File: aligner_swsse_ee_i16.cpp

package info (click to toggle)
hisat2 2.1.0-2
  • links: PTS, VCS
  • area: main
  • in suites: bullseye, buster, sid
  • size: 13,756 kB
  • sloc: cpp: 86,309; python: 12,230; sh: 2,171; perl: 936; makefile: 375
file content (1911 lines) | stat: -rw-r--r-- 63,619 bytes parent folder | download
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
/*
 * Copyright 2011, Ben Langmead <langmea@cs.jhu.edu>
 *
 * This file is part of Bowtie 2.
 *
 * Bowtie 2 is free software: you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation, either version 3 of the License, or
 * (at your option) any later version.
 *
 * Bowtie 2 is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with Bowtie 2.  If not, see <http://www.gnu.org/licenses/>.
 */

/**
 * aligner_sw_sse.cpp
 *
 * Versions of key alignment functions that use vector instructions to
 * accelerate dynamic programming.  Based chiefly on the striped Smith-Waterman
 * paper and implementation by Michael Farrar.  See:
 *
 * Farrar M. Striped Smith-Waterman speeds database searches six times over
 * other SIMD implementations. Bioinformatics. 2007 Jan 15;23(2):156-61.
 * http://sites.google.com/site/farrarmichael/smith-waterman
 *
 * While the paper describes an implementation of Smith-Waterman, we extend it
 * do end-to-end read alignment as well as local alignment.  The change
 * required for this is minor: we simply let vmax be the maximum element in the
 * score domain rather than the minimum.
 *
 * The vectorized dynamic programming implementation lacks some features that
 * make it hard to adapt to solving the entire dynamic-programming alignment
 * problem.  For instance:
 *
 * - It doesn't respect gap barriers on either end of the read
 * - It just gives a maximum; not enough information to backtrace without
 *   redoing some alignment
 * - It's a little difficult to handle st_ and en_, especially st_.
 * - The query profile mechanism makes handling of ambiguous reference bases a
 *   little tricky (16 cols in query profile lookup table instead of 5)
 *
 * Given the drawbacks, it is tempting to use SSE dynamic programming as a
 * filter rather than as an aligner per se.  Here are a few ideas for how it
 * can be extended to handle more of the alignment problem:
 *
 * - Save calculated scores to a big array as we go.  We return to this array
 *   to find and backtrace from good solutions.
 */

#include <limits>
#include "aligner_sw.h"

static const size_t NBYTES_PER_REG  = 16;
static const size_t NWORDS_PER_REG  = 8;
static const size_t NBITS_PER_WORD  = 16;
static const size_t NBYTES_PER_WORD = 2;

// In 16-bit end-to-end mode, we have the option of using signed saturated
// arithmetic.  Because we have signed arithmetic, there's no need to add/subtract
// bias when building an applying the query profile.  The lowest value we can
// use is 0x8000, and the greatest is 0x7fff.

typedef int16_t TCScore;

/**
 * Build query profile look up tables for the read.  The query profile look
 * up table is organized as a 1D array indexed by [i][j] where i is the
 * reference character in the current DP column (0=A, 1=C, etc), and j is
 * the segment of the query we're currently working on.
 */
void SwAligner::buildQueryProfileEnd2EndSseI16(bool fw) {
	bool& done = fw ? sseI16fwBuilt_ : sseI16rcBuilt_;
	if(done) {
		return;
	}
	done = true;
	const BTDnaString* rd = fw ? rdfw_ : rdrc_;
	const BTString* qu = fw ? qufw_ : qurc_;
    // daehwan - allows to align a portion of a read, not the whole
	// const size_t len = rd->length();
    const size_t len = dpRows();
	const size_t seglen = (len + (NWORDS_PER_REG-1)) / NWORDS_PER_REG;
	// How many __m128i's are needed
	size_t n128s =
		64 +                    // slack bytes, for alignment?
		(seglen * ALPHA_SIZE)   // query profile data
		* 2;                    // & gap barrier data
	assert_gt(n128s, 0);
	SSEData& d = fw ? sseI16fw_ : sseI16rc_;
	d.profbuf_.resizeNoCopy(n128s);
	assert(!d.profbuf_.empty());
	d.maxPen_      = d.maxBonus_ = 0;
	d.lastIter_    = d.lastWord_ = 0;
	d.qprofStride_ = d.gbarStride_ = 2;
	d.bias_ = 0; // no bias when words are signed
	// For each reference character A, C, G, T, N ...
	for(size_t refc = 0; refc < ALPHA_SIZE; refc++) {
		// For each segment ...
		for(size_t i = 0; i < seglen; i++) {
			size_t j = i;
			int16_t *qprofWords =
				reinterpret_cast<int16_t*>(d.profbuf_.ptr() + (refc * seglen * 2) + (i * 2));
			int16_t *gbarWords =
				reinterpret_cast<int16_t*>(d.profbuf_.ptr() + (refc * seglen * 2) + (i * 2) + 1);
			// For each sub-word (byte) ...
			for(size_t k = 0; k < NWORDS_PER_REG; k++) {
				int sc = 0;
				*gbarWords = 0;
				if(j < len) {
					int readc = (*rd)[j];
					int readq = (*qu)[j];
					sc = sc_->score(readc, (int)(1 << refc), readq - 33);
					size_t j_from_end = len - j - 1;
					if(j < (size_t)sc_->gapbar ||
					   j_from_end < (size_t)sc_->gapbar)
					{
						// Inside the gap barrier
						*gbarWords = 0x8000; // add this twice
					}
				}
				if(refc == 0 && j == len-1) {
					// Remember which 128-bit word and which smaller word has
					// the final row
					d.lastIter_ = i;
					d.lastWord_ = k;
				}
				if(sc < 0) {
					if((size_t)(-sc) > d.maxPen_) {
						d.maxPen_ = (size_t)(-sc);
					}
				} else {
					if((size_t)sc > d.maxBonus_) {
						d.maxBonus_ = (size_t)sc;
					}
				}
				*qprofWords = (int16_t)sc;
				gbarWords++;
				qprofWords++;
				j += seglen; // update offset into query
			}
		}
	}
}

#ifndef NDEBUG
/**
 * Return true iff the cell has sane E/F/H values w/r/t its predecessors.
 */
static bool cellOkEnd2EndI16(
	SSEData& d,
	size_t row,
	size_t col,
	int refc,
	int readc,
	int readq,
	const Scoring& sc)     // scoring scheme
{
	TCScore floorsc = 0x8000;
	TCScore ceilsc = MAX_I64;
	TAlScore offsetsc = -0x7fff;
	TAlScore sc_h_cur = (TAlScore)d.mat_.helt(row, col);
	TAlScore sc_e_cur = (TAlScore)d.mat_.eelt(row, col);
	TAlScore sc_f_cur = (TAlScore)d.mat_.felt(row, col);
	if(sc_h_cur > floorsc) {
		sc_h_cur += offsetsc;
	}
	if(sc_e_cur > floorsc) {
		sc_e_cur += offsetsc;
	}
	if(sc_f_cur > floorsc) {
		sc_f_cur += offsetsc;
	}
	bool gapsAllowed = true;
	size_t rowFromEnd = d.mat_.nrow() - row - 1;
	if(row < (size_t)sc.gapbar || rowFromEnd < (size_t)sc.gapbar) {
		gapsAllowed = false;
	}
	bool e_left_trans = false, h_left_trans = false;
	bool f_up_trans   = false, h_up_trans = false;
	bool h_diag_trans = false;
	if(gapsAllowed) {
		TAlScore sc_h_left = floorsc;
		TAlScore sc_e_left = floorsc;
		TAlScore sc_h_up   = floorsc;
		TAlScore sc_f_up   = floorsc;
		if(col > 0 && sc_e_cur > floorsc && sc_e_cur <= ceilsc) {
			sc_h_left = d.mat_.helt(row, col-1) + offsetsc;
			sc_e_left = d.mat_.eelt(row, col-1) + offsetsc;
			e_left_trans = (sc_e_left > floorsc && sc_e_cur == sc_e_left - sc.readGapExtend());
			h_left_trans = (sc_h_left > floorsc && sc_e_cur == sc_h_left - sc.readGapOpen());
			assert(e_left_trans || h_left_trans);
		}
		if(row > 0 && sc_f_cur > floorsc && sc_f_cur <= ceilsc) {
			sc_h_up = d.mat_.helt(row-1, col) + offsetsc;
			sc_f_up = d.mat_.felt(row-1, col) + offsetsc;
			f_up_trans = (sc_f_up > floorsc && sc_f_cur == sc_f_up - sc.refGapExtend());
			h_up_trans = (sc_h_up > floorsc && sc_f_cur == sc_h_up - sc.refGapOpen());
			assert(f_up_trans || h_up_trans);
		}
	} else {
		assert_geq(floorsc, sc_e_cur);
		assert_geq(floorsc, sc_f_cur);
	}
	if(col > 0 && row > 0 && sc_h_cur > floorsc && sc_h_cur <= ceilsc) {
		TAlScore sc_h_upleft = d.mat_.helt(row-1, col-1) + offsetsc;
		TAlScore sc_diag = sc.score(readc, (int)refc, readq - 33);
		h_diag_trans = sc_h_cur == sc_h_upleft + sc_diag;
	}
	assert(
		sc_h_cur <= floorsc ||
		e_left_trans ||
		h_left_trans ||
		f_up_trans   ||
		h_up_trans   ||
		h_diag_trans ||
		sc_h_cur > ceilsc ||
		row == 0 ||
		col == 0);
	return true;
}
#endif /*ndef NDEBUG*/

#ifdef NDEBUG

#define assert_all_eq0(x)
#define assert_all_gt(x, y)
#define assert_all_gt_lo(x)
#define assert_all_lt(x, y)
#define assert_all_lt_hi(x)

#else

#define assert_all_eq0(x) { \
	__m128i z = _mm_setzero_si128(); \
	__m128i tmp = _mm_setzero_si128(); \
	z = _mm_xor_si128(z, z); \
	tmp = _mm_cmpeq_epi16(x, z); \
	assert_eq(0xffff, _mm_movemask_epi8(tmp)); \
}

