1 /* This file is part of the Vc library.
3 Copyright (C) 2011 Matthias Kretz <kretz@kde.org>
5 Vc is free software: you can redistribute it and/or modify
6 it under the terms of the GNU Lesser General Public License as
7 published by the Free Software Foundation, either version 3 of
8 the License, or (at your option) any later version.
10 Vc is distributed in the hope that it will be useful, but
11 WITHOUT ANY WARRANTY; without even the implied warranty of
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 GNU Lesser General Public License for more details.
15 You should have received a copy of the GNU Lesser General Public
16 License along with Vc. If not, see <http://www.gnu.org/licenses/>.
20 #include <Vc/avx/intrinsics.h>
21 #include <Vc/avx/casts.h>
22 #include <Vc/avx/sorthelper.h>
23 #include <Vc/avx/macros.h>
31 template<> m128i SortHelper<short>::sort(VTArg _x)
33 m128i lo, hi, y, x = _x;
35 y = _mm_shufflelo_epi16(_mm_shufflehi_epi16(x, _MM_SHUFFLE(2, 3, 0, 1)), _MM_SHUFFLE(2, 3, 0, 1));
36 lo = _mm_min_epi16(x, y);
37 hi = _mm_max_epi16(x, y);
38 x = _mm_blend_epi16(lo, hi, 0xaa);
40 // merge left and right quads
41 y = _mm_shufflelo_epi16(_mm_shufflehi_epi16(x, _MM_SHUFFLE(0, 1, 2, 3)), _MM_SHUFFLE(0, 1, 2, 3));
42 lo = _mm_min_epi16(x, y);
43 hi = _mm_max_epi16(x, y);
44 x = _mm_blend_epi16(lo, hi, 0xcc);
45 y = _mm_srli_si128(x, 2);
46 lo = _mm_min_epi16(x, y);
47 hi = _mm_max_epi16(x, y);
48 x = _mm_blend_epi16(lo, _mm_slli_si128(hi, 2), 0xaa);
50 // merge quads into octs
51 y = _mm_shuffle_epi32(x, _MM_SHUFFLE(1, 0, 3, 2));
52 y = _mm_shufflelo_epi16(y, _MM_SHUFFLE(0, 1, 2, 3));
53 lo = _mm_min_epi16(x, y);
54 hi = _mm_max_epi16(x, y);
56 x = _mm_unpacklo_epi16(lo, hi);
57 y = _mm_srli_si128(x, 8);
58 lo = _mm_min_epi16(x, y);
59 hi = _mm_max_epi16(x, y);
61 x = _mm_unpacklo_epi16(lo, hi);
62 y = _mm_srli_si128(x, 8);
63 lo = _mm_min_epi16(x, y);
64 hi = _mm_max_epi16(x, y);
66 return _mm_unpacklo_epi16(lo, hi);
68 template<> m128i SortHelper<unsigned short>::sort(VTArg _x)
70 m128i lo, hi, y, x = _x;
72 y = _mm_shufflelo_epi16(_mm_shufflehi_epi16(x, _MM_SHUFFLE(2, 3, 0, 1)), _MM_SHUFFLE(2, 3, 0, 1));
73 lo = _mm_min_epu16(x, y);
