2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2014-2018, The GROMACS development team.
5 * Copyright (c) 2019,2020,2021, 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.
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.
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.
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.
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.
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.
37 #ifndef GMX_SIMD_IMPL_X86_AVX_512_UTIL_DOUBLE_H
38 #define GMX_SIMD_IMPL_X86_AVX_512_UTIL_DOUBLE_H
45 #include <immintrin.h>
47 #include "gromacs/utility/basedefinitions.h"
49 #include "impl_x86_avx_512_general.h"
50 #include "impl_x86_avx_512_simd_double.h"
55 static const int c_simdBestPairAlignmentDouble = 2;
59 // Multiply function optimized for powers of 2, for which it is done by
60 // shifting. Currently up to 8 is accelerated. Could be accelerated for any
61 // number with a constexpr log2 function.
63 static inline SimdDInt32 fastMultiply(SimdDInt32 x)
67 return _mm256_slli_epi32(x.simdInternal_, 1);
71 return _mm256_slli_epi32(x.simdInternal_, 2);
75 return _mm256_slli_epi32(x.simdInternal_, 3);
84 static inline void gmx_simdcall gatherLoadBySimdIntTranspose(const double*, SimdDInt32)
86 // Nothing to do. Termination of recursion.
89 /* This is an internal helper function used by decr3Hsimd(...).
91 inline void gmx_simdcall decrHsimd(double* m, SimdDouble a)
95 assert(std::size_t(m) % 32 == 0);
97 a.simdInternal_ = _mm512_add_pd(a.simdInternal_,
98 _mm512_shuffle_f64x2(a.simdInternal_, a.simdInternal_, 0xEE));
99 t = _mm256_load_pd(m);
100 t = _mm256_sub_pd(t, _mm512_castpd512_pd256(a.simdInternal_));
101 _mm256_store_pd(m, t);
106 template<int align, typename... Targs>
107 static inline void gmx_simdcall
108 gatherLoadBySimdIntTranspose(const double* base, SimdDInt32 offset, SimdDouble* v, Targs... Fargs)
112 offset = fastMultiply<align>(offset);
114 constexpr size_t scale = sizeof(double);
115 v->simdInternal_ = _mm512_i32gather_pd(offset.simdInternal_, base, scale);
116 gatherLoadBySimdIntTranspose<1>(base + 1, offset, Fargs...);
119 template<int align, typename... Targs>
120 static inline void gmx_simdcall
121 gatherLoadUBySimdIntTranspose(const double* base, SimdDInt32 offset, SimdDouble* v, Targs... Fargs)
123 gatherLoadBySimdIntTranspose<align>(base, offset, v, Fargs...);
126 template<int align, typename... Targs>
127 static inline void gmx_simdcall
128 gatherLoadTranspose(const double* base, const std::int32_t offset[], SimdDouble* v, Targs... Fargs)
130 gatherLoadBySimdIntTranspose<align>(base, simdLoad(offset, SimdDInt32Tag()), v, Fargs...);
133 template<int align, typename... Targs>
134 static inline void gmx_simdcall
135 gatherLoadUTranspose(const double* base, const std::int32_t offset[], SimdDouble* v, Targs... Fargs)
137 gatherLoadBySimdIntTranspose<align>(base, simdLoad(offset, SimdDInt32Tag()), v, Fargs...);
141 static inline void gmx_simdcall transposeScatterStoreU(double* base,
142 const std::int32_t offset[],
147 SimdDInt32 simdoffset = simdLoad(offset, SimdDInt32Tag());