#define assert_all_gt(x, y) { \
	__m128i tmp = _mm_cmpgt_epi16(x, y); \
	assert_eq(0xffff, _mm_movemask_epi8(tmp)); \
}

#define assert_all_gt_lo(x) { \
	__m128i z = _mm_setzero_si128(); \
	__m128i tmp = _mm_setzero_si128(); \
	z = _mm_xor_si128(z, z); \
	tmp = _mm_cmpgt_epi16(x, z); \
	assert_eq(0xffff, _mm_movemask_epi8(tmp)); \
}

#define assert_all_lt(x, y) { \
	__m128i tmp = _mm_cmplt_epi16(x, y); \
	assert_eq(0xffff, _mm_movemask_epi8(tmp)); \
}

#define assert_all_leq(x, y) { \
	__m128i tmp = _mm_cmpgt_epi16(x, y); \
	assert_eq(0x0000, _mm_movemask_epi8(tmp)); \
}

#define assert_all_lt_hi(x) { \
	__m128i z = _mm_setzero_si128(); \
	__m128i tmp = _mm_setzero_si128(); \
	z = _mm_cmpeq_epi16(z, z); \
	z = _mm_srli_epi16(z, 1); \
	tmp = _mm_cmplt_epi16(x, z); \
	assert_eq(0xffff, _mm_movemask_epi8(tmp)); \
}
#endif

/**
 * Aligns by filling a dynamic programming matrix with the SSE-accelerated,
 * banded DP approach of Farrar.  As it goes, it determines which cells we
 * might backtrace from and tallies the best (highest-scoring) N backtrace
 * candidate cells per diagonal.  Also returns the alignment score of the best
 * alignment in the matrix.
 *
 * This routine does *not* maintain a matrix holding the entire matrix worth of
 * scores, nor does it maintain any other dense O(mn) data structure, as this
 * would quickly exhaust memory for queries longer than about 10,000 kb.
 * Instead, in the fill stage it maintains two columns worth of scores at a
 * time (current/previous, or right/left) - these take O(m) space.  When
 * finished with the current column, it determines which cells from the
 * previous column, if any, are candidates we might backtrace from to find a
 * full alignment.  A candidate cell has a score that rises above the threshold
 * and isn't improved upon by a match in the next column.  The best N
 * candidates per diagonal are stored in a O(m + n) data structure.
 */
TAlScore SwAligner::alignGatherEE16(int& flag, bool debug) {
	assert_leq(rdf_, rd_->length());
	assert_leq(rdf_, qu_->length());
	assert_lt(rfi_, rff_);
	assert_lt(rdi_, rdf_);
	assert_eq(rd_->length(), qu_->length());
	assert_geq(sc_->gapbar, 1);
	assert(repOk());
#ifndef NDEBUG
	for(size_t i = (size_t)rfi_; i < (size_t)rff_; i++) {
		assert_range(0, 16, (int)rf_[i]);
	}
#endif

	SSEData& d = fw_ ? sseI16fw_ : sseI16rc_;
	SSEMetrics& met = extend_ ? sseI16ExtendMet_ : sseI16MateMet_;
	if(!debug) met.dp++;
	buildQueryProfileEnd2EndSseI16(fw_);
	assert(!d.profbuf_.empty());

	assert_eq(0, d.maxBonus_);
	size_t iter =
		(dpRows() + (NWORDS_PER_REG-1)) / NWORDS_PER_REG; // iter = segLen
	
	// Now set up the score vectors.  We just need two columns worth, which
	// we'll call "left" and "right".
	d.vecbuf_.resize(4 * 2 * iter);
	d.vecbuf_.zero();
	__m128i *vbuf_l = d.vecbuf_.ptr();
	__m128i *vbuf_r = d.vecbuf_.ptr() + (4 * iter);
	
	// This is the data structure that holds candidate cells per diagonal.
	const size_t ndiags = rff_ - rfi_ + dpRows() - 1;
	if(!debug) {
		btdiag_.init(ndiags, 2);
	}

	// Data structure that holds checkpointed anti-diagonals
	TAlScore perfectScore = sc_->perfectScore(dpRows());
	bool checkpoint = true;
	bool cpdebug = false;
#ifndef NDEBUG
	cpdebug = dpRows() < 1000;
#endif
	cper_.init(
		dpRows(),      // # rows
		rff_ - rfi_,   // # columns
		cperPerPow2_,  // checkpoint every 1 << perpow2 diags (& next)
		perfectScore,  // perfect score (for sanity checks)
		false,         // matrix cells have 8-bit scores?
		cperTri_,      // triangular mini-fills?
		false,         // alignment is local?
		cpdebug);      // save all cells for debugging?
		
	// Many thanks to Michael Farrar for releasing his striped Smith-Waterman
	// implementation:
	//
	//  http://sites.google.com/site/farrarmichael/smith-waterman
	//
	// Much of the implmentation below is adapted from Michael's code.

	// Set all elts to reference gap open penalty
	__m128i rfgapo   = _mm_setzero_si128();
	__m128i rfgape   = _mm_setzero_si128();
	__m128i rdgapo   = _mm_setzero_si128();
	__m128i rdgape   = _mm_setzero_si128();
	__m128i vlo      = _mm_setzero_si128();
	__m128i vhi      = _mm_setzero_si128();
	__m128i vhilsw   = _mm_setzero_si128();
	__m128i vlolsw   = _mm_setzero_si128();
	__m128i ve       = _mm_setzero_si128();
	__m128i vf       = _mm_setzero_si128();
	__m128i vh       = _mm_setzero_si128();
	__m128i vhd      = _mm_setzero_si128();
	__m128i vhdtmp   = _mm_setzero_si128();
	__m128i vtmp     = _mm_setzero_si128();

	assert_gt(sc_->refGapOpen(), 0);
	assert_leq(sc_->refGapOpen(), MAX_I16);
	rfgapo = _mm_insert_epi16(rfgapo, sc_->refGapOpen(), 0);
	rfgapo = _mm_shufflelo_epi16(rfgapo, 0);
	rfgapo = _mm_shuffle_epi32(rfgapo, 0);
	
	// Set all elts to reference gap extension penalty
	assert_gt(sc_->refGapExtend(), 0);
	assert_leq(sc_->refGapExtend(), MAX_I16);
	assert_leq(sc_->refGapExtend(), sc_->refGapOpen());
	rfgape = _mm_insert_epi16(rfgape, sc_->refGapExtend(), 0);
	rfgape = _mm_shufflelo_epi16(rfgape, 0);
	rfgape = _mm_shuffle_epi32(rfgape, 0);

	// Set all elts to read gap open penalty
	assert_gt(sc_->readGapOpen(), 0);
	assert_leq(sc_->readGapOpen(), MAX_I16);
	rdgapo = _mm_insert_epi16(rdgapo, sc_->readGapOpen(), 0);
	rdgapo = _mm_shufflelo_epi16(rdgapo, 0);
	rdgapo = _mm_shuffle_epi32(rdgapo, 0);
	
	// Set all elts to read gap extension penalty
	assert_gt(sc_->readGapExtend(), 0);
	assert_leq(sc_->readGapExtend(), MAX_I16);
	assert_leq(sc_->readGapExtend(), sc_->readGapOpen());
	rdgape = _mm_insert_epi16(rdgape, sc_->readGapExtend(), 0);
	rdgape = _mm_shufflelo_epi16(rdgape, 0);
	rdgape = _mm_shuffle_epi32(rdgape, 0);

	// Set all elts to 0x8000 (min value for signed 16-bit)
	vlo = _mm_cmpeq_epi16(vlo, vlo);             // all elts = 0xffff
	vlo = _mm_slli_epi16(vlo, NBITS_PER_WORD-1); // all elts = 0x8000
	
	// Set all elts to 0x7fff (max value for signed 16-bit)
	vhi = _mm_cmpeq_epi16(vhi, vhi);             // all elts = 0xffff
	vhi = _mm_srli_epi16(vhi, 1);                // all elts = 0x7fff
	
	// vlolsw: topmost (least sig) word set to 0x8000, all other words=0
	vlolsw = _mm_shuffle_epi32(vlo, 0);
	vlolsw = _mm_srli_si128(vlolsw, NBYTES_PER_REG - NBYTES_PER_WORD);
	
	// vhilsw: topmost (least sig) word set to 0x7fff, all other words=0
	vhilsw = _mm_shuffle_epi32(vhi, 0);
	vhilsw = _mm_srli_si128(vhilsw, NBYTES_PER_REG - NBYTES_PER_WORD);
	
	// Points to a long vector of __m128i where each element is a block of
	// contiguous cells in the E, F or H matrix.  If the index % 3 == 0, then
	// the block of cells is from the E matrix.  If index % 3 == 1, they're
	// from the F matrix.  If index % 3 == 2, then they're from the H matrix.
	// Blocks of cells are organized in the same interleaved manner as they are
	// calculated by the Farrar algorithm.
	const __m128i *pvScore; // points into the query profile

	const size_t colstride = ROWSTRIDE_2COL * iter;
	
	// Initialize the H and E vectors in the first matrix column
	__m128i *pvELeft = vbuf_l + 0; __m128i *pvERight = vbuf_r + 0;
	/* __m128i *pvFLeft = vbuf_l + 1; */ __m128i *pvFRight = vbuf_r + 1;
	__m128i *pvHLeft = vbuf_l + 2; __m128i *pvHRight = vbuf_r + 2;
	
	// Maximum score in final row
	bool found = false;
	TCScore lrmax = MIN_I16;
	
	for(size_t i = 0; i < iter; i++) {
		_mm_store_si128(pvERight, vlo); pvERight += ROWSTRIDE_2COL;
		// Could initialize Hs to high or low.  If high, cells in the lower
		// triangle will have somewhat more legitiate scores, but still won't
		// be exhaustively scored.
		_mm_store_si128(pvHRight, vlo); pvHRight += ROWSTRIDE_2COL;
	}
	
	assert_gt(sc_->gapbar, 0);
	size_t nfixup = 0;

