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.
41 #include "gromacs/math/utilities.h"
42 #include "gromacs/simd/simd.h"
43 #include "gromacs/utility/basedefinitions.h"
45 #include "testutils/testasserts.h"
61 /*! \addtogroup module_simd */
64 # if GMX_SIMD_HAVE_REAL
66 /*! \brief Test fixture for floating-point tests (identical to the generic \ref SimdTest) */
67 typedef SimdTest SimdFloatingpointTest;
69 TEST_F(SimdFloatingpointTest, setZero)
71 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(0.0), setZero());
74 TEST_F(SimdFloatingpointTest, set)
77 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(c1), SimdReal(c1));
78 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(c0), SimdReal(*p));
81 TEST_F(SimdFloatingpointTest, add)
83 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 + c3, c1 + c4, c2 + c5), rSimd_c0c1c2 + rSimd_c3c4c5);
86 TEST_F(SimdFloatingpointTest, maskAdd)
88 SimdBool m = setSimdRealFrom3R(c6, 0, c7) != setZero();
89 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 + c3, c1 + 0.0, c2 + c5),
90 maskAdd(rSimd_c0c1c2, rSimd_c3c4c5, m));
93 TEST_F(SimdFloatingpointTest, sub)
95 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 - c3, c1 - c4, c2 - c5), rSimd_c0c1c2 - rSimd_c3c4c5);
98 TEST_F(SimdFloatingpointTest, mul)
100 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 * c3, c1 * c4, c2 * c5), rSimd_c0c1c2 * rSimd_c3c4c5);
103 TEST_F(SimdFloatingpointTest, maskzMul)
105 SimdBool m = setSimdRealFrom3R(c1, 0, c1) != setZero();
106 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 * c3, 0.0, c2 * c5),
107 maskzMul(rSimd_c0c1c2, rSimd_c3c4c5, m));
110 TEST_F(SimdFloatingpointTest, fma)
112 // The last bit of FMA operations depends on hardware, so we don't require exact match
113 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 * c3 + c6, c1 * c4 + c7, c2 * c5 + c8),
114 fma(rSimd_c0c1c2, rSimd_c3c4c5, rSimd_c6c7c8));
118 TEST_F(SimdFloatingpointTest, maskzFma)
120 SimdBool m = setSimdRealFrom3R(c2, 0, c3) != setZero();
121 // The last bit of FMA operations depends on hardware, so we don't require exact match
122 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 * c3 + c6, 0.0, c2 * c5 + c8),
123 maskzFma(rSimd_c0c1c2, rSimd_c3c4c5, rSimd_c6c7c8, m));
126 TEST_F(SimdFloatingpointTest, fms)
128 // The last bit of FMA operations depends on hardware, so we don't require exact match
129 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 * c3 - c6, c1 * c4 - c7, c2 * c5 - c8),
130 fms(rSimd_c0c1c2, rSimd_c3c4c5, rSimd_c6c7c8));
133 TEST_F(SimdFloatingpointTest, fnma)
135 // The last bit of FMA operations depends on hardware, so we don't require exact match
136 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c6 - c0 * c3, c7 - c1 * c4, c8 - c2 * c5),
137 fnma(rSimd_c0c1c2, rSimd_c3c4c5, rSimd_c6c7c8));
140 TEST_F(SimdFloatingpointTest, fnms)
142 // The last bit of FMA operations depends on hardware, so we don't require exact match
143 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(-c0 * c3 - c6, -c1 * c4 - c7, -c2 * c5 - c8),
144 fnms(rSimd_c0c1c2, rSimd_c3c4c5, rSimd_c6c7c8));
147 TEST_F(SimdFloatingpointTest, abs)
149 GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, abs(rSimd_c0c1c2)); // fabs(x)=x
150 GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, abs(rSimd_m0m1m2)); // fabs(-x)=x
153 TEST_F(SimdFloatingpointTest, neg)
155 GMX_EXPECT_SIMD_REAL_EQ(rSimd_m0m1m2, -(rSimd_c0c1c2)); // fneg(x)=-x