151 simdoffset = fastMultiply<align>(simdoffset);
154 constexpr size_t scale = sizeof(double);
155 _mm512_i32scatter_pd(base, simdoffset.simdInternal_, v0.simdInternal_, scale);
156 _mm512_i32scatter_pd(&(base[1]), simdoffset.simdInternal_, v1.simdInternal_, scale);
157 _mm512_i32scatter_pd(&(base[2]), simdoffset.simdInternal_, v2.simdInternal_, scale);
161 static inline void gmx_simdcall
162 transposeScatterIncrU(double* base, const std::int32_t offset[], SimdDouble v0, SimdDouble v1, SimdDouble v2)
164 __m512d t[4], t5, t6, t7, t8;
165 alignas(GMX_SIMD_ALIGNMENT) std::int64_t o[8];
166 // TODO: should use fastMultiply
167 _mm512_store_epi64(o,
168 _mm512_cvtepi32_epi64(_mm256_mullo_epi32(
169 _mm256_load_si256((const __m256i*)(offset)), _mm256_set1_epi32(align))));
170 t5 = _mm512_unpacklo_pd(v0.simdInternal_, v1.simdInternal_);
171 t6 = _mm512_unpackhi_pd(v0.simdInternal_, v1.simdInternal_);
172 t7 = _mm512_unpacklo_pd(v2.simdInternal_, _mm512_setzero_pd());
173 t8 = _mm512_unpackhi_pd(v2.simdInternal_, _mm512_setzero_pd());
174 t[0] = _mm512_mask_permutex_pd(t5, avx512Int2Mask(0xCC), t7, 0x4E);
175 t[1] = _mm512_mask_permutex_pd(t6, avx512Int2Mask(0xCC), t8, 0x4E);
176 t[2] = _mm512_mask_permutex_pd(t7, avx512Int2Mask(0x33), t5, 0x4E);
177 t[3] = _mm512_mask_permutex_pd(t8, avx512Int2Mask(0x33), t6, 0x4E);
180 for (int i = 0; i < 4; i++)
182 _mm512_mask_storeu_pd(base + o[0 + i],
184 _mm512_castpd256_pd512(_mm256_add_pd(_mm256_loadu_pd(base + o[0 + i]),
185 _mm512_castpd512_pd256(t[i]))));
186 _mm512_mask_storeu_pd(
189 _mm512_castpd256_pd512(_mm256_add_pd(_mm256_loadu_pd(base + o[4 + i]),
190 _mm512_extractf64x4_pd(t[i], 1))));
197 for (int i = 0; i < 4; i++)
201 _mm256_add_pd(_mm256_load_pd(base + o[0 + i]), _mm512_castpd512_pd256(t[i])));
202 _mm256_store_pd(base + o[4 + i],
203 _mm256_add_pd(_mm256_load_pd(base + o[4 + i]),
204 _mm512_extractf64x4_pd(t[i], 1)));
209 for (int i = 0; i < 4; i++)
213 _mm256_add_pd(_mm256_loadu_pd(base + o[0 + i]), _mm512_castpd512_pd256(t[i])));
214 _mm256_storeu_pd(base + o[4 + i],
215 _mm256_add_pd(_mm256_loadu_pd(base + o[4 + i]),
216 _mm512_extractf64x4_pd(t[i], 1)));
223 static inline void gmx_simdcall
224 transposeScatterDecrU(double* base, const std::int32_t offset[], SimdDouble v0, SimdDouble v1, SimdDouble v2)
226 __m512d t[4], t5, t6, t7, t8;
227 alignas(GMX_SIMD_ALIGNMENT) std::int64_t o[8];
228 // TODO: should use fastMultiply
229 _mm512_store_epi64(o,
230 _mm512_cvtepi32_epi64(_mm256_mullo_epi32(
231 _mm256_load_si256((const __m256i*)(offset)), _mm256_set1_epi32(align))));
232 t5 = _mm512_unpacklo_pd(v0.simdInternal_, v1.simdInternal_);
233 t6 = _mm512_unpackhi_pd(v0.simdInternal_, v1.simdInternal_);
234 t7 = _mm512_unpacklo_pd(v2.simdInternal_, _mm512_setzero_pd());
235 t8 = _mm512_unpackhi_pd(v2.simdInternal_, _mm512_setzero_pd());
236 t[0] = _mm512_mask_permutex_pd(t5, avx512Int2Mask(0xCC), t7, 0x4E);
237 t[2] = _mm512_mask_permutex_pd(t7, avx512Int2Mask(0x33), t5, 0x4E);
238 t[1] = _mm512_mask_permutex_pd(t6, avx512Int2Mask(0xCC), t8, 0x4E);
239 t[3] = _mm512_mask_permutex_pd(t8, avx512Int2Mask(0x33), t6, 0x4E);
