2 * This file is part of the GROMACS molecular simulation package.
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.
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.
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.
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.
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.
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.
38 #include "gromacs/math/utilities.h"
50 /*! \addtogroup module_simd */
53 #ifdef GMX_SIMD_HAVE_REAL
55 /*! \brief Test fixture for floating-point tests (identical to the generic \ref SimdTest) */
56 typedef SimdTest SimdFloatingpointTest;
58 TEST_F(SimdFloatingpointTest, gmxSimdSetZeroR)
60 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(0.0), gmx_simd_setzero_r());
63 TEST_F(SimdFloatingpointTest, gmxSimdSet1R)
65 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(1.0), gmx_simd_set1_r(1.0));
68 TEST_F(SimdFloatingpointTest, gmxSimdLoad1R)
71 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(r), gmx_simd_load1_r(&r));
74 TEST_F(SimdFloatingpointTest, gmxSimdAddR)
76 GMX_EXPECT_SIMD_REAL_EQ(rSimd_5_7_9,
77 gmx_simd_add_r(rSimd_1_2_3, rSimd_4_5_6)); // 1+4=5, 2+5=7, 3+6=9
80 TEST_F(SimdFloatingpointTest, gmxSimdSubR)
82 GMX_EXPECT_SIMD_REAL_EQ(rSimd_4_5_6,
83 gmx_simd_sub_r(rSimd_5_7_9, rSimd_1_2_3)); // 5-1=4, 7-2=5, 9-3=6
86 TEST_F(SimdFloatingpointTest, gmxSimdMulR)
88 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(4, 10, 18),
89 gmx_simd_mul_r(rSimd_1_2_3, rSimd_4_5_6));
92 TEST_F(SimdFloatingpointTest, gmxSimdFmaddR)
94 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(11, 18, 27),
95 gmx_simd_fmadd_r(rSimd_1_2_3, rSimd_4_5_6, rSimd_7_8_9)); // 1*4+7, etc.
98 TEST_F(SimdFloatingpointTest, gmxSimdFmsubR)
100 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-3, 2, 9),
101 gmx_simd_fmsub_r(rSimd_1_2_3, rSimd_4_5_6, rSimd_7_8_9)); // 1*4-7, etc.
104 TEST_F(SimdFloatingpointTest, gmxSimdFnmaddR)
106 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(3, -2, -9),
107 gmx_simd_fnmadd_r(rSimd_1_2_3, rSimd_4_5_6, rSimd_7_8_9)); // -1*4+7, etc.
110 TEST_F(SimdFloatingpointTest, gmxSimdFnmsubR)
112 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-11, -18, -27),
113 gmx_simd_fnmsub_r(rSimd_1_2_3, rSimd_4_5_6, rSimd_7_8_9)); // -1*4-7, etc.
116 TEST_F(SimdFloatingpointTest, gmxSimdFabsR)
118 GMX_EXPECT_SIMD_REAL_EQ(rSimd_1_2_3, gmx_simd_fabs_r(rSimd_1_2_3)); // fabs(x)=x
119 GMX_EXPECT_SIMD_REAL_EQ(rSimd_1_2_3, gmx_simd_fabs_r(rSimd_m1_m2_m3)); // fabs(-x)=x
122 TEST_F(SimdFloatingpointTest, gmxSimdFnegR)
124 GMX_EXPECT_SIMD_REAL_EQ(rSimd_m1_m2_m3, gmx_simd_fneg_r(rSimd_1_2_3)); // fneg(x)=-x
125 GMX_EXPECT_SIMD_REAL_EQ(rSimd_1_2_3, gmx_simd_fneg_r(rSimd_m1_m2_m3)); // fneg(-x)=x
128 #ifdef GMX_SIMD_HAVE_LOGICAL
129 /* 1.3333282470703125 has mantissa 0101010101010101 (followed by zeros)
130 * 1.79998779296875 has mantissa 1100110011001100 (followed by zeros)
131 * 1.26666259765625 has mantissa 0100010001000100 (followed by zeros)
132 * 1.8666534423828125 has mantissa 1101110111011101 (followed by zeros)
134 * Since all of them have the same exponent (2^0), the exponent will
135 * not change with AND or OR operations.
