2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2014,2015,2016,2017,2018,2019, 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.
39 #include "gromacs/math/utilities.h"
40 #include "gromacs/simd/simd.h"
41 #include "gromacs/utility/basedefinitions.h"
43 #include "testutils/testasserts.h"
58 /*! \addtogroup module_simd */
61 # if GMX_SIMD4_HAVE_REAL
63 /*! \brief Test fixture for SIMD4 floating-point operations (identical to the SIMD4 \ref Simd4Test) */
64 typedef Simd4Test Simd4FloatingpointTest;
66 TEST_F(Simd4FloatingpointTest, setZero)
68 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(0.0), setZero());
71 TEST_F(Simd4FloatingpointTest, set)
73 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(c1), Simd4Real(c1));
76 TEST_F(Simd4FloatingpointTest, add)
78 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0 + c3, c1 + c4, c2 + c5), rSimd4_c0c1c2 + rSimd4_c3c4c5);
81 TEST_F(Simd4FloatingpointTest, sub)
83 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0 - c3, c1 - c4, c2 - c5), rSimd4_c0c1c2 - rSimd4_c3c4c5);
86 TEST_F(Simd4FloatingpointTest, mul)
88 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0 * c3, c1 * c4, c2 * c5), rSimd4_c0c1c2 * rSimd4_c3c4c5);
91 TEST_F(Simd4FloatingpointTest, fma)
93 // The last bit of FMA operations depends on hardware, so we don't require exact match
94 GMX_EXPECT_SIMD4_REAL_NEAR(setSimd4RealFrom3R(c0 * c3 + c6, c1 * c4 + c7, c2 * c5 + c8),
95 fma(rSimd4_c0c1c2, rSimd4_c3c4c5, rSimd4_c6c7c8));
98 TEST_F(Simd4FloatingpointTest, fms)
100 // The last bit of FMA operations depends on hardware, so we don't require exact match
101 GMX_EXPECT_SIMD4_REAL_NEAR(setSimd4RealFrom3R(c0 * c3 - c6, c1 * c4 - c7, c2 * c5 - c8),
102 fms(rSimd4_c0c1c2, rSimd4_c3c4c5, rSimd4_c6c7c8));
105 TEST_F(Simd4FloatingpointTest, fnma)
107 // The last bit of FMA operations depends on hardware, so we don't require exact match
108 GMX_EXPECT_SIMD4_REAL_NEAR(setSimd4RealFrom3R(c6 - c0 * c3, c7 - c1 * c4, c8 - c2 * c5),
109 fnma(rSimd4_c0c1c2, rSimd4_c3c4c5, rSimd4_c6c7c8));
112 TEST_F(Simd4FloatingpointTest, fnms)
114 // The last bit of FMA operations depends on hardware, so we don't require exact match
115 GMX_EXPECT_SIMD4_REAL_NEAR(setSimd4RealFrom3R(-c0 * c3 - c6, -c1 * c4 - c7, -c2 * c5 - c8),
116 fnms(rSimd4_c0c1c2, rSimd4_c3c4c5, rSimd4_c6c7c8));
119 TEST_F(Simd4FloatingpointTest, abs)
121 GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_c0c1c2, abs(rSimd4_c0c1c2)); // fabs(x)=x
122 GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_c0c1c2, abs(rSimd4_m0m1m2)); // fabs(-x)=x
125 TEST_F(Simd4FloatingpointTest, neg)
127 GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_m0m1m2, -(rSimd4_c0c1c2)); // fneg(x)=-x
128 GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_c0c1c2, -(rSimd4_m0m1m2)); // fneg(-x)=x
131 # if GMX_SIMD_HAVE_LOGICAL
132 TEST_F(Simd4FloatingpointTest, and)
134 GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_logicalResultAnd, (rSimd4_logicalA & rSimd4_logicalB));
137 TEST_F(Simd4FloatingpointTest, or)
139 GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_logicalResultOr, (rSimd4_logicalA | rSimd4_logicalB));
142 TEST_F(Simd4FloatingpointTest, xor)
144 /* Test xor by taking xor with a number and its negative. This should result
145 * in only the sign bit being set. We then use this bit change the sign of
148 Simd4Real signbit = Simd4Real(c1) ^ Simd4Real(-c1);
149 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-c2, c3, -c4), signbit ^ setSimd4RealFrom3R(c2, -c3, c4));
152 TEST_F(Simd4FloatingpointTest, andNot)
154 /* Use xor (which we already tested, so fix that first if both tests fail)
155 * to extract the sign bit, and then use andnot to take absolute values.