156 GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, -(rSimd_m0m1m2)); // fneg(-x)=x
159 # if GMX_SIMD_HAVE_LOGICAL
160 TEST_F(SimdFloatingpointTest, and)
162 GMX_EXPECT_SIMD_REAL_EQ(rSimd_logicalResultAnd, (rSimd_logicalA & rSimd_logicalB));
165 TEST_F(SimdFloatingpointTest, or)
167 GMX_EXPECT_SIMD_REAL_EQ(rSimd_logicalResultOr, (rSimd_logicalA | rSimd_logicalB));
170 TEST_F(SimdFloatingpointTest, xor)
172 /* Test xor by taking xor with a number and its negative. This should result
173 * in only the sign bit being set. We then use this bit change the sign of
176 SimdReal signbit = SimdReal(c1) ^ SimdReal(-c1);
177 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-c2, c3, -c4), (signbit ^ setSimdRealFrom3R(c2, -c3, c4)));
180 TEST_F(SimdFloatingpointTest, andNot)
182 /* Use xor (which we already tested, so fix that first if both tests fail)
183 * to extract the sign bit, and then use andnot to take absolute values.
185 SimdReal signbit = SimdReal(c1) ^ SimdReal(-c1);
186 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c2, c3, c4),
187 andNot(signbit, setSimdRealFrom3R(-c2, c3, -c4)));
192 TEST_F(SimdFloatingpointTest, max)
194 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c3, c1, c4), max(rSimd_c0c1c2, rSimd_c3c0c4));
195 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c3, c1, c4), max(rSimd_c3c0c4, rSimd_c0c1c2));
196 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-c0, -c0, -c2), max(rSimd_m0m1m2, rSimd_m3m0m4));
197 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-c0, -c0, -c2), max(rSimd_m3m0m4, rSimd_m0m1m2));
200 TEST_F(SimdFloatingpointTest, min)
202 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c0, c0, c2), min(rSimd_c0c1c2, rSimd_c3c0c4));
203 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c0, c0, c2), min(rSimd_c3c0c4, rSimd_c0c1c2));
204 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-c3, -c1, -c4), min(rSimd_m0m1m2, rSimd_m3m0m4));
205 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-c3, -c1, -c4), min(rSimd_m3m0m4, rSimd_m0m1m2));
208 TEST_F(SimdFloatingpointTest, round)
210 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(2), round(rSimd_2p25));
211 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(4), round(rSimd_3p75));
212 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(-2), round(rSimd_m2p25));
213 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(-4), round(rSimd_m3p75));
216 TEST_F(SimdFloatingpointTest, roundMode)
218 /* Rounding mode needs to be consistent between round and cvtR2I */
219 SimdReal x0 = setSimdRealFrom3R(0.5, 11.5, 99.5);
220 SimdReal x1 = setSimdRealFrom3R(-0.5, -11.5, -99.5);
222 GMX_EXPECT_SIMD_REAL_EQ(round(x0), cvtI2R(cvtR2I(x0)));
223 GMX_EXPECT_SIMD_REAL_EQ(round(x1), cvtI2R(cvtR2I(x1)));
226 TEST_F(SimdFloatingpointTest, trunc)
228 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(2), trunc(rSimd_2p25));
229 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(3), trunc(rSimd_3p75));
230 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(-2), trunc(rSimd_m2p25));
231 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(-3), trunc(rSimd_m3p75));
234 // We explicitly test the exponent/mantissa routines with double precision data,
235 // since these usually rely on direct manipulation and shift of the SIMD registers,
236 // where it is easy to make mistakes with single vs double precision.
