2 Stockfish, a UCI chess playing engine derived from Glaurung 2.1
3 Copyright (C) 2004-2023 The Stockfish developers (see AUTHORS file)
5 Stockfish is free software: you can redistribute it and/or modify
6 it under the terms of the GNU General Public License as published by
7 the Free Software Foundation, either version 3 of the License, or
8 (at your option) any later version.
10 Stockfish is distributed in the hope that it will be useful,
11 but WITHOUT ANY WARRANTY; without even the implied warranty of
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 GNU General Public License for more details.
15 You should have received a copy of the GNU General Public License
16 along with this program. If not, see <http://www.gnu.org/licenses/>.
19 #ifndef STOCKFISH_SIMD_H_INCLUDED
20 #define STOCKFISH_SIMD_H_INCLUDED
23 # include <immintrin.h>
25 #elif defined(USE_SSE41)
26 # include <smmintrin.h>
28 #elif defined(USE_SSSE3)
29 # include <tmmintrin.h>
31 #elif defined(USE_SSE2)
32 # include <emmintrin.h>
34 #elif defined(USE_MMX)
35 # include <mmintrin.h>
37 #elif defined(USE_NEON)
38 # include <arm_neon.h>
41 namespace Stockfish::Simd {
43 #if defined (USE_AVX512)
45 [[maybe_unused]] static int m512_hadd(__m512i sum, int bias) {
46 return _mm512_reduce_add_epi32(sum) + bias;
51 sum0 = [zmm0.i128[0], zmm0.i128[1], zmm0.i128[2], zmm0.i128[3]]
52 sum1 = [zmm1.i128[0], zmm1.i128[1], zmm1.i128[2], zmm1.i128[3]]
53 sum2 = [zmm2.i128[0], zmm2.i128[1], zmm2.i128[2], zmm2.i128[3]]
54 sum3 = [zmm3.i128[0], zmm3.i128[1], zmm3.i128[2], zmm3.i128[3]]
58 reduce_add_epi32(zmm0.i128[0]), reduce_add_epi32(zmm1.i128[0]), reduce_add_epi32(zmm2.i128[0]), reduce_add_epi32(zmm3.i128[0]),
59 reduce_add_epi32(zmm0.i128[1]), reduce_add_epi32(zmm1.i128[1]), reduce_add_epi32(zmm2.i128[1]), reduce_add_epi32(zmm3.i128[1]),
60 reduce_add_epi32(zmm0.i128[2]), reduce_add_epi32(zmm1.i128[2]), reduce_add_epi32(zmm2.i128[2]), reduce_add_epi32(zmm3.i128[2]),
61 reduce_add_epi32(zmm0.i128[3]), reduce_add_epi32(zmm1.i128[3]), reduce_add_epi32(zmm2.i128[3]), reduce_add_epi32(zmm3.i128[3])
64 [[maybe_unused]] static __m512i m512_hadd128x16_interleave(
65 __m512i sum0, __m512i sum1, __m512i sum2, __m512i sum3) {
67 __m512i sum01a = _mm512_unpacklo_epi32(sum0, sum1);
68 __m512i sum01b = _mm512_unpackhi_epi32(sum0, sum1);
70 __m512i sum23a = _mm512_unpacklo_epi32(sum2, sum3);
71 __m512i sum23b = _mm512_unpackhi_epi32(sum2, sum3);
73 __m512i sum01 = _mm512_add_epi32(sum01a, sum01b);
74 __m512i sum23 = _mm512_add_epi32(sum23a, sum23b);
76 __m512i sum0123a = _mm512_unpacklo_epi64(sum01, sum23);
77 __m512i sum0123b = _mm512_unpackhi_epi64(sum01, sum23);
79 return _mm512_add_epi32(sum0123a, sum0123b);
82 [[maybe_unused]] static __m128i m512_haddx4(
83 __m512i sum0, __m512i sum1, __m512i sum2, __m512i sum3,
