 #define gmx_sub_pr        _mm256_sub_ps
 196 #define gmx_mul_pr        _mm256_mul_ps
 197 #define gmx_max_pr        _mm256_max_ps
 198 /* Not-equal (ordered, non-signaling)  */
 199 #define gmx_cmpneq_pr(x,y)  _mm256_cmp_ps(x,y,0x0c)
 200 /* Less-than (ordered, non-signaling)  */
 201 #define gmx_cmplt_pr(x,y) _mm256_cmp_ps(x,y,0x11)
 202 #define gmx_and_pr        _mm256_and_ps
 203 #define gmx_or_pr         _mm256_or_ps
 204 #define gmx_andnot_pr     _mm256_andnot_ps
 205
 206 #define gmx_floor_pr      _mm256_floor_ps
 207 #define gmx_blendv_pr     _mm256_blendv_ps
 208
 209 #define gmx_movemask_pr   _mm256_movemask_ps
 210
 211 #define gmx_mm_castsi256_pr _mm256_castsi256_ps
 212
 213 #define gmx_cvttpr_epi32  _mm256_cvttps_epi32
 214
 215 #define gmx_invsqrt_pr    gmx_mm256_invsqrt_ps
 216 #define gmx_calc_rsq_pr   gmx_mm256_calc_rsq_ps
 217 #define gmx_sum4_pr       gmx_mm256_sum4_ps
 218
 219 #else
 220
 221 #include "gmx_x86_simd_double.h"
 222
 223 #define GMX_X86_SIMD_WIDTH_HERE  4
 224
 225 #define gmx_mm_pr  __m256d
 226
 227 #define gmx_load_pr       _mm256_load_pd
 228 #define gmx_load1_pr(x)   _mm256_set1_pd((x)[0])
 229 #define gmx_set1_pr       _mm256_set1_pd
 230 #define gmx_setzero_pr    _mm256_setzero_pd
 231 #define gmx_store_pr      _mm256_store_pd
 232 #define gmx_storeu_pr     _mm256_storeu_pd
 233
 234 #define gmx_add_pr        _mm256_add_pd
 235 #define gmx_sub_pr        _mm256_sub_pd
 236 #define gmx_mul_pr        _mm256_mul_pd
 237 #define gmx_max_pr        _mm256_max_pd
 238 /* Not-equal (ordered, non-signaling)  */
 239 #define gmx_cmpneq_pr(x,y)  _mm256_cmp_pd(x,y,0x0c)
 240 /* Less-than (ordered, non-signaling)  */
 241 #define gmx_cmplt_pr(x,y) _mm256_cmp_pd(x,y,0x11)
 242 #define gmx_and_pr        _mm256_and_pd
 243 #define gmx_or_pr         _mm256_or_pd
 244 #define gmx_andnot_pr     _mm256_andnot_pd
 245
 246 #define gmx_floor_pr      _mm256_floor_pd
 247 #define gmx_blendv_pr     _mm256_blendv_pd
 248
 249 #define gmx_movemask_pr   _mm256_movemask_pd
 250
 251 #define gmx_mm_castsi256_pr _mm256_castsi256_pd
 252
 253 #define gmx_cvttpr_epi32  _mm256_cvttpd_epi32
 254
 255 #define gmx_invsqrt_pr    gmx_mm256_invsqrt_pd
 256 #define gmx_calc_rsq_pr   gmx_mm256_calc_rsq_pd
 257 #define gmx_sum4_pr       gmx_mm256_sum4_pd
 258
 259 #endif
 260
 261 #endif /* GMX_MM256_HERE */