238 TEST_F(SimdFloatingpointTest, frexp)
243 fraction = frexp(rSimd_Exp, &exponent);
245 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(0.609548660288905419513128, 0.5833690139241746175358116,
246 -0.584452007502232362412542),
248 GMX_EXPECT_SIMD_INT_EQ(setSimdIntFrom3I(61, -40, 55), exponent);
251 # if GMX_SIMD_HAVE_DOUBLE && GMX_DOUBLE
252 fraction = frexp(rSimd_ExpDouble, &exponent);
254 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(0.6206306194761728178832527, 0.5236473618795619566768096,
255 -0.9280331023751380303821179),
257 GMX_EXPECT_SIMD_INT_EQ(setSimdIntFrom3I(588, -461, 673), exponent);
261 TEST_F(SimdFloatingpointTest, ldexp)
263 SimdReal one = setSimdRealFrom1R(1.0);
265 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(pow(2.0, 60.0), pow(2.0, -41.0), pow(2.0, 54.0)),
266 ldexp<MathOptimization::Unsafe>(one, setSimdIntFrom3I(60, -41, 54)));
267 # if GMX_SIMD_HAVE_DOUBLE && GMX_DOUBLE
268 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(pow(2.0, 587.0), pow(2.0, -462.0), pow(2.0, 672.0)),
269 ldexp<MathOptimization::Unsafe>(one, setSimdIntFrom3I(587, -462, 672)));
271 // The default safe version must be able to handle very negative arguments too
272 GMX_EXPECT_SIMD_REAL_EQ(setZero(), ldexp(one, setSimdIntFrom3I(-2000, -1000000, -1000000000)));
276 * We do extensive 1/sqrt(x) and 1/x accuracy testing in the math module, so
277 * we just make sure the lookup instructions appear to work here
280 TEST_F(SimdFloatingpointTest, rsqrt)
282 SimdReal x = setSimdRealFrom3R(4.0, M_PI, 1234567890.0);
283 SimdReal ref = setSimdRealFrom3R(0.5, 1.0 / std::sqrt(M_PI), 1.0 / std::sqrt(1234567890.0));
284 int shiftbits = std::numeric_limits<real>::digits - GMX_SIMD_RSQRT_BITS;
291 /* Set the allowed ulp error as 2 to the power of the number of bits in
292 * the mantissa that do not have to be correct after the table lookup.
294 setUlpTol(1LL << shiftbits);
295 GMX_EXPECT_SIMD_REAL_NEAR(ref, rsqrt(x));
298 TEST_F(SimdFloatingpointTest, maskzRsqrt)
300 SimdReal x = setSimdRealFrom3R(M_PI, -4.0, 0.0);
301 // simdCmpLe is tested separately further down
302 SimdBool m = setZero() < x;
303 SimdReal ref = setSimdRealFrom3R(1.0 / std::sqrt(M_PI), 0.0, 0.0);
304 int shiftbits = std::numeric_limits<real>::digits - GMX_SIMD_RSQRT_BITS;
311 /* Set the allowed ulp error as 2 to the power of the number of bits in
312 * the mantissa that do not have to be correct after the table lookup.
314 setUlpTol(1LL << shiftbits);
315 GMX_EXPECT_SIMD_REAL_NEAR(ref, maskzRsqrt(x, m));
318 TEST_F(SimdFloatingpointTest, rcp)
320 SimdReal x = setSimdRealFrom3R(4.0, M_PI, 1234567890.0);
321 SimdReal ref = setSimdRealFrom3R(0.25, 1.0 / M_PI, 1.0 / 1234567890.0);
322 int shiftbits = std::numeric_limits<real>::digits - GMX_SIMD_RCP_BITS;
329 /* Set the allowed ulp error as 2 to the power of the number of bits in
330 * the mantissa that do not have to be correct after the table lookup.