86 __m512i sum = m512_hadd128x16_interleave(sum0, sum1, sum2, sum3);
88 __m256i sum256lo = _mm512_castsi512_si256(sum);
89 __m256i sum256hi = _mm512_extracti64x4_epi64(sum, 1);
91 sum256lo = _mm256_add_epi32(sum256lo, sum256hi);
93 __m128i sum128lo = _mm256_castsi256_si128(sum256lo);
94 __m128i sum128hi = _mm256_extracti128_si256(sum256lo, 1);
96 return _mm_add_epi32(_mm_add_epi32(sum128lo, sum128hi), bias);
99 [[maybe_unused]] static void m512_add_dpbusd_epi32(
104 # if defined (USE_VNNI)
105 acc = _mm512_dpbusd_epi32(acc, a, b);
107 __m512i product0 = _mm512_maddubs_epi16(a, b);
108 product0 = _mm512_madd_epi16(product0, _mm512_set1_epi16(1));
109 acc = _mm512_add_epi32(acc, product0);
113 [[maybe_unused]] static void m512_add_dpbusd_epi32x2(
115 __m512i a0, __m512i b0,
116 __m512i a1, __m512i b1) {
118 # if defined (USE_VNNI)
119 acc = _mm512_dpbusd_epi32(acc, a0, b0);
120 acc = _mm512_dpbusd_epi32(acc, a1, b1);
122 __m512i product0 = _mm512_maddubs_epi16(a0, b0);
123 __m512i product1 = _mm512_maddubs_epi16(a1, b1);
124 product0 = _mm512_madd_epi16(product0, _mm512_set1_epi16(1));
125 product1 = _mm512_madd_epi16(product1, _mm512_set1_epi16(1));
126 acc = _mm512_add_epi32(acc, _mm512_add_epi32(product0, product1));
132 #if defined (USE_AVX2)
134 [[maybe_unused]] static int m256_hadd(__m256i sum, int bias) {
135 __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(sum), _mm256_extracti128_si256(sum, 1));
136 sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_PERM_BADC));
137 sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_PERM_CDAB));
138 return _mm_cvtsi128_si32(sum128) + bias;
141 [[maybe_unused]] static __m128i m256_haddx4(
142 __m256i sum0, __m256i sum1, __m256i sum2, __m256i sum3,
145 sum0 = _mm256_hadd_epi32(sum0, sum1);
146 sum2 = _mm256_hadd_epi32(sum2, sum3);
148 sum0 = _mm256_hadd_epi32(sum0, sum2);
150 __m128i sum128lo = _mm256_castsi256_si128(sum0);
151 __m128i sum128hi = _mm256_extracti128_si256(sum0, 1);
153 return _mm_add_epi32(_mm_add_epi32(sum128lo, sum128hi), bias);
156 [[maybe_unused]] static void m256_add_dpbusd_epi32(
161 # if defined (USE_VNNI)
162 acc = _mm256_dpbusd_epi32(acc, a, b);
164 __m256i product0 = _mm256_maddubs_epi16(a, b);
165 product0 = _mm256_madd_epi16(product0, _mm256_set1_epi16(1));
166 acc = _mm256_add_epi32(acc, product0);
170 [[maybe_unused]] static void m256_add_dpbusd_epi32x2(
172 __m256i a0, __m256i b0,
173 __m256i a1, __m256i b1) {
175 # if defined (USE_VNNI)
176 acc = _mm256_dpbusd_epi32(acc, a0, b0);
177 acc = _mm256_dpbusd_epi32(acc, a1, b1);
179 __m256i product0 = _mm256_maddubs_epi16(a0, b0);
180 __m256i product1 = _mm256_maddubs_epi16(a1, b1);
181 product0 = _mm256_madd_epi16(product0, _mm256_set1_epi16(1));
182 product1 = _mm256_madd_epi16(product1, _mm256_set1_epi16(1));