74 hi = _mm_max_epu16(x, y);
75 x = _mm_blend_epi16(lo, hi, 0xaa);
77 // merge left and right quads
78 y = _mm_shufflelo_epi16(_mm_shufflehi_epi16(x, _MM_SHUFFLE(0, 1, 2, 3)), _MM_SHUFFLE(0, 1, 2, 3));
79 lo = _mm_min_epu16(x, y);
80 hi = _mm_max_epu16(x, y);
81 x = _mm_blend_epi16(lo, hi, 0xcc);
82 y = _mm_srli_si128(x, 2);
83 lo = _mm_min_epu16(x, y);
84 hi = _mm_max_epu16(x, y);
85 x = _mm_blend_epi16(lo, _mm_slli_si128(hi, 2), 0xaa);
87 // merge quads into octs
88 y = _mm_shuffle_epi32(x, _MM_SHUFFLE(1, 0, 3, 2));
89 y = _mm_shufflelo_epi16(y, _MM_SHUFFLE(0, 1, 2, 3));
90 lo = _mm_min_epu16(x, y);
91 hi = _mm_max_epu16(x, y);
93 x = _mm_unpacklo_epi16(lo, hi);
94 y = _mm_srli_si128(x, 8);
95 lo = _mm_min_epu16(x, y);
96 hi = _mm_max_epu16(x, y);
98 x = _mm_unpacklo_epi16(lo, hi);
99 y = _mm_srli_si128(x, 8);
100 lo = _mm_min_epu16(x, y);
101 hi = _mm_max_epu16(x, y);
103 return _mm_unpacklo_epi16(lo, hi);
106 template<> m256i SortHelper<int>::sort(VTArg _hgfedcba)
108 VectorType hgfedcba = _hgfedcba;
109 const m128i hgfe = hi128(hgfedcba);
110 const m128i dcba = lo128(hgfedcba);
111 m128i l = _mm_min_epi32(hgfe, dcba); // ↓hd ↓gc ↓fb ↓ea
112 m128i h = _mm_max_epi32(hgfe, dcba); // ↑hd ↑gc ↑fb ↑ea
114 m128i x = _mm_unpacklo_epi32(l, h); // ↑fb ↓fb ↑ea ↓ea
115 m128i y = _mm_unpackhi_epi32(l, h); // ↑hd ↓hd ↑gc ↓gc
117 l = _mm_min_epi32(x, y); // ↓(↑fb,↑hd) ↓hfdb ↓(↑ea,↑gc) ↓geca
118 h = _mm_max_epi32(x, y); // ↑hfdb ↑(↓fb,↓hd) ↑geca ↑(↓ea,↓gc)
120 x = _mm_min_epi32(l, Reg::permute<X2, X2, X0, X0>(h)); // 2(hfdb) 1(hfdb) 2(geca) 1(geca)
121 y = _mm_max_epi32(h, Reg::permute<X3, X3, X1, X1>(l)); // 4(hfdb) 3(hfdb) 4(geca) 3(geca)
123 m128i b = Reg::shuffle<Y0, Y1, X0, X1>(y, x); // b3 <= b2 <= b1 <= b0
124 m128i a = _mm_unpackhi_epi64(x, y); // a3 >= a2 >= a1 >= a0
126 if (VC_IS_UNLIKELY(_mm_extract_epi32(x, 2) >= _mm_extract_epi32(y, 1))) {
127 return concat(Reg::permute<X0, X1, X2, X3>(b), a);
128 } else if (VC_IS_UNLIKELY(_mm_extract_epi32(x, 0) >= _mm_extract_epi32(y, 3))) {
129 return concat(a, Reg::permute<X0, X1, X2, X3>(b));
133 l = _mm_min_epi32(a, b); // ↓a3b3 ↓a2b2 ↓a1b1 ↓a0b0
134 h = _mm_max_epi32(a, b); // ↑a3b3 ↑a2b2 ↑a1b1 ↑a0b0