157 Simd4Real signbit = Simd4Real(c1) ^ Simd4Real(-c1);
158 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c2, c3, c4),
159 andNot(signbit, setSimd4RealFrom3R(-c2, c3, -c4)));
164 TEST_F(Simd4FloatingpointTest, max)
166 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c3, c1, c4), max(rSimd4_c0c1c2, rSimd4_c3c0c4));
167 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c3, c1, c4), max(rSimd4_c3c0c4, rSimd4_c0c1c2));
168 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-c0, -c0, -c2), max(rSimd4_m0m1m2, rSimd4_m3m0m4));
169 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-c0, -c0, -c2), max(rSimd4_m3m0m4, rSimd4_m0m1m2));
172 TEST_F(Simd4FloatingpointTest, min)
174 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0, c0, c2), min(rSimd4_c0c1c2, rSimd4_c3c0c4));
175 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0, c0, c2), min(rSimd4_c3c0c4, rSimd4_c0c1c2));
176 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-c3, -c1, -c4), min(rSimd4_m0m1m2, rSimd4_m3m0m4));
177 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-c3, -c1, -c4), min(rSimd4_m3m0m4, rSimd4_m0m1m2));
180 TEST_F(Simd4FloatingpointTest, round)
182 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(2), round(Simd4Real(2.25)));
183 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(4), round(Simd4Real(3.75)));
184 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-2), round(Simd4Real(-2.25)));
185 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-4), round(Simd4Real(-3.75)));
188 TEST_F(Simd4FloatingpointTest, trunc)
190 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(2), trunc(rSimd4_2p25));
191 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(3), trunc(rSimd4_3p75));
192 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-2), trunc(rSimd4_m2p25));
193 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-3), trunc(rSimd4_m3p75));
196 /* We do extensive 1/sqrt(x) and 1/x accuracy testing in the tests for
197 * the SIMD math functions, so we just make sure the lookup instructions
198 * appear to work for a few values here.
200 TEST_F(Simd4FloatingpointTest, gmxSimd4RsqrtR)
202 Simd4Real x = setSimd4RealFrom3R(4.0, M_PI, 1234567890.0);
203 Simd4Real ref = setSimd4RealFrom3R(0.5, 1.0 / std::sqrt(M_PI), 1.0 / std::sqrt(1234567890.0));
204 int shiftbits = std::numeric_limits<real>::digits - GMX_SIMD_RSQRT_BITS;
211 // The allowed Ulp deviation is 2 to the power of the number of mantissa
212 // digits, minus the number of bits provided by the table lookup
213 setUlpTol(1LL << shiftbits);
214 GMX_EXPECT_SIMD4_REAL_NEAR(ref, rsqrt(x));
217 TEST_F(Simd4FloatingpointTest, cmpEqAndSelectByMask)
219 Simd4Bool eq = rSimd4_c4c6c8 == rSimd4_c6c7c8;
220 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(0, 0, c2), selectByMask(rSimd4_c0c1c2, eq));
223 TEST_F(Simd4FloatingpointTest, selectByNotMask)
225 Simd4Bool eq = rSimd4_c4c6c8 == rSimd4_c6c7c8;
226 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0, c1, 0), selectByNotMask(rSimd4_c0c1c2, eq));
229 TEST_F(Simd4FloatingpointTest, cmpNe)
231 Simd4Bool eq = rSimd4_c4c6c8 != rSimd4_c6c7c8;
232 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0, c1, 0), selectByMask(rSimd4_c0c1c2, eq));
235 TEST_F(Simd4FloatingpointTest, cmpLe)
237 Simd4Bool le = rSimd4_c4c6c8 <= rSimd4_c6c7c8;
238 GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_c0c1c2, selectByMask(rSimd4_c0c1c2, le));
241 TEST_F(Simd4FloatingpointTest, cmpLt)
243 Simd4Bool lt = rSimd4_c4c6c8 < rSimd4_c6c7c8;
244 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0, c1, 0), selectByMask(rSimd4_c0c1c2, lt));
247 TEST_F(Simd4FloatingpointTest, andB)
249 Simd4Bool eq = rSimd4_c4c6c8 == rSimd4_c6c7c8;
250 Simd4Bool le = rSimd4_c4c6c8 <= rSimd4_c6c7c8;
251 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(0, 0, c2), selectByMask(rSimd4_c0c1c2, (eq && le)));
254 TEST_F(Simd4FloatingpointTest, orB)
256 Simd4Bool eq = rSimd4_c4c6c8 == rSimd4_c6c7c8;
257 Simd4Bool lt = rSimd4_c4c6c8 < rSimd4_c6c7c8;
258 GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_c0c1c2, selectByMask(rSimd4_c0c1c2, (eq || lt)));
261 TEST_F(Simd4FloatingpointTest, anyTrue)
265 /* this test is a bit tricky since we don't know the simd width.