242 for (int i = 0; i < 4; i++)
244 _mm512_mask_storeu_pd(base + o[0 + i],
246 _mm512_castpd256_pd512(_mm256_sub_pd(_mm256_loadu_pd(base + o[0 + i]),
247 _mm512_castpd512_pd256(t[i]))));
248 _mm512_mask_storeu_pd(
251 _mm512_castpd256_pd512(_mm256_sub_pd(_mm256_loadu_pd(base + o[4 + i]),
252 _mm512_extractf64x4_pd(t[i], 1))));
259 for (int i = 0; i < 4; i++)
263 _mm256_sub_pd(_mm256_load_pd(base + o[0 + i]), _mm512_castpd512_pd256(t[i])));
264 _mm256_store_pd(base + o[4 + i],
265 _mm256_sub_pd(_mm256_load_pd(base + o[4 + i]),
266 _mm512_extractf64x4_pd(t[i], 1)));
271 for (int i = 0; i < 4; i++)
275 _mm256_sub_pd(_mm256_loadu_pd(base + o[0 + i]), _mm512_castpd512_pd256(t[i])));
276 _mm256_storeu_pd(base + o[4 + i],
277 _mm256_sub_pd(_mm256_loadu_pd(base + o[4 + i]),
278 _mm512_extractf64x4_pd(t[i], 1)));
284 static inline void gmx_simdcall expandScalarsToTriplets(SimdDouble scalar,
285 SimdDouble* triplets0,
286 SimdDouble* triplets1,
287 SimdDouble* triplets2)
289 triplets0->simdInternal_ = _mm512_castsi512_pd(_mm512_permutexvar_epi32(
290 _mm512_set_epi32(5, 4, 5, 4, 3, 2, 3, 2, 3, 2, 1, 0, 1, 0, 1, 0),
291 _mm512_castpd_si512(scalar.simdInternal_)));
292 triplets1->simdInternal_ = _mm512_castsi512_pd(_mm512_permutexvar_epi32(
293 _mm512_set_epi32(11, 10, 9, 8, 9, 8, 9, 8, 7, 6, 7, 6, 7, 6, 5, 4),
294 _mm512_castpd_si512(scalar.simdInternal_)));
295 triplets2->simdInternal_ = _mm512_castsi512_pd(_mm512_permutexvar_epi32(
296 _mm512_set_epi32(15, 14, 15, 14, 15, 14, 13, 12, 13, 12, 13, 12, 11, 10, 11, 10),
297 _mm512_castpd_si512(scalar.simdInternal_)));
301 static inline double gmx_simdcall
302 reduceIncr4ReturnSum(double* m, SimdDouble v0, SimdDouble v1, SimdDouble v2, SimdDouble v3)
307 assert(std::size_t(m) % 32 == 0);
309 t0 = _mm512_add_pd(v0.simdInternal_, _mm512_permute_pd(v0.simdInternal_, 0x55));
310 t2 = _mm512_add_pd(v2.simdInternal_, _mm512_permute_pd(v2.simdInternal_, 0x55));
311 t0 = _mm512_mask_add_pd(
312 t0, avx512Int2Mask(0xAA), v1.simdInternal_, _mm512_permute_pd(v1.simdInternal_, 0x55));
313 t2 = _mm512_mask_add_pd(
314 t2, avx512Int2Mask(0xAA), v3.simdInternal_, _mm512_permute_pd(v3.simdInternal_, 0x55));
315 t0 = _mm512_add_pd(t0, _mm512_shuffle_f64x2(t0, t0, 0x4E));
316 t0 = _mm512_mask_add_pd(t0, avx512Int2Mask(0xF0), t2, _mm512_shuffle_f64x2(t2, t2, 0x4E));
317 t0 = _mm512_add_pd(t0, _mm512_shuffle_f64x2(t0, t0, 0xB1));
318 t0 = _mm512_mask_shuffle_f64x2(t0, avx512Int2Mask(0x0C), t0, t0, 0xEE);
320 t3 = _mm512_castpd512_pd256(t0);
321 t4 = _mm256_load_pd(m);
322 t4 = _mm256_add_pd(t4, t3);
323 _mm256_store_pd(m, t4);
325 t0 = _mm512_add_pd(t0, _mm512_permutex_pd(t0, 0x4E));
326 t0 = _mm512_add_pd(t0, _mm512_permutex_pd(t0, 0xB1));
328 return _mm_cvtsd_f64(_mm512_castpd512_pd128(t0));
331 static inline SimdDouble gmx_simdcall loadDualHsimd(const double* m0, const double* m1)
333 assert(std::size_t(m0) % 32 == 0);
334 assert(std::size_t(m1) % 32 == 0);
336 return { _mm512_insertf64x4(_mm512_castpd256_pd512(_mm256_load_pd(m0)), _mm256_load_pd(m1), 1) };
339 static inline SimdDouble gmx_simdcall loadDuplicateHsimd(const double* m)
341 assert(std::size_t(m) % 32 == 0);