332 setUlpTol(1LL << shiftbits);
333 GMX_EXPECT_SIMD_REAL_NEAR(ref, rcp(x));
336 TEST_F(SimdFloatingpointTest, maskzRcp)
338 SimdReal x = setSimdRealFrom3R(M_PI, 0.0, -1234567890.0);
339 SimdBool m = (x != setZero());
340 SimdReal ref = setSimdRealFrom3R(1.0 / M_PI, 0.0, -1.0 / 1234567890.0);
341 int shiftbits = std::numeric_limits<real>::digits - GMX_SIMD_RCP_BITS;
348 /* Set the allowed ulp error as 2 to the power of the number of bits in
349 * the mantissa that do not have to be correct after the table lookup.
351 setUlpTol(1LL << shiftbits);
352 GMX_EXPECT_SIMD_REAL_NEAR(ref, maskzRcp(x, m));
355 TEST_F(SimdFloatingpointTest, cmpEqAndSelectByMask)
357 SimdBool eq = rSimd_c4c6c8 == rSimd_c6c7c8;
358 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(0, 0, c2), selectByMask(rSimd_c0c1c2, eq));
361 TEST_F(SimdFloatingpointTest, selectByNotMask)
363 SimdBool eq = rSimd_c4c6c8 == rSimd_c6c7c8;
364 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c0, c1, 0), selectByNotMask(rSimd_c0c1c2, eq));
367 TEST_F(SimdFloatingpointTest, cmpNe)
369 SimdBool eq = rSimd_c4c6c8 != rSimd_c6c7c8;
370 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c0, c1, 0), selectByMask(rSimd_c0c1c2, eq));
373 TEST_F(SimdFloatingpointTest, cmpLe)
375 SimdBool le = rSimd_c4c6c8 <= rSimd_c6c7c8;
376 GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, selectByMask(rSimd_c0c1c2, le));
379 TEST_F(SimdFloatingpointTest, cmpLt)
381 SimdBool lt = rSimd_c4c6c8 < rSimd_c6c7c8;
382 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c0, c1, 0), selectByMask(rSimd_c0c1c2, lt));
385 # if GMX_SIMD_HAVE_INT32_LOGICAL || GMX_SIMD_HAVE_LOGICAL
386 TEST_F(SimdFloatingpointTest, testBits)
388 SimdBool eq = testBits(setSimdRealFrom3R(c1, 0, c1));
389 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c0, 0, c2), selectByMask(rSimd_c0c1c2, eq));
391 // Test if we detect only the sign bit being set
392 eq = testBits(setSimdRealFrom1R(GMX_REAL_NEGZERO));
393 GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, selectByMask(rSimd_c0c1c2, eq));
397 TEST_F(SimdFloatingpointTest, andB)
399 SimdBool eq = rSimd_c4c6c8 == rSimd_c6c7c8;
400 SimdBool le = rSimd_c4c6c8 <= rSimd_c6c7c8;
401 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(0, 0, c2), selectByMask(rSimd_c0c1c2, (eq && le)));
404 TEST_F(SimdFloatingpointTest, orB)
406 SimdBool eq = rSimd_c4c6c8 == rSimd_c6c7c8;
407 SimdBool lt = rSimd_c4c6c8 < rSimd_c6c7c8;
408 GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, selectByMask(rSimd_c0c1c2, (eq || lt)));
411 TEST_F(SimdFloatingpointTest, anyTrueB)
413 alignas(GMX_SIMD_ALIGNMENT) std::array<real, GMX_SIMD_REAL_WIDTH> mem{};
415 // Test the false case
416 EXPECT_FALSE(anyTrue(setZero() < load<SimdReal>(mem.data())));
418 // Test each bit (these should all be true)
419 for (int i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
423 EXPECT_TRUE(anyTrue(setZero() < load<SimdReal>(mem.data())))