	// Fill in the table as usual but instead of using the same gap-penalty
	// vector for each iteration of the inner loop, load words out of a
	// pre-calculated gap vector parallel to the query profile.  The pre-
	// calculated gap vectors enforce the gap barrier constraint by making it
	// infinitely costly to introduce a gap in barrier rows.
	//
	// AND use a separate loop to fill in the first row of the table, enforcing
	// the st_ constraints in the process.  This is awkward because it
	// separates the processing of the first row from the others and might make
	// it difficult to use the first-row results in the next row, but it might
	// be the simplest and least disruptive way to deal with the st_ constraint.
	
	for(size_t i = (size_t)rfi_; i < (size_t)rff_; i++) {
		// Swap left and right; vbuf_l is the vector on the left, which we
		// generally load from, and vbuf_r is the vector on the right, which we
		// generally store to.
		swap(vbuf_l, vbuf_r);
		pvELeft = vbuf_l + 0; pvERight = vbuf_r + 0;
		/* pvFLeft = vbuf_l + 1; */ pvFRight = vbuf_r + 1;
		pvHLeft = vbuf_l + 2; pvHRight = vbuf_r + 2;
		
		// Fetch the appropriate query profile.  Note that elements of rf_ must
		// be numbers, not masks.
		const int refc = (int)rf_[i];
		
		// Fetch the appropriate query profile
		size_t off = (size_t)firsts5[refc] * iter * 2;
		pvScore = d.profbuf_.ptr() + off; // even elts = query profile, odd = gap barrier
		
		// Set all cells to low value
		vf = _mm_cmpeq_epi16(vf, vf);
		vf = _mm_slli_epi16(vf, NBITS_PER_WORD-1);
		vf = _mm_or_si128(vf, vlolsw);
		
		// Load H vector from the final row of the previous column
		vh = _mm_load_si128(pvHLeft + colstride - ROWSTRIDE_2COL);
		// Shift 2 bytes down so that topmost (least sig) cell gets 0
		vh = _mm_slli_si128(vh, NBYTES_PER_WORD);
		// Fill topmost (least sig) cell with high value
		vh = _mm_or_si128(vh, vhilsw);
		
		// For each character in the reference text:
		size_t j;
		for(j = 0; j < iter; j++) {
			// Load cells from E, calculated previously
			ve = _mm_load_si128(pvELeft);
			vhd = _mm_load_si128(pvHLeft);
			assert_all_lt(ve, vhi);
			pvELeft += ROWSTRIDE_2COL;
			
			// Store cells in F, calculated previously
			vf = _mm_adds_epi16(vf, pvScore[1]); // veto some ref gap extensions
			vf = _mm_adds_epi16(vf, pvScore[1]); // veto some ref gap extensions
			_mm_store_si128(pvFRight, vf);
			pvFRight += ROWSTRIDE_2COL;
			
			// Factor in query profile (matches and mismatches)
			vh = _mm_adds_epi16(vh, pvScore[0]);
			
			// Update H, factoring in E and F
			vh = _mm_max_epi16(vh, vf);
			
			// Update vE value
			vhdtmp = vhd;
			vhd = _mm_subs_epi16(vhd, rdgapo);
			vhd = _mm_adds_epi16(vhd, pvScore[1]); // veto some read gap opens
			vhd = _mm_adds_epi16(vhd, pvScore[1]); // veto some read gap opens
			ve = _mm_subs_epi16(ve, rdgape);
			ve = _mm_max_epi16(ve, vhd);
			vh = _mm_max_epi16(vh, ve);

			// Save the new vH values
			_mm_store_si128(pvHRight, vh);
			pvHRight += ROWSTRIDE_2COL;
			vtmp = vh;
			assert_all_lt(ve, vhi);
			
			// Load the next h value
			vh = vhdtmp;
			pvHLeft += ROWSTRIDE_2COL;

			// Save E values
			_mm_store_si128(pvERight, ve);
			pvERight += ROWSTRIDE_2COL;
			
			// Update vf value
			vtmp = _mm_subs_epi16(vtmp, rfgapo);
			vf = _mm_subs_epi16(vf, rfgape);
			assert_all_lt(vf, vhi);
			vf = _mm_max_epi16(vf, vtmp);
			
			pvScore += 2; // move on to next query profile / gap veto
		}
		// pvHStore, pvELoad, pvEStore have all rolled over to the next column
		pvFRight -= colstride; // reset to start of column
		vtmp = _mm_load_si128(pvFRight);
		
		pvHRight -= colstride; // reset to start of column
		vh = _mm_load_si128(pvHRight);
		
		pvScore = d.profbuf_.ptr() + off + 1; // reset veto vector
		
		// vf from last row gets shifted down by one to overlay the first row
		// rfgape has already been subtracted from it.
		vf = _mm_slli_si128(vf, NBYTES_PER_WORD);
		vf = _mm_or_si128(vf, vlolsw);
		
		vf = _mm_adds_epi16(vf, *pvScore); // veto some ref gap extensions
		vf = _mm_adds_epi16(vf, *pvScore); // veto some ref gap extensions
		vf = _mm_max_epi16(vtmp, vf);
		vtmp = _mm_cmpgt_epi16(vf, vtmp);
		int cmp = _mm_movemask_epi8(vtmp);
		
		// If any element of vtmp is greater than H - gap-open...
		j = 0;
		while(cmp != 0x0000) {
			// Store this vf
			_mm_store_si128(pvFRight, vf);
			pvFRight += ROWSTRIDE_2COL;
			
			// Update vh w/r/t new vf
			vh = _mm_max_epi16(vh, vf);
			
			// Save vH values
			_mm_store_si128(pvHRight, vh);
			pvHRight += ROWSTRIDE_2COL;
			
			pvScore += 2;
			
			assert_lt(j, iter);
			if(++j == iter) {
				pvFRight -= colstride;
				vtmp = _mm_load_si128(pvFRight);   // load next vf ASAP
				pvHRight -= colstride;
				vh = _mm_load_si128(pvHRight);     // load next vh ASAP
				pvScore = d.profbuf_.ptr() + off + 1;
				j = 0;
				vf = _mm_slli_si128(vf, NBYTES_PER_WORD);
				vf = _mm_or_si128(vf, vlolsw);
			} else {
				vtmp = _mm_load_si128(pvFRight);   // load next vf ASAP
				vh = _mm_load_si128(pvHRight);     // load next vh ASAP
			}
			
			// Update F with another gap extension
			vf = _mm_subs_epi16(vf, rfgape);
			vf = _mm_adds_epi16(vf, *pvScore); // veto some ref gap extensions
			vf = _mm_adds_epi16(vf, *pvScore); // veto some ref gap extensions
			vf = _mm_max_epi16(vtmp, vf);
			vtmp = _mm_cmpgt_epi16(vf, vtmp);
			cmp = _mm_movemask_epi8(vtmp);
			nfixup++;
		}

		
		// Check in the last row for the maximum so far
		__m128i *vtmp = vbuf_r + 2 /* H */ + (d.lastIter_ * ROWSTRIDE_2COL);
		// Note: we may not want to extract from the final row
		TCScore lr = ((TCScore*)(vtmp))[d.lastWord_];
		found = true;
		if(lr > lrmax) {
			lrmax = lr;
		}
		
		// Now we'd like to know whether the bottommost element of the right
		// column is a candidate we might backtrace from.  First question is:
		// did it exceed the minimum score threshold?
		TAlScore score = (TAlScore)(lr - 0x7fff);
		if(lr == MIN_I16) {
			score = MIN_I64;
		}
		if(!debug && score >= minsc_) {
			DpBtCandidate cand(dpRows() - 1, i - rfi_, score);
			btdiag_.add(i - rfi_, cand);
		}

		// Save some elements to checkpoints
		if(checkpoint) {
			
			__m128i *pvE = vbuf_r + 0;
			__m128i *pvF = vbuf_r + 1;
			__m128i *pvH = vbuf_r + 2;
			size_t coli = i - rfi_;
			if(coli < cper_.locol_) cper_.locol_ = coli;
			if(coli > cper_.hicol_) cper_.hicol_ = coli;
			
