src/gromacs/simd/tests/simd4_floatingpoint.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2014, 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 #ifdef HAVE_CONFIG_H
  36 #include <config.h>
  37 #endif
  38
  39 #include <math.h>
  40 #include "gromacs/math/utilities.h"
  41
  42 #include "simd4.h"
  43
  44 namespace gmx
  45 {
  46 namespace test
  47 {
  48 namespace
  49 {
  50
  51 /*! \cond internal */
  52 /*! \addtogroup module_simd */
  53 /*! \{ */
  54
  55 #ifdef GMX_SIMD4_HAVE_REAL
  56
  57 /*! \brief Test fixture for SIMD4 floating-point operations (identical to the SIMD4 \ref Simd4Test) */
  58 typedef Simd4Test Simd4FloatingpointTest;
  59
  60 TEST_F(Simd4FloatingpointTest, gmxSimd4SetZeroR)
  61 {
  62     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(0.0), gmx_simd4_setzero_r());
  63 }
  64
  65 TEST_F(Simd4FloatingpointTest, gmxSimd4Set1R)
  66 {
  67     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(1.0), gmx_simd4_set1_r(1.0));
  68 }
  69
  70 TEST_F(Simd4FloatingpointTest, gmxSimd4Load1R)
  71 {
  72     real r = 2.0;
  73     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(r), gmx_simd4_load1_r(&r));
  74 }
  75
  76 TEST_F(Simd4FloatingpointTest, gmxSimd4AddR)
  77 {
  78     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_5_7_9, gmx_simd4_add_r(rSimd4_1_2_3, rSimd4_4_5_6)); // 1+4=5, 2+5=7, 3+6=9
  79 }
  80
  81 TEST_F(Simd4FloatingpointTest, gmxSimd4SubR)
  82 {
  83     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_4_5_6, gmx_simd4_sub_r(rSimd4_5_7_9, rSimd4_1_2_3)); // 5-1=4, 7-2=5, 9-3=6
  84 }
  85
  86 TEST_F(Simd4FloatingpointTest, gmxSimd4MulR)
  87 {
  88     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(4, 10, 18), gmx_simd4_mul_r(rSimd4_1_2_3, rSimd4_4_5_6));
  89 }
  90
  91 TEST_F(Simd4FloatingpointTest, gmxSimd4FmaddR)
  92 {
  93     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(11, 18, 27), gmx_simd4_fmadd_r(rSimd4_1_2_3, rSimd4_4_5_6, rSimd4_7_8_9)); // 1*4+7, etc.
  94 }
  95
  96 TEST_F(Simd4FloatingpointTest, gmxSimd4FmsubR)
  97 {
  98     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-3, 2, 9), gmx_simd4_fmsub_r(rSimd4_1_2_3, rSimd4_4_5_6, rSimd4_7_8_9)); // 1*4-7, etc.
  99 }
 100
 101 TEST_F(Simd4FloatingpointTest, gmxSimd4FnmaddR)
 102 {
 103     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(3, -2, -9), gmx_simd4_fnmadd_r(rSimd4_1_2_3, rSimd4_4_5_6, rSimd4_7_8_9)); // -1*4+7, etc.
 104 }
 105
 106 TEST_F(Simd4FloatingpointTest, gmxSimd4FnmsubR)
 107 {
 108     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-11, -18, -27), gmx_simd4_fnmsub_r(rSimd4_1_2_3, rSimd4_4_5_6, rSimd4_7_8_9)); // -1*4-7, etc.