136 a = _mm_unpacklo_epi32(l, h); // ↑a1b1 ↓a1b1 ↑a0b0 ↓a0b0
137 b = _mm_unpackhi_epi32(l, h); // ↑a3b3 ↓a3b3 ↑a2b2 ↓a2b2
138 l = _mm_min_epi32(a, b); // ↓(↑a1b1,↑a3b3) ↓a1b3 ↓(↑a0b0,↑a2b2) ↓a0b2
139 h = _mm_max_epi32(a, b); // ↑a3b1 ↑(↓a1b1,↓a3b3) ↑a2b0 ↑(↓a0b0,↓a2b2)
141 a = _mm_unpacklo_epi32(l, h); // ↑a2b0 ↓(↑a0b0,↑a2b2) ↑(↓a0b0,↓a2b2) ↓a0b2
142 b = _mm_unpackhi_epi32(l, h); // ↑a3b1 ↓(↑a1b1,↑a3b3) ↑(↓a1b1,↓a3b3) ↓a1b3
143 l = _mm_min_epi32(a, b); // ↓(↑a2b0,↑a3b1) ↓(↑a0b0,↑a2b2,↑a1b1,↑a3b3) ↓(↑(↓a0b0,↓a2b2) ↑(↓a1b1,↓a3b3)) ↓a0b3
144 h = _mm_max_epi32(a, b); // ↑a3b0 ↑(↓(↑a0b0,↑a2b2) ↓(↑a1b1,↑a3b3)) ↑(↓a0b0,↓a2b2,↓a1b1,↓a3b3) ↑(↓a0b2,↓a1b3)
146 return concat(_mm_unpacklo_epi32(l, h), _mm_unpackhi_epi32(l, h));
149 template<> m256i SortHelper<unsigned int>::sort(VTArg _hgfedcba)
151 VectorType hgfedcba = _hgfedcba;
152 const m128i hgfe = hi128(hgfedcba);
153 const m128i dcba = lo128(hgfedcba);
154 m128i l = _mm_min_epu32(hgfe, dcba); // ↓hd ↓gc ↓fb ↓ea
155 m128i h = _mm_max_epu32(hgfe, dcba); // ↑hd ↑gc ↑fb ↑ea
157 m128i x = _mm_unpacklo_epi32(l, h); // ↑fb ↓fb ↑ea ↓ea
158 m128i y = _mm_unpackhi_epi32(l, h); // ↑hd ↓hd ↑gc ↓gc
160 l = _mm_min_epu32(x, y); // ↓(↑fb,↑hd) ↓hfdb ↓(↑ea,↑gc) ↓geca
161 h = _mm_max_epu32(x, y); // ↑hfdb ↑(↓fb,↓hd) ↑geca ↑(↓ea,↓gc)
163 x = _mm_min_epu32(l, Reg::permute<X2, X2, X0, X0>(h)); // 2(hfdb) 1(hfdb) 2(geca) 1(geca)
164 y = _mm_max_epu32(h, Reg::permute<X3, X3, X1, X1>(l)); // 4(hfdb) 3(hfdb) 4(geca) 3(geca)
166 m128i b = Reg::shuffle<Y0, Y1, X0, X1>(y, x); // b3 <= b2 <= b1 <= b0
167 m128i a = _mm_unpackhi_epi64(x, y); // a3 >= a2 >= a1 >= a0
169 if (VC_IS_UNLIKELY(_mm_extract_epu32(x, 2) >= _mm_extract_epu32(y, 1))) {
170 return concat(Reg::permute<X0, X1, X2, X3>(b), a);
171 } else if (VC_IS_UNLIKELY(_mm_extract_epu32(x, 0) >= _mm_extract_epu32(y, 3))) {
172 return concat(a, Reg::permute<X0, X1, X2, X3>(b));
176 l = _mm_min_epu32(a, b); // ↓a3b3 ↓a2b2 ↓a1b1 ↓a0b0
177 h = _mm_max_epu32(a, b); // ↑a3b3 ↑a2b2 ↑a1b1 ↑a0b0
179 a = _mm_unpacklo_epi32(l, h); // ↑a1b1 ↓a1b1 ↑a0b0 ↓a0b0
180 b = _mm_unpackhi_epi32(l, h); // ↑a3b3 ↓a3b3 ↑a2b2 ↓a2b2
181 l = _mm_min_epu32(a, b); // ↓(↑a1b1,↑a3b3) ↓a1b3 ↓(↑a0b0,↑a2b2) ↓a0b2
