src/gromacs/simd/tests/simd.cpp

   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 #include "gmxpre.h"
  37
  38 #include "simd.h"
  39
  40 #include "gromacs/simd/simd.h"
  41 #include "gromacs/utility/basedefinitions.h"
  42
  43 #if GMX_SIMD
  44
  45 namespace gmx
  46 {
  47 namespace test
  48 {
  49
  50 /*! \cond internal */
  51 /*! \addtogroup module_simd */
  52 /*! \{ */
  53
  54 /* Unfortunately we cannot keep static SIMD constants in the test fixture class.
  55  * The problem is that SIMD memory need to be aligned, and in particular
  56  * this applies to automatic storage of variables in classes. For SSE registers
  57  * this means 16-byte alignment (which seems to work), but AVX requires 32-bit
  58  * alignment. At least both gcc-4.7.3 and Apple clang-5.0 (OS X 10.9) fail to
  59  * align these variables when they are stored as data in a class.
  60  *
  61  * In theory we could set some of these on-the-fly e.g. with setSimdRealFrom3R()
  62  * instead (although that would mean repeating code between tests), but many of
  63  * the constants depend on the current precision not to mention they
  64  * occasionally have many digits that need to be exactly right, and keeping
  65  * them in a single place makes sure they are consistent.
  66  */
  67 #    if GMX_SIMD_HAVE_REAL
  68 const SimdReal rSimd_c0c1c2 = setSimdRealFrom3R(c0, c1, c2);
  69 const SimdReal rSimd_c3c4c5 = setSimdRealFrom3R(c3, c4, c5);
  70 const SimdReal rSimd_c6c7c8 = setSimdRealFrom3R(c6, c7, c8);
  71 const SimdReal rSimd_c3c0c4 = setSimdRealFrom3R(c3, c0, c4);
  72 const SimdReal rSimd_c4c6c8 = setSimdRealFrom3R(c4, c6, c8);
  73 const SimdReal rSimd_c7c2c3 = setSimdRealFrom3R(c7, c2, c3);
  74 const SimdReal rSimd_m0m1m2 = setSimdRealFrom3R(-c0, -c1, -c2);
  75 const SimdReal rSimd_m3m0m4 = setSimdRealFrom3R(-c3, -c0, -c4);
  76
  77 const SimdReal rSimd_2p25  = setSimdRealFrom1R(2.25);
  78 const SimdReal rSimd_3p25  = setSimdRealFrom1R(3.25);
  79 const SimdReal rSimd_3p75  = setSimdRealFrom1R(3.75);
  80 const SimdReal rSimd_m2p25 = setSimdRealFrom1R(-2.25);
  81 const SimdReal rSimd_m3p25 = setSimdRealFrom1R(-3.25);
  82 const SimdReal rSimd_m3p75 = setSimdRealFrom1R(-3.75);
  83 const SimdReal rSimd_Exp   = setSimdRealFrom3R(1.4055235171027452623914516e+18,
  84                                              5.3057102734253445623914516e-13,
  85                                              -2.1057102745623934534514516e+16);
  86
  87 #        if GMX_SIMD_HAVE_DOUBLE && GMX_DOUBLE
  88 // Make sure we also test exponents outside single precision when we use double
  89 const SimdReal rSimd_ExpDouble1 =
  90         setSimdRealFrom3R(0.0, 8.794495252903116023030553e-140, -3.637060701570496477655022e+202);
  91 const SimdReal rSimd_ExpDouble2 =
  92         setSimdRealFrom3R(6.287393598732017379054414e+176, 0.0, -3.637060701570496477655022e+202);
  93 #        endif // GMX_SIMD_HAVE_DOUBLE && GMX_DOUBLE
  94
  95 #        if GMX_SIMD_HAVE_LOGICAL
  96 // The numbers below all have exponent (2^0), which will not change with AND/OR operations.