 109 }
 110
 111 TEST_F(Simd4FloatingpointTest, gmxSimd4FabsR)
 112 {
 113     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_1_2_3, gmx_simd4_fabs_r(rSimd4_1_2_3));    // fabs(x)=x
 114     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_1_2_3, gmx_simd4_fabs_r(rSimd4_m1_m2_m3)); // fabs(-x)=x
 115 }
 116
 117 TEST_F(Simd4FloatingpointTest, gmxSimd4FnegR)
 118 {
 119     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_m1_m2_m3, gmx_simd4_fneg_r(rSimd4_1_2_3));   // fneg(x)=-x
 120     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_1_2_3,   gmx_simd4_fneg_r(rSimd4_m1_m2_m3)); // fneg(-x)=x
 121 }
 122
 123 #ifdef GMX_SIMD4_HAVE_LOGICAL
 124 TEST_F(Simd4FloatingpointTest, gmxSimd4AndR)
 125 {
 126     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_Bits3, gmx_simd4_and_r(rSimd4_Bits1, rSimd4_Bits2)); // Bits1 & Bits2 = Bits3
 127 }
 128
 129 TEST_F(Simd4FloatingpointTest, gmxSimd4AndnotR)
 130 {
 131     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_Bits4, gmx_simd4_andnot_r(rSimd4_Bits1, rSimd4_Bits2)); // (~Bits1) & Bits2 = Bits3
 132 }
 133
 134 TEST_F(Simd4FloatingpointTest, gmxSimd4OrR)
 135 {
 136     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_Bits5, gmx_simd4_or_r(rSimd4_Bits1, rSimd4_Bits2)); // Bits1 | Bits2 = Bits3
 137 }
 138
 139 TEST_F(Simd4FloatingpointTest, gmxSimd4XorR)
 140 {
 141     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_Bits6, gmx_simd4_xor_r(rSimd4_Bits1, rSimd4_Bits2)); // Bits1 ^ Bits2 = Bits3
 142 }
 143 #endif
 144
 145 TEST_F(Simd4FloatingpointTest, gmxSimd4MaxR)
 146 {
 147     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(3, 2, 4), gmx_simd4_max_r(rSimd4_1_2_3, rSimd4_3_1_4));
 148     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(3, 2, 4), gmx_simd4_max_r(rSimd4_3_1_4, rSimd4_1_2_3));
 149     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-1, -1, -3), gmx_simd4_max_r(rSimd4_m1_m2_m3, rSimd4_m3_m1_m4));
 150     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-1, -1, -3), gmx_simd4_max_r(rSimd4_m3_m1_m4, rSimd4_m1_m2_m3));
 151 }
 152
 153 TEST_F(Simd4FloatingpointTest, gmxSimd4MinR)
 154 {
 155     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(1, 1, 3), gmx_simd4_min_r(rSimd4_1_2_3, rSimd4_3_1_4));
 156     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(1, 1, 3), gmx_simd4_min_r(rSimd4_3_1_4, rSimd4_1_2_3));
 157     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-3, -2, -4), gmx_simd4_min_r(rSimd4_m1_m2_m3, rSimd4_m3_m1_m4));
 158     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-3, -2, -4), gmx_simd4_min_r(rSimd4_m3_m1_m4, rSimd4_m1_m2_m3));
 159 }
 160
 161 TEST_F(Simd4FloatingpointTest, gmxSimd4RoundR)
 162 {
 163     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(2), gmx_simd4_round_r(gmx_simd4_set1_r(2.25)));
 164     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(4), gmx_simd4_round_r(gmx_simd4_set1_r(3.75)));
 165     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-2), gmx_simd4_round_r(gmx_simd4_set1_r(-2.25)));
 166     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-4), gmx_simd4_round_r(gmx_simd4_set1_r(-3.75)));
 167 }
 168
 169 TEST_F(Simd4FloatingpointTest, gmxSimd4TruncR)
 170 {
 171     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(2), gmx_simd4_trunc_r(rSimd4_2p25));
 172     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(3), gmx_simd4_trunc_r(rSimd4_3p75));
 173     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-2), gmx_simd4_trunc_r(rSimd4_m2p25));
 174     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-3), gmx_simd4_trunc_r(rSimd4_m3p75));
 175 }
 176
 177 /* We do extensive 1/sqrt(x) and 1/x accuracy testing in the tests for
 178  * the SIMD math functions, so we just make sure the lookup instructions
 179  * appear to work for a few values here.
 180  */
 181 TEST_F(Simd4FloatingpointTest, gmxSimd4RsqrtR)
 182 {
 183     gmx_simd4_real_t x      = setSimd4RealFrom3R(4.0, M_PI, 1234567890.0);
 184     gmx_simd4_real_t ref    = setSimd4RealFrom3R(0.5, 1.0/sqrt(M_PI), 1.0/sqrt(1234567890.0));