182 h = _mm_max_epu32(a, b); // ↑a3b1 ↑(↓a1b1,↓a3b3) ↑a2b0 ↑(↓a0b0,↓a2b2)
184 a = _mm_unpacklo_epi32(l, h); // ↑a2b0 ↓(↑a0b0,↑a2b2) ↑(↓a0b0,↓a2b2) ↓a0b2
185 b = _mm_unpackhi_epi32(l, h); // ↑a3b1 ↓(↑a1b1,↑a3b3) ↑(↓a1b1,↓a3b3) ↓a1b3
186 l = _mm_min_epu32(a, b); // ↓(↑a2b0,↑a3b1) ↓(↑a0b0,↑a2b2,↑a1b1,↑a3b3) ↓(↑(↓a0b0,↓a2b2) ↑(↓a1b1,↓a3b3)) ↓a0b3
187 h = _mm_max_epu32(a, b); // ↑a3b0 ↑(↓(↑a0b0,↑a2b2) ↓(↑a1b1,↑a3b3)) ↑(↓a0b0,↓a2b2,↓a1b1,↓a3b3) ↑(↓a0b2,↓a1b3)
189 return concat(_mm_unpacklo_epi32(l, h), _mm_unpackhi_epi32(l, h));
192 template<> m256 SortHelper<float>::sort(VTArg _hgfedcba)
194 VectorType hgfedcba = _hgfedcba;
195 const m128 hgfe = hi128(hgfedcba);
196 const m128 dcba = lo128(hgfedcba);
197 m128 l = _mm_min_ps(hgfe, dcba); // ↓hd ↓gc ↓fb ↓ea
198 m128 h = _mm_max_ps(hgfe, dcba); // ↑hd ↑gc ↑fb ↑ea
200 m128 x = _mm_unpacklo_ps(l, h); // ↑fb ↓fb ↑ea ↓ea
201 m128 y = _mm_unpackhi_ps(l, h); // ↑hd ↓hd ↑gc ↓gc
203 l = _mm_min_ps(x, y); // ↓(↑fb,↑hd) ↓hfdb ↓(↑ea,↑gc) ↓geca
204 h = _mm_max_ps(x, y); // ↑hfdb ↑(↓fb,↓hd) ↑geca ↑(↓ea,↓gc)
206 x = _mm_min_ps(l, Reg::permute<X2, X2, X0, X0>(h)); // 2(hfdb) 1(hfdb) 2(geca) 1(geca)
207 y = _mm_max_ps(h, Reg::permute<X3, X3, X1, X1>(l)); // 4(hfdb) 3(hfdb) 4(geca) 3(geca)
209 m128 a = _mm_castpd_ps(_mm_unpackhi_pd(_mm_castps_pd(x), _mm_castps_pd(y))); // a3 >= a2 >= a1 >= a0
210 m128 b = Reg::shuffle<Y0, Y1, X0, X1>(y, x); // b3 <= b2 <= b1 <= b0
213 l = _mm_min_ps(a, b); // ↓a3b3 ↓a2b2 ↓a1b1 ↓a0b0
214 h = _mm_max_ps(a, b); // ↑a3b3 ↑a2b2 ↑a1b1 ↑a0b0
216 a = _mm_unpacklo_ps(l, h); // ↑a1b1 ↓a1b1 ↑a0b0 ↓a0b0
217 b = _mm_unpackhi_ps(l, h); // ↑a3b3 ↓a3b3 ↑a2b2 ↓a2b2
218 l = _mm_min_ps(a, b); // ↓(↑a1b1,↑a3b3) ↓a1b3 ↓(↑a0b0,↑a2b2) ↓a0b2
219 h = _mm_max_ps(a, b); // ↑a3b1 ↑(↓a1b1,↓a3b3) ↑a2b0 ↑(↓a0b0,↓a2b2)
221 a = _mm_unpacklo_ps(l, h); // ↑a2b0 ↓(↑a0b0,↑a2b2) ↑(↓a0b0,↓a2b2) ↓a0b2
222 b = _mm_unpackhi_ps(l, h); // ↑a3b1 ↓(↑a1b1,↑a3b3) ↑(↓a1b1,↓a3b3) ↓a1b3
223 l = _mm_min_ps(a, b); // ↓(↑a2b0,↑a3b1) ↓(↑a0b0,↑a2b2,↑a1b1,↑a3b3) ↓(↑(↓a0b0,↓a2b2) ↑(↓a1b1,↓a3b3)) ↓a0b3
224 h = _mm_max_ps(a, b); // ↑a3b0 ↑(↓(↑a0b0,↑a2b2) ↓(↑a1b1,↑a3b3)) ↑(↓a0b0,↓a2b2,↓a1b1,↓a3b3) ↑(↓a0b2,↓a1b3)
226 return concat(_mm_unpacklo_ps(l, h), _mm_unpackhi_ps(l, h));