343 return { _mm512_broadcast_f64x4(_mm256_load_pd(m)) };
346 static inline SimdDouble gmx_simdcall loadU1DualHsimd(const double* m)
348 return { _mm512_insertf64x4(
349 _mm512_broadcastsd_pd(_mm_load_sd(m)), _mm256_broadcastsd_pd(_mm_load_sd(m + 1)), 1) };
353 static inline void gmx_simdcall storeDualHsimd(double* m0, double* m1, SimdDouble a)
355 assert(std::size_t(m0) % 32 == 0);
356 assert(std::size_t(m1) % 32 == 0);
358 _mm256_store_pd(m0, _mm512_castpd512_pd256(a.simdInternal_));
359 _mm256_store_pd(m1, _mm512_extractf64x4_pd(a.simdInternal_, 1));
362 static inline void gmx_simdcall incrDualHsimd(double* m0, double* m1, SimdDouble a)
364 assert(std::size_t(m0) % 32 == 0);
365 assert(std::size_t(m1) % 32 == 0);
370 x = _mm256_load_pd(m0);
371 x = _mm256_add_pd(x, _mm512_castpd512_pd256(a.simdInternal_));
372 _mm256_store_pd(m0, x);
375 x = _mm256_load_pd(m1);
376 x = _mm256_add_pd(x, _mm512_extractf64x4_pd(a.simdInternal_, 1));
377 _mm256_store_pd(m1, x);
380 static inline void gmx_simdcall decr3Hsimd(double* m, SimdDouble a0, SimdDouble a1, SimdDouble a2)
383 decrHsimd(m + GMX_SIMD_DOUBLE_WIDTH / 2, a1);
384 decrHsimd(m + GMX_SIMD_DOUBLE_WIDTH, a2);
388 static inline void gmx_simdcall gatherLoadTransposeHsimd(const double* base0,
390 const std::int32_t offset[],
398 assert(std::size_t(offset) % 16 == 0);
399 assert(std::size_t(base0) % 16 == 0);
400 assert(std::size_t(base1) % 16 == 0);
402 idx0 = _mm_load_si128(reinterpret_cast<const __m128i*>(offset));
404 static_assert(align == 2 || align == 4, "If more are needed use fastMultiply");
405 idx0 = _mm_slli_epi32(idx0, align == 2 ? 1 : 2);
407 idx1 = _mm_add_epi32(idx0, _mm_set1_epi32(1));
409 idx = _mm256_inserti128_si256(_mm256_castsi128_si256(idx0), idx1, 1);
411 constexpr size_t scale = sizeof(double);
412 tmp1 = _mm512_i32gather_pd(idx, base0, scale); // TODO: Might be faster to use invidual loads
413 tmp2 = _mm512_i32gather_pd(idx, base1, scale);
415 v0->simdInternal_ = _mm512_shuffle_f64x2(tmp1, tmp2, 0x44);
416 v1->simdInternal_ = _mm512_shuffle_f64x2(tmp1, tmp2, 0xEE);
419 static inline double gmx_simdcall reduceIncr4ReturnSumHsimd(double* m, SimdDouble v0, SimdDouble v1)
424 assert(std::size_t(m) % 32 == 0);
426 t0 = _mm512_add_pd(v0.simdInternal_, _mm512_permutex_pd(v0.simdInternal_, 0x4E));
427 t0 = _mm512_mask_add_pd(
428 t0, avx512Int2Mask(0xCC), v1.simdInternal_, _mm512_permutex_pd(v1.simdInternal_, 0x4E));
429 t0 = _mm512_add_pd(t0, _mm512_permutex_pd(t0, 0xB1));
430 t0 = _mm512_mask_shuffle_f64x2(t0, avx512Int2Mask(0xAA), t0, t0, 0xEE);
432 t2 = _mm512_castpd512_pd256(t0);
433 t3 = _mm256_load_pd(m);
434 t3 = _mm256_add_pd(t3, t2);
435 _mm256_store_pd(m, t3);
437 t0 = _mm512_add_pd(t0, _mm512_permutex_pd(t0, 0x4E));
438 t0 = _mm512_add_pd(t0, _mm512_permutex_pd(t0, 0xB1));
440 return _mm_cvtsd_f64(_mm512_castpd512_pd128(t0));
443 static inline SimdDouble gmx_simdcall loadU4NOffset(const double* m, int offset)
445 return { _mm512_insertf64x4(
446 _mm512_castpd256_pd512(_mm256_loadu_pd(m)), _mm256_loadu_pd(m + offset), 1) };
451 #endif // GMX_SIMD_IMPL_X86_AVX_512_UTIL_DOUBLE_H