			if(cperTri_) {
				size_t rc_mod = coli & cper_.lomask_;
				assert_lt(rc_mod, cper_.per_);
				int64_t row = -(int64_t)rc_mod-1;
				int64_t row_mod = row;
				int64_t row_div = 0;
				size_t idx = coli >> cper_.perpow2_;
				size_t idxrow = idx * cper_.nrow_;
				assert_eq(4, ROWSTRIDE_2COL);
				bool done = false;
				while(true) {
					row += (cper_.per_ - 2);
					row_mod += (cper_.per_ - 2);
					for(size_t j = 0; j < 2; j++) {
						row++;
						row_mod++;
						if(row >= 0 && (size_t)row < cper_.nrow_) {
							// Update row divided by iter_ and mod iter_
							while(row_mod >= (int64_t)iter) {
								row_mod -= (int64_t)iter;
								row_div++;
							}
							size_t delt = idxrow + row;
							size_t vecoff = (row_mod << 5) + row_div;
							assert_lt(row_div, 8);
							int16_t h_sc = ((int16_t*)pvH)[vecoff];
							int16_t e_sc = ((int16_t*)pvE)[vecoff];
							int16_t f_sc = ((int16_t*)pvF)[vecoff];
							if(h_sc != MIN_I16) h_sc -= 0x7fff;
							if(e_sc != MIN_I16) e_sc -= 0x7fff;
							if(f_sc != MIN_I16) f_sc -= 0x7fff;
							assert_leq(h_sc, cper_.perf_);
							assert_leq(e_sc, cper_.perf_);
							assert_leq(f_sc, cper_.perf_);
							CpQuad *qdiags = ((j == 0) ? cper_.qdiag1s_.ptr() : cper_.qdiag2s_.ptr());
							qdiags[delt].sc[0] = h_sc;
							qdiags[delt].sc[1] = e_sc;
							qdiags[delt].sc[2] = f_sc;
						} // if(row >= 0 && row < nrow_)
						else if(row >= 0 && (size_t)row >= cper_.nrow_) {
							done = true;
							break;
						}
					} // end of loop over anti-diags
					if(done) {
						break;
					}
					idx++;
					idxrow += cper_.nrow_;
				}
			} else {
				// If this is the first column, take this opportunity to
				// pre-calculate the coordinates of the elements we're going to
				// checkpoint.
				if(coli == 0) {
					size_t cpi    = cper_.per_-1;
					size_t cpimod = cper_.per_-1;
					size_t cpidiv = 0;
					cper_.commitMap_.clear();
					while(cpi < cper_.nrow_) {
						while(cpimod >= iter) {
							cpimod -= iter;
							cpidiv++;
						}
						size_t vecoff = (cpimod << 5) + cpidiv;
						cper_.commitMap_.push_back(vecoff);
						cpi += cper_.per_;
						cpimod += cper_.per_;
					}
				}
				// Save all the rows
				size_t rowoff = 0;
				size_t sz = cper_.commitMap_.size();
				for(size_t i = 0; i < sz; i++, rowoff += cper_.ncol_) {
					size_t vecoff = cper_.commitMap_[i];
					int16_t h_sc = ((int16_t*)pvH)[vecoff];
					int16_t e_sc = ((int16_t*)pvE)[vecoff];
					int16_t f_sc = ((int16_t*)pvF)[vecoff];
					if(h_sc != MIN_I16) h_sc -= 0x7fff;
					if(e_sc != MIN_I16) e_sc -= 0x7fff;
					if(f_sc != MIN_I16) f_sc -= 0x7fff;
					assert_leq(h_sc, cper_.perf_);
					assert_leq(e_sc, cper_.perf_);
					assert_leq(f_sc, cper_.perf_);
					CpQuad& dst = cper_.qrows_[rowoff + coli];
					dst.sc[0] = h_sc;
					dst.sc[1] = e_sc;
					dst.sc[2] = f_sc;
				}
				// Is this a column we'd like to checkpoint?
				if((coli & cper_.lomask_) == cper_.lomask_) {
					// Save the column using memcpys
					assert_gt(coli, 0);
					size_t wordspercol = cper_.niter_ * ROWSTRIDE_2COL;
					size_t coloff = (coli >> cper_.perpow2_) * wordspercol;
					__m128i *dst = cper_.qcols_.ptr() + coloff;
					memcpy(dst, vbuf_r, sizeof(__m128i) * wordspercol);
				}
			}
			if(cper_.debug_) {
				// Save the column using memcpys
				size_t wordspercol = cper_.niter_ * ROWSTRIDE_2COL;
				size_t coloff = coli * wordspercol;
				__m128i *dst = cper_.qcolsD_.ptr() + coloff;
				memcpy(dst, vbuf_r, sizeof(__m128i) * wordspercol);
			}
		}
	}
	
	// Update metrics
	if(!debug) {
		size_t ninner = (rff_ - rfi_) * iter;
		met.col   += (rff_ - rfi_);             // DP columns
		met.cell  += (ninner * NWORDS_PER_REG); // DP cells
		met.inner += ninner;                    // DP inner loop iters
		met.fixup += nfixup;                    // DP fixup loop iters
	}

	flag = 0;

	// Did we find a solution?
	TAlScore score = MIN_I64;
	if(!found) {
		flag = -1; // no
		if(!debug) met.dpfail++;
		return MIN_I64;
	} else {
		score = (TAlScore)(lrmax - 0x7fff);
		if(score < minsc_) {
			flag = -1; // no
			if(!debug) met.dpfail++;
			return score;
		}
	}
	
	// Could we have saturated?
	if(lrmax == MIN_I16) {
		flag = -2; // yes
		if(!debug) met.dpsat++;
		return MIN_I64;
	}
	
	// Now take all the backtrace candidates in the btdaig_ structure and
	// dump them into the btncand_ array.  They'll be sorted later.
	if(!debug) {
		btdiag_.dump(btncand_);
		assert(!btncand_.empty());
	}
	
	// Return largest score
	if(!debug) met.dpsucc++;
	return score;
}

/**
 * Solve the current alignment problem using SSE instructions that operate on 8
 * signed 16-bit values packed into a single 128-bit register.
 */
TAlScore SwAligner::alignNucleotidesEnd2EndSseI16(int& flag, bool debug) {
	assert_leq(rdf_, rd_->length());
	assert_leq(rdf_, qu_->length());
	assert_lt(rfi_, rff_);
	assert_lt(rdi_, rdf_);
	assert_eq(rd_->length(), qu_->length());
	assert_geq(sc_->gapbar, 1);
	assert(repOk());
#ifndef NDEBUG
	for(size_t i = (size_t)rfi_; i < (size_t)rff_; i++) {
		assert_range(0, 16, (int)rf_[i]);
	}
#endif

	SSEData& d = fw_ ? sseI16fw_ : sseI16rc_;
	SSEMetrics& met = extend_ ? sseI16ExtendMet_ : sseI16MateMet_;
	if(!debug) met.dp++;
	buildQueryProfileEnd2EndSseI16(fw_);
	assert(!d.profbuf_.empty());

	assert_eq(0, d.maxBonus_);
	size_t iter =
		(dpRows() + (NWORDS_PER_REG-1)) / NWORDS_PER_REG; // iter = segLen

	// Many thanks to Michael Farrar for releasing his striped Smith-Waterman
	// implementation:
	//
	//  http://sites.google.com/site/farrarmichael/smith-waterman
	//
	// Much of the implmentation below is adapted from Michael's code.

	// Set all elts to reference gap open penalty
	__m128i rfgapo   = _mm_setzero_si128();
	__m128i rfgape   = _mm_setzero_si128();
	__m128i rdgapo   = _mm_setzero_si128();
	__m128i rdgape   = _mm_setzero_si128();
	__m128i vlo      = _mm_setzero_si128();
	__m128i vhi      = _mm_setzero_si128();
	__m128i vhilsw   = _mm_setzero_si128();
	__m128i vlolsw   = _mm_setzero_si128();
	__m128i ve       = _mm_setzero_si128();
	__m128i vf       = _mm_setzero_si128();
	__m128i vh       = _mm_setzero_si128();
#if 0
	__m128i vhd      = _mm_setzero_si128();
	__m128i vhdtmp   = _mm_setzero_si128();
#endif
	__m128i vtmp     = _mm_setzero_si128();

	assert_gt(sc_->refGapOpen(), 0);
	assert_leq(sc_->refGapOpen(), MAX_I16);
	rfgapo = _mm_insert_epi16(rfgapo, sc_->refGapOpen(), 0);
	rfgapo = _mm_shufflelo_epi16(rfgapo, 0);
	rfgapo = _mm_shuffle_epi32(rfgapo, 0);
	
	// Set all elts to reference gap extension penalty
	assert_gt(sc_->refGapExtend(), 0);
	assert_leq(sc_->refGapExtend(), MAX_I16);
	assert_leq(sc_->refGapExtend(), sc_->refGapOpen());
	rfgape = _mm_insert_epi16(rfgape, sc_->refGapExtend(), 0);
	rfgape = _mm_shufflelo_epi16(rfgape, 0);
	rfgape = _mm_shuffle_epi32(rfgape, 0);

	// Set all elts to read gap open penalty
	assert_gt(sc_->readGapOpen(), 0);
	assert_leq(sc_->readGapOpen(), MAX_I16);
	rdgapo = _mm_insert_epi16(rdgapo, sc_->readGapOpen(), 0);
	rdgapo = _mm_shufflelo_epi16(rdgapo, 0);
	rdgapo = _mm_shuffle_epi32(rdgapo, 0);
	
	// Set all elts to read gap extension penalty
	assert_gt(sc_->readGapExtend(), 0);
	assert_leq(sc_->readGapExtend(), MAX_I16);
	assert_leq(sc_->readGapExtend(), sc_->readGapOpen());
	rdgape = _mm_insert_epi16(rdgape, sc_->readGapExtend(), 0);
	rdgape = _mm_shufflelo_epi16(rdgape, 0);
	rdgape = _mm_shuffle_epi32(rdgape, 0);

	// Set all elts to 0x8000 (min value for signed 16-bit)
	vlo = _mm_cmpeq_epi16(vlo, vlo);             // all elts = 0xffff
	vlo = _mm_slli_epi16(vlo, NBITS_PER_WORD-1); // all elts = 0x8000
	
	// Set all elts to 0x7fff (max value for signed 16-bit)
	vhi = _mm_cmpeq_epi16(vhi, vhi);             // all elts = 0xffff
	vhi = _mm_srli_epi16(vhi, 1);                // all elts = 0x7fff
	
	// vlolsw: topmost (least sig) word set to 0x8000, all other words=0
	vlolsw = _mm_shuffle_epi32(vlo, 0);
	vlolsw = _mm_srli_si128(vlolsw, NBYTES_PER_REG - NBYTES_PER_WORD);
	
	// vhilsw: topmost (least sig) word set to 0x7fff, all other words=0
	vhilsw = _mm_shuffle_epi32(vhi, 0);
	vhilsw = _mm_srli_si128(vhilsw, NBYTES_PER_REG - NBYTES_PER_WORD);
	
	// Points to a long vector of __m128i where each element is a block of
	// contiguous cells in the E, F or H matrix.  If the index % 3 == 0, then
	// the block of cells is from the E matrix.  If index % 3 == 1, they're
	// from the F matrix.  If index % 3 == 2, then they're from the H matrix.
	// Blocks of cells are organized in the same interleaved manner as they are
	// calculated by the Farrar algorithm.
	const __m128i *pvScore; // points into the query profile

	d.mat_.init(dpRows(), rff_ - rfi_, NWORDS_PER_REG);
	const size_t colstride = d.mat_.colstride();
	assert_eq(ROWSTRIDE, colstride / iter);
	
	// Initialize the H and E vectors in the first matrix column
	__m128i *pvHTmp = d.mat_.tmpvec(0, 0);
	__m128i *pvETmp = d.mat_.evec(0, 0);
	