266 * We cannot check for truth values for "any" element beyond the first,
267 * since that part of the data will not be used if simd width is 1.
269 eq = (rSimd4_c4c6c8 == setSimd4RealFrom3R(c4, 0, 0));
270 EXPECT_TRUE(anyTrue(eq));
272 eq = (rSimd4_c0c1c2 == rSimd4_c3c4c5);
273 EXPECT_FALSE(anyTrue(eq));
276 TEST_F(Simd4FloatingpointTest, blend)
278 Simd4Bool lt = rSimd4_c4c6c8 < rSimd4_c6c7c8;
279 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c3, c4, c2), blend(rSimd4_c0c1c2, rSimd4_c3c4c5, lt));
282 TEST_F(Simd4FloatingpointTest, reduce)
284 // The horizontal sum of the SIMD variable depends on the width, so
285 // simply store it an extra time and calculate what the sum should be
286 std::vector<real> v = simd4Real2Vector(rSimd4_c3c4c5);
289 for (int i = 0; i < GMX_SIMD4_WIDTH; i++)
294 EXPECT_REAL_EQ_TOL(sum, reduce(rSimd4_c3c4c5), defaultRealTolerance());
298 TEST_F(Simd4FloatingpointTest, dotProduct)
300 real res = c0 * c3 + c1 * c4 + c2 * c5;
302 EXPECT_REAL_EQ_TOL(res, dotProduct(rSimd4_c0c1c2, rSimd4_c3c4c5), defaultRealTolerance());
305 TEST_F(Simd4FloatingpointTest, transpose)
307 Simd4Real v0, v1, v2, v3;
310 alignas(GMX_SIMD_ALIGNMENT) real p0[4 * GMX_SIMD4_WIDTH];
311 real* p1 = p0 + GMX_SIMD4_WIDTH;
312 real* p2 = p0 + 2 * GMX_SIMD4_WIDTH;
313 real* p3 = p0 + 3 * GMX_SIMD4_WIDTH;
315 // Assign data with tens as row, single-digit as column
316 for (i = 0; i < 4; i++)
318 // Scale by 1+100*eps to use low bits tii
319 p0[i] = (0 * 10 + i * 1) * (1.0 + 100 * GMX_REAL_EPS);
320 p1[i] = (1 * 10 + i * 1) * (1.0 + 100 * GMX_REAL_EPS);
321 p2[i] = (2 * 10 + i * 1) * (1.0 + 100 * GMX_REAL_EPS);
322 p3[i] = (3 * 10 + i * 1) * (1.0 + 100 * GMX_REAL_EPS);
330 transpose(&v0, &v1, &v2, &v3);
337 for (i = 0; i < 4; i++)
339 EXPECT_REAL_EQ_TOL((i * 10 + 0) * (1.0 + 100 * GMX_REAL_EPS), p0[i], defaultRealTolerance());
340 EXPECT_REAL_EQ_TOL((i * 10 + 1) * (1.0 + 100 * GMX_REAL_EPS), p1[i], defaultRealTolerance());
341 EXPECT_REAL_EQ_TOL((i * 10 + 2) * (1.0 + 100 * GMX_REAL_EPS), p2[i], defaultRealTolerance());
342 EXPECT_REAL_EQ_TOL((i * 10 + 3) * (1.0 + 100 * GMX_REAL_EPS), p3[i], defaultRealTolerance());
346 # endif // GMX_SIMD4_HAVE_REAL