229 template<> m256 SortHelper<sfloat>::sort(VTArg hgfedcba)
231 return SortHelper<float>::sort(hgfedcba);
234 template<> void SortHelper<double>::sort(m256d &VC_RESTRICT x, m256d &VC_RESTRICT y)
236 m256d l = _mm256_min_pd(x, y); // ↓x3y3 ↓x2y2 ↓x1y1 ↓x0y0
237 m256d h = _mm256_max_pd(x, y); // ↑x3y3 ↑x2y2 ↑x1y1 ↑x0y0
238 x = _mm256_unpacklo_pd(l, h); // ↑x2y2 ↓x2y2 ↑x0y0 ↓x0y0
239 y = _mm256_unpackhi_pd(l, h); // ↑x3y3 ↓x3y3 ↑x1y1 ↓x1y1
240 l = _mm256_min_pd(x, y); // ↓(↑x2y2,↑x3y3) ↓x3x2y3y2 ↓(↑x0y0,↑x1y1) ↓x1x0y1y0
241 h = _mm256_max_pd(x, y); // ↑x3x2y3y2 ↑(↓x2y2,↓x3y3) ↑x1x0y1y0 ↑(↓x0y0,↓x1y1)
242 x = _mm256_unpacklo_pd(l, h); // ↑(↓x2y2,↓x3y3) ↓x3x2y3y2 ↑(↓x0y0,↓x1y1) ↓x1x0y1y0
243 y = _mm256_unpackhi_pd(h, l); // ↓(↑x2y2,↑x3y3) ↑x3x2y3y2 ↓(↑x0y0,↑x1y1) ↑x1x0y1y0
244 l = _mm256_min_pd(x, y); // ↓(↑(↓x2y2,↓x3y3) ↓(↑x2y2,↑x3y3)) ↓x3x2y3y2 ↓(↑(↓x0y0,↓x1y1) ↓(↑x0y0,↑x1y1)) ↓x1x0y1y0
245 h = _mm256_max_pd(x, y); // ↑(↑(↓x2y2,↓x3y3) ↓(↑x2y2,↑x3y3)) ↑x3x2y3y2 ↑(↑(↓x0y0,↓x1y1) ↓(↑x0y0,↑x1y1)) ↑x1x0y1y0
246 m256d a = Reg::permute<X2, X3, X1, X0>(Reg::permute128<X0, X1>(h, h)); // h0 h1 h3 h2
247 m256d b = Reg::permute<X2, X3, X1, X0>(l); // l2 l3 l1 l0
249 // a3 >= a2 >= b1 >= b0
250 // b3 <= b2 <= a1 <= a0
253 l = _mm256_min_pd(a, b); // ↓a3b3 ↓a2b2 ↓a1b1 ↓a0b0
254 h = _mm256_min_pd(a, b); // ↑a3b3 ↑a2b2 ↑a1b1 ↑a0b0
256 x = _mm256_unpacklo_pd(l, h); // ↑a2b2 ↓a2b2 ↑a0b0 ↓a0b0
257 y = _mm256_unpackhi_pd(l, h); // ↑a3b3 ↓a3b3 ↑a1b1 ↓a1b1
258 l = _mm256_min_pd(x, y); // ↓(↑a2b2,↑a3b3) ↓a2b3 ↓(↑a0b0,↑a1b1) ↓a1b0
259 h = _mm256_min_pd(x, y); // ↑a3b2 ↑(↓a2b2,↓a3b3) ↑a0b1 ↑(↓a0b0,↓a1b1)
261 x = Reg::permute128<Y0, X0>(l, h); // ↑a0b1 ↑(↓a0b0,↓a1b1) ↓(↑a0b0,↑a1b1) ↓a1b0
262 y = Reg::permute128<Y1, X1>(l, h); // ↑a3b2 ↑(↓a2b2,↓a3b3) ↓(↑a2b2,↑a3b3) ↓a2b3
263 l = _mm256_min_pd(x, y); // ↓(↑a0b1,↑a3b2) ↓(↑(↓a0b0,↓a1b1) ↑(↓a2b2,↓a3b3)) ↓(↑a0b0,↑a1b1,↑a2b2,↑a3b3) ↓b0b3
264 h = _mm256_min_pd(x, y); // ↑a0a3 ↑(↓a0b0,↓a1b1,↓a2b2,↓a3b3) ↑(↓(↑a0b0,↑a1b1) ↓(↑a2b2,↑a3b3)) ↑(↓a1b0,↓a2b3)
266 x = _mm256_unpacklo_pd(l, h); // h2 l2 h0 l0
267 y = _mm256_unpackhi_pd(l, h); // h3 l3 h1 l1
269 template<> m256d SortHelper<double>::sort(VTArg _dcba)