	// Maximum score in final row
	bool found = false;
	TCScore lrmax = MIN_I16;
	
	for(size_t i = 0; i < iter; i++) {
		_mm_store_si128(pvETmp, vlo);
		// Could initialize Hs to high or low.  If high, cells in the lower
		// triangle will have somewhat more legitiate scores, but still won't
		// be exhaustively scored.
		_mm_store_si128(pvHTmp, vlo);
		pvETmp += ROWSTRIDE;
		pvHTmp += ROWSTRIDE;
	}
	// These are swapped just before the innermost loop
	__m128i *pvHStore = d.mat_.hvec(0, 0);
	__m128i *pvHLoad  = d.mat_.tmpvec(0, 0);
	__m128i *pvELoad  = d.mat_.evec(0, 0);
	__m128i *pvEStore = d.mat_.evecUnsafe(0, 1);
	__m128i *pvFStore = d.mat_.fvec(0, 0);
	__m128i *pvFTmp   = NULL;
	
	assert_gt(sc_->gapbar, 0);
	size_t nfixup = 0;
	
	// Fill in the table as usual but instead of using the same gap-penalty
	// vector for each iteration of the inner loop, load words out of a
	// pre-calculated gap vector parallel to the query profile.  The pre-
	// calculated gap vectors enforce the gap barrier constraint by making it
	// infinitely costly to introduce a gap in barrier rows.
	//
	// AND use a separate loop to fill in the first row of the table, enforcing
	// the st_ constraints in the process.  This is awkward because it
	// separates the processing of the first row from the others and might make
	// it difficult to use the first-row results in the next row, but it might
	// be the simplest and least disruptive way to deal with the st_ constraint.
	
	colstop_ = rff_ - 1;
	lastsolcol_ = 0;
	
	for(size_t i = (size_t)rfi_; i < (size_t)rff_; i++) {
		assert(pvFStore == d.mat_.fvec(0, i - rfi_));
		assert(pvHStore == d.mat_.hvec(0, i - rfi_));
		
		// Fetch the appropriate query profile.  Note that elements of rf_ must
		// be numbers, not masks.
		const int refc = (int)rf_[i];
		size_t off = (size_t)firsts5[refc] * iter * 2;
		pvScore = d.profbuf_.ptr() + off; // even elts = query profile, odd = gap barrier
		
		// Set all cells to low value
		vf = _mm_cmpeq_epi16(vf, vf);
		vf = _mm_slli_epi16(vf, NBITS_PER_WORD-1);
		vf = _mm_or_si128(vf, vlolsw);
		
		// Load H vector from the final row of the previous column
		vh = _mm_load_si128(pvHLoad + colstride - ROWSTRIDE);
		// Shift 2 bytes down so that topmost (least sig) cell gets 0
		vh = _mm_slli_si128(vh, NBYTES_PER_WORD);
		// Fill topmost (least sig) cell with high value
		vh = _mm_or_si128(vh, vhilsw);
		
		// For each character in the reference text:
		size_t j;
		for(j = 0; j < iter; j++) {
			// Load cells from E, calculated previously
			ve = _mm_load_si128(pvELoad);
#if 0
			vhd = _mm_load_si128(pvHLoad);
#endif
			assert_all_lt(ve, vhi);
			pvELoad += ROWSTRIDE;
			
			// Store cells in F, calculated previously
			vf = _mm_adds_epi16(vf, pvScore[1]); // veto some ref gap extensions
			vf = _mm_adds_epi16(vf, pvScore[1]); // veto some ref gap extensions
			_mm_store_si128(pvFStore, vf);
			pvFStore += ROWSTRIDE;
			
			// Factor in query profile (matches and mismatches)
			vh = _mm_adds_epi16(vh, pvScore[0]);
			
			// Update H, factoring in E and F
			vh = _mm_max_epi16(vh, ve);
			vh = _mm_max_epi16(vh, vf);
			
			// Save the new vH values
			_mm_store_si128(pvHStore, vh);
			pvHStore += ROWSTRIDE;
			
			// Update vE value
			vtmp = vh;
#if 0
			vhdtmp = vhd;
			vhd = _mm_subs_epi16(vhd, rdgapo);
			vhd = _mm_adds_epi16(vhd, pvScore[1]); // veto some read gap opens
			vhd = _mm_adds_epi16(vhd, pvScore[1]); // veto some read gap opens
			ve = _mm_subs_epi16(ve, rdgape);
			ve = _mm_max_epi16(ve, vhd);
#else
			vh = _mm_subs_epi16(vh, rdgapo);
			vh = _mm_adds_epi16(vh, pvScore[1]); // veto some read gap opens
			vh = _mm_adds_epi16(vh, pvScore[1]); // veto some read gap opens
			ve = _mm_subs_epi16(ve, rdgape);
			ve = _mm_max_epi16(ve, vh);
#endif
			assert_all_lt(ve, vhi);
			
			// Load the next h value
#if 0
			vh = vhdtmp;
#else
			vh = _mm_load_si128(pvHLoad);
#endif
			pvHLoad += ROWSTRIDE;
			
			// Save E values
			_mm_store_si128(pvEStore, ve);
			pvEStore += ROWSTRIDE;
			
			// Update vf value
			vtmp = _mm_subs_epi16(vtmp, rfgapo);
			vf = _mm_subs_epi16(vf, rfgape);
			assert_all_lt(vf, vhi);
			vf = _mm_max_epi16(vf, vtmp);
			
			pvScore += 2; // move on to next query profile / gap veto
		}
		// pvHStore, pvELoad, pvEStore have all rolled over to the next column
		pvFTmp = pvFStore;
		pvFStore -= colstride; // reset to start of column
		vtmp = _mm_load_si128(pvFStore);
		
		pvHStore -= colstride; // reset to start of column
		vh = _mm_load_si128(pvHStore);
		
#if 0
#else
		pvEStore -= colstride; // reset to start of column
		ve = _mm_load_si128(pvEStore);
#endif
		
		pvHLoad = pvHStore;    // new pvHLoad = pvHStore
		pvScore = d.profbuf_.ptr() + off + 1; // reset veto vector
		
		// vf from last row gets shifted down by one to overlay the first row
		// rfgape has already been subtracted from it.
		vf = _mm_slli_si128(vf, NBYTES_PER_WORD);
		vf = _mm_or_si128(vf, vlolsw);
		
		vf = _mm_adds_epi16(vf, *pvScore); // veto some ref gap extensions
		vf = _mm_adds_epi16(vf, *pvScore); // veto some ref gap extensions
		vf = _mm_max_epi16(vtmp, vf);
		vtmp = _mm_cmpgt_epi16(vf, vtmp);
		int cmp = _mm_movemask_epi8(vtmp);
		
		// If any element of vtmp is greater than H - gap-open...
		j = 0;
		while(cmp != 0x0000) {
			// Store this vf
			_mm_store_si128(pvFStore, vf);
			pvFStore += ROWSTRIDE;
			
			// Update vh w/r/t new vf
			vh = _mm_max_epi16(vh, vf);
			
			// Save vH values
			_mm_store_si128(pvHStore, vh);
			pvHStore += ROWSTRIDE;
			
			// Update E in case it can be improved using our new vh
#if 0
#else
			vh = _mm_subs_epi16(vh, rdgapo);
			vh = _mm_adds_epi16(vh, *pvScore); // veto some read gap opens
			vh = _mm_adds_epi16(vh, *pvScore); // veto some read gap opens
			ve = _mm_max_epi16(ve, vh);
			_mm_store_si128(pvEStore, ve);
			pvEStore += ROWSTRIDE;
#endif
			pvScore += 2;
			
			assert_lt(j, iter);
			if(++j == iter) {
				pvFStore -= colstride;
				vtmp = _mm_load_si128(pvFStore);   // load next vf ASAP
				pvHStore -= colstride;
				vh = _mm_load_si128(pvHStore);     // load next vh ASAP
#if 0
#else
				pvEStore -= colstride;
				ve = _mm_load_si128(pvEStore);     // load next ve ASAP
#endif
				pvScore = d.profbuf_.ptr() + off + 1;
				j = 0;
				vf = _mm_slli_si128(vf, NBYTES_PER_WORD);
				vf = _mm_or_si128(vf, vlolsw);
			} else {
				vtmp = _mm_load_si128(pvFStore);   // load next vf ASAP
				vh = _mm_load_si128(pvHStore);     // load next vh ASAP
#if 0
#else
				ve = _mm_load_si128(pvEStore);     // load next vh ASAP
#endif
			}
			
			// Update F with another gap extension
			vf = _mm_subs_epi16(vf, rfgape);
			vf = _mm_adds_epi16(vf, *pvScore); // veto some ref gap extensions
			vf = _mm_adds_epi16(vf, *pvScore); // veto some ref gap extensions
			vf = _mm_max_epi16(vtmp, vf);
			vtmp = _mm_cmpgt_epi16(vf, vtmp);
			cmp = _mm_movemask_epi8(vtmp);
			nfixup++;
		}

#ifndef NDEBUG
		if((rand() & 15) == 0) {
			// This is a work-intensive sanity check; each time we finish filling
			// a column, we check that each H, E, and F is sensible.
			for(size_t k = 0; k < dpRows(); k++) {
				assert(cellOkEnd2EndI16(
					d,
					k,                   // row
					i - rfi_,            // col
					refc,                // reference mask
					(int)(*rd_)[rdi_+k], // read char
					(int)(*qu_)[rdi_+k], // read quality
					*sc_));              // scoring scheme
			}
		}
#endif
		
		__m128i *vtmp = d.mat_.hvec(d.lastIter_, i-rfi_);
		// Note: we may not want to extract from the final row
		TCScore lr = ((TCScore*)(vtmp))[d.lastWord_];
		found = true;
		if(lr > lrmax) {
			lrmax = lr;
		}

		// pvELoad and pvHLoad are already where they need to be
		
		// Adjust the load and store vectors here.  
		pvHStore = pvHLoad + colstride;
		pvEStore = pvELoad + colstride;
		pvFStore = pvFTmp;
	}
	
	// Update metrics
	if(!debug) {
		size_t ninner = (rff_ - rfi_) * iter;
		met.col   += (rff_ - rfi_);             // DP columns
		met.cell  += (ninner * NWORDS_PER_REG); // DP cells
		met.inner += ninner;                    // DP inner loop iters
		met.fixup += nfixup;                    // DP fixup loop iters
	}
	
	flag = 0;
	