 185
 186     // The allowed Ulp deviation is 2 to the power of the number of mantissa
 187     // digits, minus the number of bits provided by the table lookup
 188     setUlpTol(1LL << (std::numeric_limits<real>::digits-GMX_SIMD_RSQRT_BITS));
 189     GMX_EXPECT_SIMD4_REAL_NEAR(ref, gmx_simd4_rsqrt_r(x));
 190 }
 191
 192 TEST_F(Simd4FloatingpointTest, gmxSimd4BoolCmpEqAndBlendZeroR)
 193 {
 194     gmx_simd4_bool_t eq   = gmx_simd4_cmpeq_r(rSimd4_5_7_9, rSimd4_7_8_9);
 195     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(0, 0, 3), gmx_simd4_blendzero_r(rSimd4_1_2_3, eq));
 196 }
 197
 198 TEST_F(Simd4FloatingpointTest, gmxSimd4BlendNotZeroR)
 199 {
 200     gmx_simd4_bool_t eq   = gmx_simd4_cmpeq_r(rSimd4_5_7_9, rSimd4_7_8_9);
 201     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(1, 2, 0), gmx_simd4_blendnotzero_r(rSimd4_1_2_3, eq));
 202 }
 203
 204 TEST_F(Simd4FloatingpointTest, gmxSimd4BoolCmpLER)
 205 {
 206     gmx_simd4_bool_t le   = gmx_simd4_cmple_r(rSimd4_5_7_9, rSimd4_7_8_9);
 207     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_1_2_3, gmx_simd4_blendzero_r(rSimd4_1_2_3, le));
 208 }
 209
 210 TEST_F(Simd4FloatingpointTest, gmxSimd4BoolCmpLTR)
 211 {
 212     gmx_simd4_bool_t lt   = gmx_simd4_cmplt_r(rSimd4_5_7_9, rSimd4_7_8_9);
 213     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(1, 2, 0), gmx_simd4_blendzero_r(rSimd4_1_2_3, lt));
 214 }
 215
 216 TEST_F(Simd4FloatingpointTest, gmxSimd4BoolAndB)
 217 {
 218     gmx_simd4_bool_t eq   = gmx_simd4_cmpeq_r(rSimd4_5_7_9, rSimd4_7_8_9);
 219     gmx_simd4_bool_t le   = gmx_simd4_cmple_r(rSimd4_5_7_9, rSimd4_7_8_9);
 220     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(0, 0, 3), gmx_simd4_blendzero_r(rSimd4_1_2_3, gmx_simd4_and_b(eq, le)));
 221 }
 222
 223 TEST_F(Simd4FloatingpointTest, gmxSimd4BoolOrB)
 224 {
 225     gmx_simd4_bool_t eq   = gmx_simd4_cmpeq_r(rSimd4_5_7_9, rSimd4_7_8_9);
 226     gmx_simd4_bool_t lt   = gmx_simd4_cmplt_r(rSimd4_5_7_9, rSimd4_7_8_9);
 227     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(1, 2, 3), gmx_simd4_blendzero_r(rSimd4_1_2_3, gmx_simd4_or_b(eq, lt)));
 228 }
 229
 230 TEST_F(Simd4FloatingpointTest, gmxSimd4AnytrueB)
 231 {
 232     gmx_simd4_bool_t eq;
 233
 234     /* this test is a bit tricky since we don't know the simd width.
 235      * We cannot check for truth values for "any" element beyond the first,
 236      * since that part of the data will not be used if simd width is 1.
 237      */
 238     eq = gmx_simd4_cmpeq_r(rSimd4_5_7_9, setSimd4RealFrom3R(5, 0, 0));
 239     EXPECT_NE(0, gmx_simd4_anytrue_b(eq));
 240
 241     eq = gmx_simd4_cmpeq_r(rSimd4_1_2_3, rSimd4_4_5_6);
 242     EXPECT_EQ(0, gmx_simd4_anytrue_b(eq));
 243 }
 244
 245 TEST_F(Simd4FloatingpointTest, gmxSimd4BlendvR)
 246 {
 247     gmx_simd4_bool_t lt   = gmx_simd4_cmplt_r(rSimd4_5_7_9, rSimd4_7_8_9);
 248     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(4, 5, 3), gmx_simd4_blendv_r(rSimd4_1_2_3, rSimd4_4_5_6, lt));
 249 }
 250
 251 TEST_F(Simd4FloatingpointTest, gmxSimd4ReduceR)
 252 {
 253     // The horizontal sum of the SIMD variable depends on the width, so
 254     // simply store it an extra time and calculate what the sum should be
 255     std::vector<real> v   = simd4Real2Vector(rSimd4_1_2_3);
 256     real              sum = 0.0;
 257
 258     for (int i = 0; i < GMX_SIMD4_WIDTH; i++)
 259     {
 260         sum += v[i];
 261     }
 262
 263     EXPECT_EQ(sum, gmx_simd4_reduce_r(rSimd4_1_2_3));
 264 }
 265
 266
 267 TEST_F(Simd4FloatingpointTest, gmxSimd4Dotproduct3R)
 268 {
 269     gmx_simd4_real_t v1 = setSimd4RealFrom3R(1, 4, 5);
 270     gmx_simd4_real_t v2 = setSimd4RealFrom3R(3, 8, 2);
 271 #    ifdef GMX_DOUBLE
 272     EXPECT_DOUBLE_EQ(45.0, gmx_simd4_dotproduct3_r(v1, v2));
 273 #    else
 274     EXPECT_FLOAT_EQ(45.0, gmx_simd4_dotproduct3_r(v1, v2));
 275 #    endif
 276 }
 277
 278 #endif      // GMX_SIMD4_HAVE_REAL
 279
 280 /*! \} */
 281 /*! \endcond */
 282
 283 }      // namespace
 284 }      // namespace
 285 }      // namespace