271 VectorType dcba = _dcba;
273 * to find the second largest number find
274 * max(min(max(ab),max(cd)), min(max(ad),max(bc)))
276 * max(max(min(ab),min(cd)), min(max(ab),max(cd)))
278 const m256d adcb = avx_cast<m256d>(concat(_mm_alignr_epi8(avx_cast<m128i>(dc), avx_cast<m128i>(ba), 8), _mm_alignr_epi8(avx_cast<m128i>(ba), avx_cast<m128i>(dc), 8)));
279 const m256d l = _mm256_min_pd(dcba, adcb); // min(ad cd bc ab)
280 const m256d h = _mm256_max_pd(dcba, adcb); // max(ad cd bc ab)
282 // max(min(h0,h2), min(h3,h1))
283 // min(max(l0,l2), max(l3,l1))
286 const m256d ll = _mm256_min_pd(h, Reg::permute128<X0, X1>(h, h)); // min(h3h1 h2h0 h1h3 h0h2)
287 //const m256d hh = _mm256_max_pd(h3 ll1_3 l1 l0, h1 ll0_2 l3 l2);
288 const m256d hh = _mm256_max_pd(
289 Reg::permute128<X1, Y0>(_mm256_unpackhi_pd(ll, h), l),
290 Reg::permute128<X0, Y1>(_mm256_blend_pd(h ll, 0x1), l));
291 _mm256_min_pd(hh0, hh1
294 //////////////////////////////////////////////////////////////////////////////////
295 // max(max(ac), max(bd))
296 // max(max(min(ac),min(bd)), min(max(ac),max(bd)))
297 // min(max(min(ac),min(bd)), min(max(ac),max(bd)))
298 // min(min(ac), min(bd))
299 m128d l = _mm_min_pd(lo128(dcba), hi128(dcba)); // min(bd) min(ac)
300 m128d h = _mm_max_pd(lo128(dcba), hi128(dcba)); // max(bd) max(ac)
301 m128d h0_l0 = _mm_unpacklo_pd(l, h);
302 m128d h1_l1 = _mm_unpackhi_pd(l, h);
303 l = _mm_min_pd(h0_l0, h1_l1);
304 h = _mm_max_pd(h0_l0, h1_l1);
306 _mm_min_pd(l, Reg::permute<X0, X0>(h)),
307 _mm_max_pd(h, Reg::permute<X1, X1>(l))
311 // unpacklo/hi: 2 cycles
316 // ----------------------
320 m256d cdab = Reg::permute<X2, X3, X0, X1>(dcba);
321 m256d l = _mm256_min_pd(dcba, cdab);
322 m256d h = _mm256_max_pd(dcba, cdab);
323 m256d maxmin_ba = Reg::permute128<X0, Y0>(l, h);
324 m256d maxmin_dc = Reg::permute128<X1, Y1>(l, h);
326 l = _mm256_min_pd(maxmin_ba, maxmin_dc);
327 h = _mm256_max_pd(maxmin_ba, maxmin_dc);
329 return _mm256_blend_pd(h, l, 0x55);
338 y = shuffle<X1, Y0, X3, Y2>(x, x);
339 l = _mm256_min_pd(x, y); // min[ab ab cd cd]
340 h = _mm256_max_pd(x, y); // max[ab ab cd cd]
347 // don't be fooled by unpack here. It works differently for AVX pd than for SSE ps