	// Did we find a solution?
	TAlScore score = MIN_I64;
	if(!found) {
		flag = -1; // no
		if(!debug) met.dpfail++;
		return MIN_I64;
	} else {
		score = (TAlScore)(lrmax - 0x7fff);
		if(score < minsc_) {
			flag = -1; // no
			if(!debug) met.dpfail++;
			return score;
		}
	}
	
	// Could we have saturated?
	if(lrmax == MIN_I16) {
		flag = -2; // yes
		if(!debug) met.dpsat++;
		return MIN_I64;
	}
	
	// Return largest score
	if(!debug) met.dpsucc++;
	return score;
}

/**
 * Given a filled-in DP table, populate the btncand_ list with candidate cells
 * that might be at the ends of valid alignments.  No need to do this unless
 * the maximum score returned by the align*() func is >= the minimum.
 *
 * Only cells that are exhaustively scored are candidates.  Those are the
 * cells inside the shape made of o's in this:
 *
 *  |-maxgaps-|
 *  *********************************    -
 *   ********************************    |
 *    *******************************    |
 *     ******************************    |
 *      *****************************    |
 *       **************************** read len
 *        ***************************    |
 *         **************************    |
 *          *************************    |
 *           ************************    |
 *            ***********oooooooooooo    -
 *            |-maxgaps-|
 *  |-readlen-|
 *  |-------skip--------|
 *
 * And it's possible for the shape to be truncated on the left and right sides.
 *
 * 
 */
bool SwAligner::gatherCellsNucleotidesEnd2EndSseI16(TAlScore best) {
	// What's the minimum number of rows that can possibly be spanned by an
	// alignment that meets the minimum score requirement?
	assert(sse16succ_);
	const size_t ncol = rff_ - rfi_;
	const size_t nrow = dpRows();
	assert_gt(nrow, 0);
	btncand_.clear();
	btncanddone_.clear();
	SSEData& d = fw_ ? sseI16fw_ : sseI16rc_;
	SSEMetrics& met = extend_ ? sseI16ExtendMet_ : sseI16MateMet_;
	assert(!d.profbuf_.empty());
	const size_t colstride = d.mat_.colstride();
	ASSERT_ONLY(bool sawbest = false);
	__m128i *pvH = d.mat_.hvec(d.lastIter_, 0);
	for(size_t j = 0; j < ncol; j++) {
		TAlScore sc = (TAlScore)(((TCScore*)pvH)[d.lastWord_] - 0x7fff);
		assert_leq(sc, best);
		ASSERT_ONLY(sawbest = (sawbest || sc == best));
		if(sc >= minsc_) {
			// Yes, this is legit
			met.gathsol++;
			btncand_.expand();
			btncand_.back().init(nrow-1, j, sc);
		}
		pvH += colstride;
	}
	assert(sawbest);
	if(!btncand_.empty()) {
		d.mat_.initMasks();
	}
	return !btncand_.empty();
}

#define MOVE_VEC_PTR_UP(vec, rowvec, rowelt) { \
	if(rowvec == 0) { \
		rowvec += d.mat_.nvecrow_; \
		vec += d.mat_.colstride_; \
		rowelt--; \
	} \
	rowvec--; \
	vec -= ROWSTRIDE; \
}

#define MOVE_VEC_PTR_LEFT(vec, rowvec, rowelt) { vec -= d.mat_.colstride_; }

#define MOVE_VEC_PTR_UPLEFT(vec, rowvec, rowelt) { \
 	MOVE_VEC_PTR_UP(vec, rowvec, rowelt); \
 	MOVE_VEC_PTR_LEFT(vec, rowvec, rowelt); \
}

#define MOVE_ALL_LEFT() { \
	MOVE_VEC_PTR_LEFT(cur_vec, rowvec, rowelt); \
	MOVE_VEC_PTR_LEFT(left_vec, left_rowvec, left_rowelt); \
	MOVE_VEC_PTR_LEFT(up_vec, up_rowvec, up_rowelt); \
	MOVE_VEC_PTR_LEFT(upleft_vec, upleft_rowvec, upleft_rowelt); \
}

#define MOVE_ALL_UP() { \
	MOVE_VEC_PTR_UP(cur_vec, rowvec, rowelt); \
	MOVE_VEC_PTR_UP(left_vec, left_rowvec, left_rowelt); \
	MOVE_VEC_PTR_UP(up_vec, up_rowvec, up_rowelt); \
	MOVE_VEC_PTR_UP(upleft_vec, upleft_rowvec, upleft_rowelt); \
}

#define MOVE_ALL_UPLEFT() { \
	MOVE_VEC_PTR_UPLEFT(cur_vec, rowvec, rowelt); \
	MOVE_VEC_PTR_UPLEFT(left_vec, left_rowvec, left_rowelt); \
	MOVE_VEC_PTR_UPLEFT(up_vec, up_rowvec, up_rowelt); \
	MOVE_VEC_PTR_UPLEFT(upleft_vec, upleft_rowvec, upleft_rowelt); \
}

#define NEW_ROW_COL(row, col) { \
	rowelt = row / d.mat_.nvecrow_; \
	rowvec = row % d.mat_.nvecrow_; \
	eltvec = (col * d.mat_.colstride_) + (rowvec * ROWSTRIDE); \
	cur_vec = d.mat_.matbuf_.ptr() + eltvec; \
	left_vec = cur_vec; \
	left_rowelt = rowelt; \
	left_rowvec = rowvec; \
	MOVE_VEC_PTR_LEFT(left_vec, left_rowvec, left_rowelt); \
	up_vec = cur_vec; \
	up_rowelt = rowelt; \
	up_rowvec = rowvec; \
	MOVE_VEC_PTR_UP(up_vec, up_rowvec, up_rowelt); \
	upleft_vec = up_vec; \
	upleft_rowelt = up_rowelt; \
	upleft_rowvec = up_rowvec; \
	MOVE_VEC_PTR_LEFT(upleft_vec, upleft_rowvec, upleft_rowelt); \
}