  97 // We also leave the last part of the mantissa as zeros, to avoid rounding issues in the compiler
  98 #            if GMX_DOUBLE
  99 const SimdReal rSimd_logicalA =
 100         setSimdRealFrom1R(1.3333333332557231188); // mantissa 01010101010101010101010101010101
 101 const SimdReal rSimd_logicalB =
 102         setSimdRealFrom1R(1.7999999998137354851); // mantissa 11001100110011001100110011001100
 103 const SimdReal rSimd_logicalResultAnd =
 104         setSimdRealFrom1R(1.266666666604578495); // mantissa 01000100010001000100010001000100
 105 const SimdReal rSimd_logicalResultOr =
 106         setSimdRealFrom1R(1.8666666664648801088); // mantissa 11011101110111011101110111011101
 107 #            else                                 // GMX_DOUBLE
 108 const SimdReal rSimd_logicalA = setSimdRealFrom1R(1.3333282470703125); // mantissa 0101010101010101
 109 const SimdReal rSimd_logicalB = setSimdRealFrom1R(1.79998779296875);   // mantissa 1100110011001100
 110 const SimdReal rSimd_logicalResultAnd = setSimdRealFrom1R(1.26666259765625); // mantissa 0100010001000100
 111 const SimdReal rSimd_logicalResultOr = setSimdRealFrom1R(1.8666534423828125); // mantissa 1101110111011101
 112 #            endif                                // GMX_DOUBLE
 113 #        endif                                    // GMX_SIMD_HAVE_LOGICAL
 114
 115 #    endif // GMX_SIMD_HAVE_REAL
 116 #    if GMX_SIMD_HAVE_INT32_ARITHMETICS
 117 const SimdInt32 iSimd_1_2_3    = setSimdIntFrom3I(1, 2, 3);
 118 const SimdInt32 iSimd_4_5_6    = setSimdIntFrom3I(4, 5, 6);
 119 const SimdInt32 iSimd_7_8_9    = setSimdIntFrom3I(7, 8, 9);
 120 const SimdInt32 iSimd_5_7_9    = setSimdIntFrom3I(5, 7, 9);
 121 const SimdInt32 iSimd_1M_2M_3M = setSimdIntFrom3I(1000000, 2000000, 3000000);
 122 const SimdInt32 iSimd_4M_5M_6M = setSimdIntFrom3I(4000000, 5000000, 6000000);
 123 const SimdInt32 iSimd_5M_7M_9M = setSimdIntFrom3I(5000000, 7000000, 9000000);
 124 #    endif
 125 #    if GMX_SIMD_HAVE_INT32_LOGICAL
 126 const SimdInt32 iSimd_0xF0F0F0F0 = setSimdIntFrom1I(0xF0F0F0F0);
 127 const SimdInt32 iSimd_0xCCCCCCCC = setSimdIntFrom1I(0xCCCCCCCC);
 128 #    endif
 129
 130 #    if GMX_SIMD_HAVE_REAL
 131 TEST(SimdTest, GmxAligned)
 132 {
 133     // Test alignment with two variables that must be aligned, and one that
 134     // doesn't have to be. The order of variables is up to the compiler, but
 135     // if it ignores alignment it is highly unlikely that both r1/r3 still end
 136     // up being aligned by mistake.
 137     alignas(GMX_SIMD_ALIGNMENT) real r1;
 138     real                             r2;
 139     alignas(GMX_SIMD_ALIGNMENT) real r3;
 140
 141     std::uint64_t addr1 = reinterpret_cast<std::uint64_t>(&r1);
 142     std::uint64_t addr2 = reinterpret_cast<std::uint64_t>(&r2);
 143     std::uint64_t addr3 = reinterpret_cast<std::uint64_t>(&r3);
 144
 145     EXPECT_EQ(0, addr1 % GMX_SIMD_ALIGNMENT);
 146     EXPECT_NE(0, addr2); // Just so r2 is not optimized away
 147     EXPECT_EQ(0, addr3 % GMX_SIMD_ALIGNMENT);
 148
 149     alignas(GMX_SIMD_ALIGNMENT) std::int32_t i1;
 150     std::int32_t                             i2;
 151     alignas(GMX_SIMD_ALIGNMENT) std::int32_t i3;
 152
 153     addr1 = reinterpret_cast<std::uint64_t>(&i1);
 154     addr2 = reinterpret_cast<std::uint64_t>(&i2);
 155     addr3 = reinterpret_cast<std::uint64_t>(&i3);
 156
 157     EXPECT_EQ(0, addr1 % GMX_SIMD_ALIGNMENT);
 158     EXPECT_NE(0, addr2); // Just so i2 is not optimized away
 159     EXPECT_EQ(0, addr3 % GMX_SIMD_ALIGNMENT);
 160 }
 161
 162
 163 ::std::vector<real> simdReal2Vector(const SimdReal simd)
 164 {