137 TEST_F(SimdFloatingpointTest, gmxSimdAndR)
139 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(1.26666259765625),
140 gmx_simd_and_r(gmx_simd_set1_r(1.3333282470703125),
141 gmx_simd_set1_r(1.79998779296875)));
144 TEST_F(SimdFloatingpointTest, gmxSimdOrR)
146 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(1.8666534423828125),
147 gmx_simd_or_r(gmx_simd_set1_r(1.3333282470703125),
148 gmx_simd_set1_r(1.79998779296875)));
151 TEST_F(SimdFloatingpointTest, gmxSimdXorR)
153 /* Test xor by taking xor with a number and its negative. This should result
154 * in only the sign bit being set. We then use this bit change the sign of
157 gmx_simd_real_t signbit = gmx_simd_xor_r(gmx_simd_set1_r(1.5), gmx_simd_set1_r(-1.5));
158 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-1, 2, -3), gmx_simd_xor_r(signbit, setSimdRealFrom3R(1, -2, 3)));
161 TEST_F(SimdFloatingpointTest, gmxSimdAndnotR)
163 /* Use xor (which we already tested, so fix that first if both tests fail)
164 * to extract the sign bit, and then use andnot to take absolute values.
166 gmx_simd_real_t signbit = gmx_simd_xor_r(gmx_simd_set1_r(1.5), gmx_simd_set1_r(-1.5));
167 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(1, 2, 3), gmx_simd_andnot_r(signbit, setSimdRealFrom3R(-1, 2, -3)));
172 TEST_F(SimdFloatingpointTest, gmxSimdMaxR)
174 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(3, 2, 4), gmx_simd_max_r(rSimd_1_2_3, rSimd_3_1_4));
175 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(3, 2, 4), gmx_simd_max_r(rSimd_3_1_4, rSimd_1_2_3));
176 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-1, -1, -3), gmx_simd_max_r(rSimd_m1_m2_m3, rSimd_m3_m1_m4));
177 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-1, -1, -3), gmx_simd_max_r(rSimd_m3_m1_m4, rSimd_m1_m2_m3));
180 TEST_F(SimdFloatingpointTest, gmxSimdMinR)
182 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(1, 1, 3), gmx_simd_min_r(rSimd_1_2_3, rSimd_3_1_4));
183 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(1, 1, 3), gmx_simd_min_r(rSimd_3_1_4, rSimd_1_2_3));
184 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-3, -2, -4), gmx_simd_min_r(rSimd_m1_m2_m3, rSimd_m3_m1_m4));
185 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-3, -2, -4), gmx_simd_min_r(rSimd_m3_m1_m4, rSimd_m1_m2_m3));
188 TEST_F(SimdFloatingpointTest, gmxSimdRoundR)
190 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(2), gmx_simd_round_r(gmx_simd_set1_r(2.25)));
191 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(4), gmx_simd_round_r(gmx_simd_set1_r(3.75)));
192 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(-2), gmx_simd_round_r(gmx_simd_set1_r(-2.25)));
193 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(-4), gmx_simd_round_r(gmx_simd_set1_r(-3.75)));
196 TEST_F(SimdFloatingpointTest, gmxSimdTruncR)
198 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(2), gmx_simd_trunc_r(rSimd_2p25));
199 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(3), gmx_simd_trunc_r(rSimd_3p75));
200 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(-2), gmx_simd_trunc_r(rSimd_m2p25));
201 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(-3), gmx_simd_trunc_r(rSimd_m3p75));
204 TEST_F(SimdFloatingpointTest, gmxSimdFractionR)
206 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(0.25), gmx_simd_fraction_r(rSimd_2p25)); // fract(2.25)=0.25
207 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(0.75), gmx_simd_fraction_r(rSimd_3p75)); // fract(3.75)=0.75
208 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(-0.25), gmx_simd_fraction_r(rSimd_m2p25)); // fract(-2.25)=-0.25
209 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(-0.75), gmx_simd_fraction_r(rSimd_m3p75)); // fract(-3.75)=-0.75
212 // We explicitly test the exponent/mantissa routines with double precision data,
213 // since these usually rely on direct manipulation and shift of the SIMD registers,
214 // where it is easy to make mistakes with single vs double precision.