/**
 * Given the dynamic programming table and a cell, trace backwards from the
 * cell and install the edits and score/penalty in the appropriate fields
 * of res.  The RandomSource is used to break ties among equally good ways
 * of tracing back.
 *
 * Whenever we enter a cell, we check whether the read/ref coordinates of
 * that cell correspond to a cell we traversed constructing a previous
 * alignment.  If so, we backtrack to the last decision point, mask out the
 * path that led to the previously observed cell, and continue along a
 * different path; or, if there are no more paths to try, we give up.
 *
 * If an alignment is found, 'off' is set to the alignment's upstream-most
 * reference character's offset into the chromosome and true is returned.
 * Otherwise, false is returned.
 */
bool SwAligner::backtraceNucleotidesEnd2EndSseI16(
	TAlScore       escore, // in: expected score
	SwResult&      res,    // out: store results (edits and scores) here
	size_t&        off,    // out: store diagonal projection of origin
	size_t&        nbts,   // out: # backtracks
	size_t         row,    // start in this row
	size_t         col,    // start in this column
	RandomSource&  rnd)    // random gen, to choose among equal paths
{
	assert_lt(row, dpRows());
	assert_lt(col, (size_t)(rff_ - rfi_));
	SSEData& d = fw_ ? sseI16fw_ : sseI16rc_;
	SSEMetrics& met = extend_ ? sseI16ExtendMet_ : sseI16MateMet_;
	met.bt++;
	assert(!d.profbuf_.empty());
	assert_lt(row, rd_->length());
	btnstack_.clear(); // empty the backtrack stack
	btcells_.clear();  // empty the cells-so-far list
	AlnScore score; score.score_ = 0;
	score.gaps_ = score.ns_ = 0;
	size_t origCol = col;
	size_t gaps = 0, readGaps = 0, refGaps = 0;
	res.alres.reset();
	EList<Edit>& ned = res.alres.ned();
	assert(ned.empty());
	assert_gt(dpRows(), row);
	size_t trimEnd = dpRows() - row - 1; 
	size_t trimBeg = 0;
	size_t ct = SSEMatrix::H; // cell type
	// Row and col in terms of where they fall in the SSE vector matrix
	size_t rowelt, rowvec, eltvec;
	size_t left_rowelt, up_rowelt, upleft_rowelt;
	size_t left_rowvec, up_rowvec, upleft_rowvec;
	__m128i *cur_vec, *left_vec, *up_vec, *upleft_vec;
	NEW_ROW_COL(row, col);
	while((int)row >= 0) {
		met.btcell++;
		nbts++;
		int readc = (*rd_)[rdi_ + row];
		int refm  = (int)rf_[rfi_ + col];
		int readq = (*qu_)[row];
		assert_leq(col, origCol);
		// Get score in this cell
		bool empty = false, reportedThru, canMoveThru, branch = false;
		int cur = SSEMatrix::H;
		if(!d.mat_.reset_[row]) {
			d.mat_.resetRow(row);
		}
		reportedThru = d.mat_.reportedThrough(row, col);
		canMoveThru = true;
		if(reportedThru) {
			canMoveThru = false;
		} else {
			empty = false;
			if(row > 0) {
				assert_gt(row, 0);
				size_t rowFromEnd = d.mat_.nrow() - row - 1;
				bool gapsAllowed = true;
				if(row < (size_t)sc_->gapbar ||
				   rowFromEnd < (size_t)sc_->gapbar)
				{
					gapsAllowed = false;
				}
				const TAlScore floorsc = MIN_I64;
				const int offsetsc = -0x7fff;
				// Move to beginning of column/row
				if(ct == SSEMatrix::E) { // AKA rdgap
					assert_gt(col, 0);
					TAlScore sc_cur = ((TCScore*)(cur_vec + SSEMatrix::E))[rowelt] + offsetsc;
					assert(gapsAllowed);
					// Currently in the E matrix; incoming transition must come from the
					// left.  It's either a gap open from the H matrix or a gap extend from
					// the E matrix.
					// TODO: save and restore origMask as well as mask
					int origMask = 0, mask = 0;
					// Get H score of cell to the left
					TAlScore sc_h_left = ((TCScore*)(left_vec + SSEMatrix::H))[left_rowelt] + offsetsc;
					if(sc_h_left > floorsc && sc_h_left - sc_->readGapOpen() == sc_cur) {
						mask |= (1 << 0);
					}
					// Get E score of cell to the left
					TAlScore sc_e_left = ((TCScore*)(left_vec + SSEMatrix::E))[left_rowelt] + offsetsc;
					if(sc_e_left > floorsc && sc_e_left - sc_->readGapExtend() == sc_cur) {
						mask |= (1 << 1);
					}
					origMask = mask;
					assert(origMask > 0 || sc_cur <= sc_->match());
					if(d.mat_.isEMaskSet(row, col)) {
						mask = (d.mat_.masks_[row][col] >> 8) & 3;
					}
					if(mask == 3) {
#if 1
						// Pick H -> E cell
						cur = SW_BT_OALL_READ_OPEN;
						d.mat_.eMaskSet(row, col, 2); // might choose E later
#else
						if(rnd.nextU2()) {
							// Pick H -> E cell
							cur = SW_BT_OALL_READ_OPEN;
							d.mat_.eMaskSet(row, col, 2); // might choose E later
						} else {
							// Pick E -> E cell
							cur = SW_BT_RDGAP_EXTEND;
							d.mat_.eMaskSet(row, col, 1); // might choose H later
						}
#endif
						branch = true;
					} else if(mask == 2) {
						// I chose the E cell
						cur = SW_BT_RDGAP_EXTEND;
						d.mat_.eMaskSet(row, col, 0); // done
					} else if(mask == 1) {
						// I chose the H cell
						cur = SW_BT_OALL_READ_OPEN;
						d.mat_.eMaskSet(row, col, 0); // done
					} else {
						empty = true;
						// It's empty, so the only question left is whether we should be
						// allowed in terimnate in this cell.  If it's got a valid score
						// then we *shouldn't* be allowed to terminate here because that
						// means it's part of a larger alignment that was already reported.
						canMoveThru = (origMask == 0);
					}
					assert(!empty || !canMoveThru);
				} else if(ct == SSEMatrix::F) { // AKA rfgap
					assert_gt(row, 0);
					assert(gapsAllowed);
					TAlScore sc_h_up = ((TCScore*)(up_vec  + SSEMatrix::H))[up_rowelt] + offsetsc;
					TAlScore sc_f_up = ((TCScore*)(up_vec  + SSEMatrix::F))[up_rowelt] + offsetsc;
					TAlScore sc_cur  = ((TCScore*)(cur_vec + SSEMatrix::F))[rowelt] + offsetsc;
					// Currently in the F matrix; incoming transition must come from above.
					// It's either a gap open from the H matrix or a gap extend from the F
					// matrix.
					// TODO: save and restore origMask as well as mask
					int origMask = 0, mask = 0;
					// Get H score of cell above
					if(sc_h_up > floorsc && sc_h_up - sc_->refGapOpen() == sc_cur) {
						mask |= (1 << 0);
					}
					// Get F score of cell above
					if(sc_f_up > floorsc && sc_f_up - sc_->refGapExtend() == sc_cur) {
						mask |= (1 << 1);
					}
					origMask = mask;
					assert(origMask > 0 || sc_cur <= sc_->match());
					if(d.mat_.isFMaskSet(row, col)) {
						mask = (d.mat_.masks_[row][col] >> 11) & 3;
					}
					if(mask == 3) {
#if 1
						// I chose the H cell
						cur = SW_BT_OALL_REF_OPEN;
						d.mat_.fMaskSet(row, col, 2); // might choose E later
#else
						if(rnd.nextU2()) {
							// I chose the H cell
							cur = SW_BT_OALL_REF_OPEN;
							d.mat_.fMaskSet(row, col, 2); // might choose E later
						} else {
							// I chose the F cell
							cur = SW_BT_RFGAP_EXTEND;
							d.mat_.fMaskSet(row, col, 1); // might choose E later
						}
#endif
						branch = true;
					} else if(mask == 2) {
						// I chose the F cell
						cur = SW_BT_RFGAP_EXTEND;
						d.mat_.fMaskSet(row, col, 0); // done
					} else if(mask == 1) {
						// I chose the H cell
						cur = SW_BT_OALL_REF_OPEN;
						d.mat_.fMaskSet(row, col, 0); // done
					} else {
						empty = true;
						// It's empty, so the only question left is whether we should be
						// allowed in terimnate in this cell.  If it's got a valid score
						// then we *shouldn't* be allowed to terminate here because that
						// means it's part of a larger alignment that was already reported.
						canMoveThru = (origMask == 0);
					}
					assert(!empty || !canMoveThru);
				} else {
					assert_eq(SSEMatrix::H, ct);
					TAlScore sc_cur      = ((TCScore*)(cur_vec + SSEMatrix::H))[rowelt]    + offsetsc;
					TAlScore sc_f_up     = ((TCScore*)(up_vec  + SSEMatrix::F))[up_rowelt] + offsetsc;
					TAlScore sc_h_up     = ((TCScore*)(up_vec  + SSEMatrix::H))[up_rowelt] + offsetsc;
					TAlScore sc_h_left   = col > 0 ? (((TCScore*)(left_vec   + SSEMatrix::H))[left_rowelt]   + offsetsc) : floorsc;
					TAlScore sc_e_left   = col > 0 ? (((TCScore*)(left_vec   + SSEMatrix::E))[left_rowelt]   + offsetsc) : floorsc;
					TAlScore sc_h_upleft = col > 0 ? (((TCScore*)(upleft_vec + SSEMatrix::H))[upleft_rowelt] + offsetsc) : floorsc;
					TAlScore sc_diag     = sc_->score(readc, refm, readq - 33);
					// TODO: save and restore origMask as well as mask
					int origMask = 0, mask = 0;
					if(gapsAllowed) {
						if(sc_h_up     > floorsc && sc_cur == sc_h_up   - sc_->refGapOpen()) {
							mask |= (1 << 0);
						}
						if(sc_h_left   > floorsc && sc_cur == sc_h_left - sc_->readGapOpen()) {
							mask |= (1 << 1);
						}
						if(sc_f_up     > floorsc && sc_cur == sc_f_up   - sc_->refGapExtend()) {
							mask |= (1 << 2);
						}
						if(sc_e_left   > floorsc && sc_cur == sc_e_left - sc_->readGapExtend()) {
							mask |= (1 << 3);
						}
					}
					if(sc_h_upleft > floorsc && sc_cur == sc_h_upleft + sc_diag) {
						mask |= (1 << 4);
					}
					origMask = mask;
					assert(origMask > 0 || sc_cur <= sc_->match());
					if(d.mat_.isHMaskSet(row, col)) {
						mask = (d.mat_.masks_[row][col] >> 2) & 31;
					}
					assert(gapsAllowed || mask == (1 << 4) || mask == 0);
					int opts = alts5[mask];
					int select = -1;
					if(opts == 1) {
						select = firsts5[mask];
						assert_geq(mask, 0);
						d.mat_.hMaskSet(row, col, 0);
					} else if(opts > 1) {
#if 1
						if(       (mask & 16) != 0) {
							select = 4; // H diag
						} else if((mask & 1) != 0) {
							select = 0; // H up
						} else if((mask & 4) != 0) {
							select = 2; // F up
						} else if((mask & 2) != 0) {
							select = 1; // H left
						} else if((mask & 8) != 0) {
							select = 3; // E left
						}
#else
						select = randFromMask(rnd, mask);
#endif
						assert_geq(mask, 0);
						mask &= ~(1 << select);
						assert(gapsAllowed || mask == (1 << 4) || mask == 0);
						d.mat_.hMaskSet(row, col, mask);
						branch = true;
					} else { /* No way to backtrack! */ }
					if(select != -1) {
						if(select == 4) {
							cur = SW_BT_OALL_DIAG;
						} else if(select == 0) {
							cur = SW_BT_OALL_REF_OPEN;
						} else if(select == 1) {
							cur = SW_BT_OALL_READ_OPEN;