348 x = _mm256_unpacklo_pd(l, h); // l_ab h_ab l_cd h_cd
349 l2 = _mm_min_pd(lo128(x), hi128(x)); // l_abcd l(h_ab hcd)
350 h2 = _mm_max_pd(lo128(x), hi128(x)); // h(l_ab l_cd) h_abcd
353 return concat(l2, h2);
355 // concat(_mm_unpacklo_pd(l2, h2), _mm_unpackhi_pd(l2, h2));
357 // I'd like to have four useful compares
358 const m128d dc = hi128(dcba);
359 const m128d ba = lo128(dcba);
360 const m256d adcb = avx_cast<m256d>(concat(_mm_alignr_epi8(avx_cast<m128i>(dc), avx_cast<m128i>(ba), 8), _mm_alignr_epi8(avx_cast<m128i>(ba), avx_cast<m128i>(dc), 8)));
362 const int extraCmp = _mm_movemask_pd(_mm_cmpgt_pd(dc, ba));
363 // 0x0: d <= b && c <= a
364 // 0x1: d <= b && c > a
365 // 0x2: d > b && c <= a
366 // 0x3: d > b && c > a
368 switch (_mm256_movemask_pd(_mm256_cmpgt_pd(dcba, adcb))) {
369 // impossible: 0x0, 0xf
370 case 0x1: // a <= b && b <= c && c <= d && d > a
372 return Reg::permute<X2, X3, X0, X1>(Reg::permute<X0, X1>(dcba, dcba));
373 case 0x2: // a <= b && b <= c && c > d && d <= a
375 return Reg::permute<X2, X3, X0, X1>(adcb);
376 case 0x3: // a <= b && b <= c && c > d && d > a
380 return Reg::permute<X2, X3, X1, X0>(Reg::permute<X0, X1>(dcba, dcba));
383 return Reg::permute<X3, X2, X0, X1>(adcb);
385 case 0x4: // a <= b && b > c && c <= d && d <= a
387 return Reg::permute<X2, X3, X0, X1>(dcba);
388 case 0x5: // a <= b && b > c && c <= d && d > a
391 case 0x0: // d <= b && c <= a
393 return shuffle<>(dcba, bcda);
394 case 0x1: // d <= b && c > a
395 case 0x2: // d > b && c <= a
396 case 0x3: // d > b && c > a
398 case 0x6: // a <= b && b > c && c > d && d <= a
400 case 0x7: // a <= b && b > c && c > d && d > a
402 return permute<X1, X0, X3, X2>(permute128<X1, X0>(bcda, bcda));
403 case 0x8: // a > b && b <= c && c <= d && d <= a
405 case 0x9: // a > b && b <= c && c <= d && d > a
407 case 0xa: // a > b && b <= c && c > d && d <= a
409 case 0xb: // a > b && b <= c && c > d && d > a
411 return permute128<X1, X0>(dcba);
412 case 0xc: // a > b && b > c && c <= d && d <= a
414 case 0xd: // a > b && b > c && c <= d && d > a
416 return permute<X1, X0, X3, X2>(bcda);
417 case 0xe: // a > b && b > c && c > d && d <= a
425 } // namespace AliRoot