183 acc = _mm256_add_epi32(acc, _mm256_add_epi32(product0, product1));
189 #if defined (USE_SSSE3)
191 [[maybe_unused]] static int m128_hadd(__m128i sum, int bias) {
192 sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0x4E)); //_MM_PERM_BADC
193 sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0xB1)); //_MM_PERM_CDAB
194 return _mm_cvtsi128_si32(sum) + bias;
197 [[maybe_unused]] static __m128i m128_haddx4(
198 __m128i sum0, __m128i sum1, __m128i sum2, __m128i sum3,
201 sum0 = _mm_hadd_epi32(sum0, sum1);
202 sum2 = _mm_hadd_epi32(sum2, sum3);
203 sum0 = _mm_hadd_epi32(sum0, sum2);
204 return _mm_add_epi32(sum0, bias);
207 [[maybe_unused]] static void m128_add_dpbusd_epi32(
212 __m128i product0 = _mm_maddubs_epi16(a, b);
213 product0 = _mm_madd_epi16(product0, _mm_set1_epi16(1));
214 acc = _mm_add_epi32(acc, product0);
217 [[maybe_unused]] static void m128_add_dpbusd_epi32x2(
219 __m128i a0, __m128i b0,
220 __m128i a1, __m128i b1) {
222 __m128i product0 = _mm_maddubs_epi16(a0, b0);
223 __m128i product1 = _mm_maddubs_epi16(a1, b1);
224 product0 = _mm_madd_epi16(product0, _mm_set1_epi16(1));
225 product1 = _mm_madd_epi16(product1, _mm_set1_epi16(1));
226 acc = _mm_add_epi32(acc, _mm_add_epi32(product0, product1));
231 #if defined (USE_NEON_DOTPROD)
233 [[maybe_unused]] static void dotprod_m128_add_dpbusd_epi32x2(
235 int8x16_t a0, int8x16_t b0,
236 int8x16_t a1, int8x16_t b1) {
238 acc = vdotq_s32(acc, a0, b0);
239 acc = vdotq_s32(acc, a1, b1);
242 [[maybe_unused]] static void dotprod_m128_add_dpbusd_epi32(
244 int8x16_t a, int8x16_t b) {
246 acc = vdotq_s32(acc, a, b);
250 #if defined (USE_NEON)
252 [[maybe_unused]] static int neon_m128_reduce_add_epi32(int32x4_t s) {
254 return vaddvq_s32(s);
256 return s[0] + s[1] + s[2] + s[3];
260 [[maybe_unused]] static int neon_m128_hadd(int32x4_t sum, int bias) {
261 return neon_m128_reduce_add_epi32(sum) + bias;
264 [[maybe_unused]] static int32x4_t neon_m128_haddx4(
265 int32x4_t sum0, int32x4_t sum1, int32x4_t sum2, int32x4_t sum3,
269 neon_m128_reduce_add_epi32(sum0),
270 neon_m128_reduce_add_epi32(sum1),
271 neon_m128_reduce_add_epi32(sum2),
272 neon_m128_reduce_add_epi32(sum3)
274 return vaddq_s32(hsums, bias);
277 [[maybe_unused]] static void neon_m128_add_dpbusd_epi32x2(
279 int8x8_t a0, int8x8_t b0,
280 int8x8_t a1, int8x8_t b1) {
282 int16x8_t product = vmull_s8(a0, b0);
283 product = vmlal_s8(product, a1, b1);
284 acc = vpadalq_s16(acc, product);
289 [[maybe_unused]] static void neon_m128_add_dpbusd_epi32(
291 int8x16_t a, int8x16_t b) {
293 int16x8_t product0 = vmull_s8(vget_low_s8(a), vget_low_s8(b));
294 int16x8_t product1 = vmull_high_s8(a, b);
295 int16x8_t sum = vpaddq_s16(product0, product1);
296 acc = vpadalq_s16(acc, sum);
301 #endif // STOCKFISH_SIMD_H_INCLUDED