 165     alignas(GMX_SIMD_ALIGNMENT) real mem[GMX_SIMD_REAL_WIDTH];
 166
 167     store(mem, simd);
 168     std::vector<real> v(mem, mem + GMX_SIMD_REAL_WIDTH);
 169
 170     return v;
 171 }
 172
 173 SimdReal vector2SimdReal(const std::vector<real>& v)
 174 {
 175     alignas(GMX_SIMD_ALIGNMENT) real mem[GMX_SIMD_REAL_WIDTH];
 176
 177     for (int i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
 178     {
 179         mem[i] = v[i % v.size()]; // repeat vector contents to fill simd width
 180     }
 181     return load<SimdReal>(mem);
 182 }
 183
 184 SimdReal setSimdRealFrom3R(real r0, real r1, real r2)
 185 {
 186     std::vector<real> v(3);
 187     v[0] = r0;
 188     v[1] = r1;
 189     v[2] = r2;
 190     return vector2SimdReal(v);
 191 }
 192
 193 SimdReal setSimdRealFrom1R(real value)
 194 {
 195     std::vector<real> v(GMX_SIMD_REAL_WIDTH);
 196     for (int i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
 197     {
 198         v[i] = value;
 199     }
 200     return vector2SimdReal(v);
 201 }
 202
 203 testing::AssertionResult SimdTest::compareSimdRealUlp(const char*    refExpr,
 204                                                       const char*    tstExpr,
 205                                                       const SimdReal ref,
 206                                                       const SimdReal tst)
 207 {
 208     return compareVectorRealUlp(refExpr, tstExpr, simdReal2Vector(ref), simdReal2Vector(tst));
 209 }
 210
 211 testing::AssertionResult SimdTest::compareSimdEq(const char*    refExpr,
 212                                                  const char*    tstExpr,
 213                                                  const SimdReal ref,
 214                                                  const SimdReal tst)
 215 {
 216     return compareVectorEq(refExpr, tstExpr, simdReal2Vector(ref), simdReal2Vector(tst));
 217 }
 218
 219 std::vector<std::int32_t> simdInt2Vector(const SimdInt32 simd)
 220 {
 221     alignas(GMX_SIMD_ALIGNMENT) std::int32_t mem[GMX_SIMD_REAL_WIDTH];
 222
 223     store(mem, simd);
 224     std::vector<std::int32_t> v(mem, mem + GMX_SIMD_REAL_WIDTH);
 225
 226     return v;
 227 }
 228
 229 SimdInt32 vector2SimdInt(const std::vector<std::int32_t>& v)
 230 {
 231     alignas(GMX_SIMD_ALIGNMENT) std::int32_t mem[GMX_SIMD_REAL_WIDTH];
 232
 233     for (int i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
 234     {
 235         mem[i] = v[i % v.size()]; // repeat vector contents to fill simd width
 236     }
 237     return load<SimdInt32>(mem);
 238 }
 239
 240 SimdInt32 setSimdIntFrom3I(int i0, int i1, int i2)
 241 {
 242     std::vector<int> v(3);
 243     v[0] = i0;
 244     v[1] = i1;
 245     v[2] = i2;
 246     return vector2SimdInt(v);
 247 }
 248
 249 SimdInt32 setSimdIntFrom1I(int value)
 250 {
 251     std::vector<int> v(GMX_SIMD_REAL_WIDTH);
 252     for (int i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
 253     {
 254         v[i] = value;
 255     }
 256     return vector2SimdInt(v);
 257 }
 258
 259 ::testing::AssertionResult SimdTest::compareSimdEq(const char*     refExpr,
 260                                                    const char*     tstExpr,
 261                                                    const SimdInt32 ref,
 262                                                    const SimdInt32 tst)
 263 {
 264     return compareVectorEq(refExpr, tstExpr, simdInt2Vector(ref), simdInt2Vector(tst));
 265 }
 266
 267 #    endif // GMX_SIMD_HAVE_REAL
 268
 269 /*! \} */
 270 /*! \endcond */
 271
 272 } // namespace test
 273 } // namespace gmx
 274
 275 #endif // GMX_SIMD