216 TEST_F(SimdFloatingpointTest, gmxSimdGetExponentR)
218 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(60.0, -41.0, 54.0), gmx_simd_get_exponent_r(rSimd_Exp));
219 #if (defined GMX_SIMD_HAVE_DOUBLE) && (defined GMX_DOUBLE)
220 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(587.0, -462.0, 672.0), gmx_simd_get_exponent_r(rSimd_ExpDouble));
224 TEST_F(SimdFloatingpointTest, gmxSimdGetMantissaR)
226 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(1.219097320577810839026256,
227 1.166738027848349235071623,
228 1.168904015004464724825084), gmx_simd_get_mantissa_r(rSimd_Exp));
229 #if (defined GMX_SIMD_HAVE_DOUBLE) && (defined GMX_DOUBLE)
230 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(1.241261238952345623563251,
231 1.047294723759123852359232,
232 1.856066204750275957395734), gmx_simd_get_mantissa_r(rSimd_ExpDouble));
236 TEST_F(SimdFloatingpointTest, gmxSimdSetExponentR)
238 gmx_simd_real_t x0 = setSimdRealFrom3R(0.5, 11.5, 99.5);
239 gmx_simd_real_t x1 = setSimdRealFrom3R(-0.5, -11.5, -99.5);
241 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(pow(2.0, 60.0), pow(2.0, -41.0), pow(2.0, 54.0)),
242 gmx_simd_set_exponent_r(setSimdRealFrom3R(60.0, -41.0, 54.0)));
243 #if (defined GMX_SIMD_HAVE_DOUBLE) && (defined GMX_DOUBLE)
244 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(pow(2.0, 587.0), pow(2.0, -462.0), pow(2.0, 672.0)),
245 gmx_simd_set_exponent_r(setSimdRealFrom3R(587.0, -462.0, 672.0)));
247 /* Rounding mode in gmx_simd_set_exponent_r() must be consistent with gmx_simd_round_r() */
248 GMX_EXPECT_SIMD_REAL_EQ(gmx_simd_set_exponent_r(gmx_simd_round_r(x0)), gmx_simd_set_exponent_r(x0));
249 GMX_EXPECT_SIMD_REAL_EQ(gmx_simd_set_exponent_r(gmx_simd_round_r(x1)), gmx_simd_set_exponent_r(x1));
253 * We do extensive 1/sqrt(x) and 1/x accuracy testing in the math module, so
254 * we just make sure the lookup instructions appear to work here
257 TEST_F(SimdFloatingpointTest, gmxSimdRsqrtR)
259 gmx_simd_real_t x = setSimdRealFrom3R(4.0, M_PI, 1234567890.0);
260 gmx_simd_real_t ref = setSimdRealFrom3R(0.5, 1.0/sqrt(M_PI), 1.0/sqrt(1234567890.0));
262 /* Set the allowed ulp error as 2 to the power of the number of bits in
263 * the mantissa that do not have to be correct after the table lookup.
265 setUlpTol(1LL << (std::numeric_limits<real>::digits-GMX_SIMD_RSQRT_BITS));
267 GMX_EXPECT_SIMD_REAL_NEAR(ref, gmx_simd_rsqrt_r(x));
270 TEST_F(SimdFloatingpointTest, gmxSimdRcpR)
272 gmx_simd_real_t x = setSimdRealFrom3R(4.0, M_PI, 1234567890.0);
273 gmx_simd_real_t ref = setSimdRealFrom3R(0.25, 1.0/M_PI, 1.0/1234567890.0);
275 /* Set the allowed ulp error as 2 to the power of the number of bits in
276 * the mantissa that do not have to be correct after the table lookup.