424 << "Not detecting true in element " << i;
428 TEST_F(SimdFloatingpointTest, blend)
430 SimdBool lt = rSimd_c4c6c8 < rSimd_c6c7c8;
431 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c3, c4, c2), blend(rSimd_c0c1c2, rSimd_c3c4c5, lt));
434 TEST_F(SimdFloatingpointTest, reduce)
436 // The horizontal sum of the SIMD variable depends on the width, so
437 // simply store it an extra time and calculate what the sum should be
438 std::vector<real> v = simdReal2Vector(rSimd_c3c4c5);
441 for (int i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
446 EXPECT_REAL_EQ_TOL(sum, reduce(rSimd_c3c4c5), defaultRealTolerance());
449 # endif // GMX_SIMD_HAVE_REAL
451 # if GMX_SIMD_HAVE_FLOAT && GMX_SIMD_HAVE_DOUBLE
452 TEST_F(SimdFloatingpointTest, cvtFloat2Double)
454 alignas(GMX_SIMD_ALIGNMENT) float f[GMX_SIMD_FLOAT_WIDTH];
455 alignas(GMX_SIMD_ALIGNMENT) double d[GMX_SIMD_FLOAT_WIDTH]; // Yes, double array length should be same as float
460 FloatingPointTolerance tolerance(defaultRealTolerance());
462 for (i = 0; i < GMX_SIMD_FLOAT_WIDTH; i++)
464 // Scale by 1+100*eps to use low bits too.
465 // Due to the conversions we want to avoid being too sensitive to fluctuations in last bit
466 f[i] = i * (1.0 + 100 * GMX_FLOAT_EPS);
469 vf = load<SimdFloat>(f);
470 # if (GMX_SIMD_FLOAT_WIDTH == 2 * GMX_SIMD_DOUBLE_WIDTH)
472 cvtF2DD(vf, &vd0, &vd1);
473 store(d + GMX_SIMD_DOUBLE_WIDTH, vd1); // Store upper part halfway through array
474 # elif (GMX_SIMD_FLOAT_WIDTH == GMX_SIMD_DOUBLE_WIDTH)
477 # error Width of float SIMD must either be identical to double, or twice the width.
479 store(d, vd0); // store lower (or whole) part from start of vector
481 for (i = 0; i < GMX_SIMD_FLOAT_WIDTH; i++)
483 EXPECT_REAL_EQ_TOL(f[i], d[i], tolerance);
487 TEST_F(SimdFloatingpointTest, cvtDouble2Float)
489 alignas(GMX_SIMD_ALIGNMENT) float f[GMX_SIMD_FLOAT_WIDTH];
490 alignas(GMX_SIMD_ALIGNMENT) double d[GMX_SIMD_FLOAT_WIDTH]; // Yes, double array length should be same as float
494 FloatingPointTolerance tolerance(defaultRealTolerance());
496 // This fills elements for pd1 too when double width is 2*single width
497 for (i = 0; i < GMX_SIMD_FLOAT_WIDTH; i++)
499 // Scale by 1+eps to use low bits too.
500 // Due to the conversions we want to avoid being too sensitive to fluctuations in last bit
501 d[i] = i * (1.0 + 100 * GMX_FLOAT_EPS);
504 vd0 = load<SimdDouble>(d);
505 # if (GMX_SIMD_FLOAT_WIDTH == 2 * GMX_SIMD_DOUBLE_WIDTH)
506 SimdDouble vd1 = load<SimdDouble>(d + GMX_SIMD_DOUBLE_WIDTH); // load upper half of data
507 vf = cvtDD2F(vd0, vd1);
508 # elif (GMX_SIMD_FLOAT_WIDTH == GMX_SIMD_DOUBLE_WIDTH)
511 # error Width of float SIMD must either be identical to double, or twice the width.
515 // This will check elements in pd1 too when double width is 2*single width
516 for (i = 0; i < GMX_SIMD_FLOAT_WIDTH; i++)
518 EXPECT_FLOAT_EQ_TOL(d[i], f[i], tolerance);
521 # endif // GMX_SIMD_HAVE_FLOAT && GMX_SIMD_HAVE_DOUBLE