						} else if(select == 2) {
							cur = SW_BT_RFGAP_EXTEND;
						} else {
							assert_eq(3, select)
							cur = SW_BT_RDGAP_EXTEND;
						}
					} else {
						empty = true;
						// It's empty, so the only question left is whether we should be
						// allowed in terimnate in this cell.  If it's got a valid score
						// then we *shouldn't* be allowed to terminate here because that
						// means it's part of a larger alignment that was already reported.
						canMoveThru = (origMask == 0);
					}
				}
				assert(!empty || !canMoveThru || ct == SSEMatrix::H);
			}
		}
		d.mat_.setReportedThrough(row, col);
		assert_eq(gaps, Edit::numGaps(ned));
		assert_leq(gaps, rdgap_ + rfgap_);
		// Cell was involved in a previously-reported alignment?
		if(!canMoveThru) {
			if(!btnstack_.empty()) {
				// Remove all the cells from list back to and including the
				// cell where the branch occurred
				btcells_.resize(btnstack_.back().celsz);
				// Pop record off the top of the stack
				ned.resize(btnstack_.back().nedsz);
				//aed.resize(btnstack_.back().aedsz);
				row      = btnstack_.back().row;
				col      = btnstack_.back().col;
				gaps     = btnstack_.back().gaps;
				readGaps = btnstack_.back().readGaps;
				refGaps  = btnstack_.back().refGaps;
				score    = btnstack_.back().score;
				ct       = btnstack_.back().ct;
				btnstack_.pop_back();
				assert(!sc_->monotone || score.score() >= escore);
				NEW_ROW_COL(row, col);
				continue;
			} else {
				// No branch points to revisit; just give up
				res.reset();
				met.btfail++; // DP backtraces failed
				return false;
			}
		}
		assert(!reportedThru);
		assert(!sc_->monotone || score.score() >= minsc_);
		if(empty || row == 0) {
			assert_eq(SSEMatrix::H, ct);
			btcells_.expand();
			btcells_.back().first = row;
			btcells_.back().second = col;
			// This cell is at the end of a legitimate alignment
			trimBeg = row;
			assert_eq(btcells_.size(), dpRows() - trimBeg - trimEnd + readGaps);
			break;
		}
		if(branch) {
			// Add a frame to the backtrack stack
			btnstack_.expand();
			btnstack_.back().init(
				ned.size(),
				0,               // aed.size()
				btcells_.size(),
				row,
				col,
				gaps,
				readGaps,
				refGaps,
				score,
				(int)ct);
		}
		btcells_.expand();
		btcells_.back().first = row;
		btcells_.back().second = col;
		switch(cur) {
			// Move up and to the left.  If the reference nucleotide in the
			// source row mismatches the read nucleotide, penalize
			// it and add a nucleotide mismatch.
			case SW_BT_OALL_DIAG: {
				assert_gt(row, 0); assert_gt(col, 0);
				// Check for color mismatch
				int readC = (*rd_)[row];
				int refNmask = (int)rf_[rfi_+col];
				assert_gt(refNmask, 0);
				int m = matchesEx(readC, refNmask);
				ct = SSEMatrix::H;
				if(m != 1) {
					Edit e(
						(int)row,
						mask2dna[refNmask],
						"ACGTN"[readC],
						EDIT_TYPE_MM);
					assert(e.repOk());
					assert(ned.empty() || ned.back().pos >= row);
					ned.push_back(e);
					int pen = QUAL2(row, col);
					score.score_ -= pen;
					assert(!sc_->monotone || score.score() >= escore);
				} else {
					// Reward a match
					int64_t bonus = sc_->match(30);
					score.score_ += bonus;
					assert(!sc_->monotone || score.score() >= escore);
				}
				if(m == -1) {
					score.ns_++;
				}
				row--; col--;
				MOVE_ALL_UPLEFT();
				assert(VALID_AL_SCORE(score));
				break;
			}
			// Move up.  Add an edit encoding the ref gap.
			case SW_BT_OALL_REF_OPEN:
			{
				assert_gt(row, 0);
				Edit e(
					(int)row,
					'-',
					"ACGTN"[(int)(*rd_)[row]],
					EDIT_TYPE_REF_GAP);
				assert(e.repOk());
				assert(ned.empty() || ned.back().pos >= row);
				ned.push_back(e);
				assert_geq(row, (size_t)sc_->gapbar);
				assert_geq((int)(rdf_-rdi_-row-1), sc_->gapbar-1);
				row--;
				ct = SSEMatrix::H;
				int pen = sc_->refGapOpen();
				score.score_ -= pen;
				assert(!sc_->monotone || score.score() >= minsc_);
				gaps++; refGaps++;
				assert_eq(gaps, Edit::numGaps(ned));
				assert_leq(gaps, rdgap_ + rfgap_);
				MOVE_ALL_UP();
				break;
			}
			// Move up.  Add an edit encoding the ref gap.
			case SW_BT_RFGAP_EXTEND:
			{
				assert_gt(row, 1);
				Edit e(
					(int)row,
					'-',
					"ACGTN"[(int)(*rd_)[row]],
					EDIT_TYPE_REF_GAP);
				assert(e.repOk());
				assert(ned.empty() || ned.back().pos >= row);
				ned.push_back(e);
				assert_geq(row, (size_t)sc_->gapbar);
				assert_geq((int)(rdf_-rdi_-row-1), sc_->gapbar-1);
				row--;
				ct = SSEMatrix::F;
				int pen = sc_->refGapExtend();
				score.score_ -= pen;
				assert(!sc_->monotone || score.score() >= minsc_);
				gaps++; refGaps++;
				assert_eq(gaps, Edit::numGaps(ned));
				assert_leq(gaps, rdgap_ + rfgap_);
				MOVE_ALL_UP();
				break;
			}
			case SW_BT_OALL_READ_OPEN:
			{
				assert_gt(col, 0);
				Edit e(
					(int)row+1,
					mask2dna[(int)rf_[rfi_+col]],
					'-',
					EDIT_TYPE_READ_GAP);
				assert(e.repOk());
				assert(ned.empty() || ned.back().pos >= row);
				ned.push_back(e);
				assert_geq(row, (size_t)sc_->gapbar);
				assert_geq((int)(rdf_-rdi_-row-1), sc_->gapbar-1);
				col--;
				ct = SSEMatrix::H;
				int pen = sc_->readGapOpen();
				score.score_ -= pen;
				assert(!sc_->monotone || score.score() >= minsc_);
				gaps++; readGaps++;
				assert_eq(gaps, Edit::numGaps(ned));
				assert_leq(gaps, rdgap_ + rfgap_);
				MOVE_ALL_LEFT();
				break;
			}
			case SW_BT_RDGAP_EXTEND:
			{
				assert_gt(col, 1);
				Edit e(
					(int)row+1,
					mask2dna[(int)rf_[rfi_+col]],
					'-',
					EDIT_TYPE_READ_GAP);
				assert(e.repOk());
				assert(ned.empty() || ned.back().pos >= row);
				ned.push_back(e);
				assert_geq(row, (size_t)sc_->gapbar);
				assert_geq((int)(rdf_-rdi_-row-1), sc_->gapbar-1);
				col--;
				ct = SSEMatrix::E;
				int pen = sc_->readGapExtend();
				score.score_ -= pen;
				assert(!sc_->monotone || score.score() >= minsc_);
				gaps++; readGaps++;
				assert_eq(gaps, Edit::numGaps(ned));
				assert_leq(gaps, rdgap_ + rfgap_);
				MOVE_ALL_LEFT();
				break;
			}
			default: throw 1;
		}
	} // while((int)row > 0)
	assert_eq(0, trimBeg);
	assert_eq(0, trimEnd);
	assert_geq(col, 0);
	assert_eq(SSEMatrix::H, ct);
	// The number of cells in the backtracs should equal the number of read
	// bases after trimming plus the number of gaps
	assert_eq(btcells_.size(), dpRows() - trimBeg - trimEnd + readGaps);
	// Check whether we went through a core diagonal and set 'reported' flag on
	// each cell
	bool overlappedCoreDiag = false;
	for(size_t i = 0; i < btcells_.size(); i++) {
		size_t rw = btcells_[i].first;
		size_t cl = btcells_[i].second;
		// Calculate the diagonal within the *trimmed* rectangle, i.e. the
		// rectangle we dealt with in align, gather and backtrack.
		int64_t diagi = cl - rw;
		// Now adjust to the diagonal within the *untrimmed* rectangle by
		// adding on the amount trimmed from the left.
		diagi += rect_->triml;
		if(diagi >= 0) {
			size_t diag = (size_t)diagi;
			if(diag >= rect_->corel && diag <= rect_->corer) {
				overlappedCoreDiag = true;
				break;
			}
		}
		assert(d.mat_.reportedThrough(rw, cl));
	}
	if(!overlappedCoreDiag) {
		// Must overlap a core diagonal.  Otherwise, we run the risk of
		// reporting an alignment that overlaps (and trumps) a higher-scoring
		// alignment that lies partially outside the dynamic programming
		// rectangle.
		res.reset();
		met.corerej++;
		return false;
	}
	int readC = (*rd_)[rdi_+row];      // get last char in read
	int refNmask = (int)rf_[rfi_+col]; // get last ref char ref involved in aln
	assert_gt(refNmask, 0);
	int m = matchesEx(readC, refNmask);
	if(m != 1) {
		Edit e((int)row, mask2dna[refNmask], "ACGTN"[readC], EDIT_TYPE_MM);
		assert(e.repOk());
		assert(ned.empty() || ned.back().pos >= row);
		ned.push_back(e);
		score.score_ -= QUAL2(row, col);
		assert_geq(score.score(), minsc_);
	} else {
		score.score_ += sc_->match(30);
	}
	if(m == -1) {
		score.ns_++;
	}
	if(score.ns_ > nceil_) {
		// Alignment has too many Ns in it!
		res.reset();
		met.nrej++;
		return false;
	}
	res.reverse();
	assert(Edit::repOk(ned, (*rd_)));
	assert_eq(score.score(), escore);
	assert_leq(gaps, rdgap_ + rfgap_);
	off = col;
	assert_lt(col + (size_t)rfi_, (size_t)rff_);
	score.gaps_ = gaps;
	res.alres.setScore(score);
	res.alres.setShape(
		refidx_,                  // ref id
		off + rfi_ + rect_->refl, // 0-based ref offset
		reflen_,                  // reference length
		fw_,                      // aligned to Watson?
		rdf_ - rdi_,              // read length
        0,                        // read ID
		true,                     // pretrim soft?
		0,                        // pretrim 5' end
		0,                        // pretrim 3' end
		true,                     // alignment trim soft?
		fw_ ? trimBeg : trimEnd,  // alignment trim 5' end
		fw_ ? trimEnd : trimBeg); // alignment trim 3' end
	size_t refns = 0;
	for(size_t i = col; i <= origCol; i++) {
		if((int)rf_[rfi_+i] > 15) {
			refns++;
		}
	}
	res.alres.setRefNs(refns);
	assert(Edit::repOk(ned, (*rd_), true, trimBeg, trimEnd));
	assert(res.repOk());
#ifndef NDEBUG
	size_t gapsCheck = 0;
	for(size_t i = 0; i < ned.size(); i++) {
		if(ned[i].isGap()) gapsCheck++;
	}
	assert_eq(gaps, gapsCheck);
	BTDnaString refstr;
	for(size_t i = col; i <= origCol; i++) {
		refstr.append(firsts5[(int)rf_[rfi_+i]]);
	}
	BTDnaString editstr;
    // daehwan
	// Edit::toRef((*rd_), ned, editstr, true, trimBeg, trimEnd);
    Edit::toRef((*rd_), ned, editstr, true, trimBeg + rdi_, trimEnd + (rd_->length() - rdf_));
	if(refstr != editstr) {
		cerr << "Decoded nucleotides and edits don't match reference:" << endl;
		cerr << "           score: " << score.score()
		     << " (" << gaps << " gaps)" << endl;
		cerr << "           edits: ";
		Edit::print(cerr, ned);
		cerr << endl;
		cerr << "    decoded nucs: " << (*rd_) << endl;
		cerr << "     edited nucs: " << editstr << endl;
		cerr << "  reference nucs: " << refstr << endl;
		assert(0);
	}
#endif
	met.btsucc++; // DP backtraces succeeded
	return true;
}