278 setUlpTol(1LL << (std::numeric_limits<real>::digits-GMX_SIMD_RCP_BITS));
280 GMX_EXPECT_SIMD_REAL_NEAR(ref, gmx_simd_rcp_r(x));
283 TEST_F(SimdFloatingpointTest, gmxSimdBoolCmpEqAndBlendZeroR)
285 gmx_simd_bool_t eq = gmx_simd_cmpeq_r(rSimd_5_7_9, rSimd_7_8_9);
286 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(0, 0, 3), gmx_simd_blendzero_r(rSimd_1_2_3, eq));
289 TEST_F(SimdFloatingpointTest, gmxSimdBlendNotZeroR)
291 gmx_simd_bool_t eq = gmx_simd_cmpeq_r(rSimd_5_7_9, rSimd_7_8_9);
292 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(1, 2, 0), gmx_simd_blendnotzero_r(rSimd_1_2_3, eq));
295 TEST_F(SimdFloatingpointTest, gmxSimdBoolCmpLER)
297 gmx_simd_bool_t le = gmx_simd_cmple_r(rSimd_5_7_9, rSimd_7_8_9);
298 GMX_EXPECT_SIMD_REAL_EQ(rSimd_1_2_3, gmx_simd_blendzero_r(rSimd_1_2_3, le));
301 TEST_F(SimdFloatingpointTest, gmxSimdBoolCmpLTR)
303 gmx_simd_bool_t lt = gmx_simd_cmplt_r(rSimd_5_7_9, rSimd_7_8_9);
304 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(1, 2, 0), gmx_simd_blendzero_r(rSimd_1_2_3, lt));
307 TEST_F(SimdFloatingpointTest, gmxSimdBoolAndB)
309 gmx_simd_bool_t eq = gmx_simd_cmpeq_r(rSimd_5_7_9, rSimd_7_8_9);
310 gmx_simd_bool_t le = gmx_simd_cmple_r(rSimd_5_7_9, rSimd_7_8_9);
311 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(0, 0, 3), gmx_simd_blendzero_r(rSimd_1_2_3, gmx_simd_and_b(eq, le)));
314 TEST_F(SimdFloatingpointTest, gmxSimdBoolOrB)
316 gmx_simd_bool_t eq = gmx_simd_cmpeq_r(rSimd_5_7_9, rSimd_7_8_9);
317 gmx_simd_bool_t lt = gmx_simd_cmplt_r(rSimd_5_7_9, rSimd_7_8_9);
318 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(1, 2, 3), gmx_simd_blendzero_r(rSimd_1_2_3, gmx_simd_or_b(eq, lt)));
321 TEST_F(SimdFloatingpointTest, gmxSimdAnytrueB)
325 /* this test is a bit tricky since we don't know the simd width.
326 * We cannot check for truth values for "any" element beyond the first,
327 * since that part of the data will not be used if simd width is 1.
329 eq = gmx_simd_cmpeq_r(rSimd_5_7_9, setSimdRealFrom3R(5, 0, 0));
330 EXPECT_NE(0, gmx_simd_anytrue_b(eq));
332 eq = gmx_simd_cmpeq_r(rSimd_1_2_3, rSimd_4_5_6);
333 EXPECT_EQ(0, gmx_simd_anytrue_b(eq));
336 TEST_F(SimdFloatingpointTest, gmxSimdBlendvR)
338 gmx_simd_bool_t lt = gmx_simd_cmplt_r(rSimd_5_7_9, rSimd_7_8_9);
339 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(4, 5, 3), gmx_simd_blendv_r(rSimd_1_2_3, rSimd_4_5_6, lt));
342 TEST_F(SimdFloatingpointTest, gmxSimdReduceR)
344 // The horizontal sum of the SIMD variable depends on the width, so
345 // simply store it an extra time and calculate what the sum should be
346 std::vector<real> v = simdReal2Vector(rSimd_4_5_6);
349 for (int i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
354 EXPECT_EQ(sum, gmx_simd_reduce_r(rSimd_4_5_6));
357 #endif // GMX_SIMD_HAVE_REAL