edlib 0.0.3 → 0.0.5
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/ext/edlib/edlib.cpp +1482 -0
- data/ext/edlib/edlibext.c +1 -7
- data/lib/edlib/version.rb +1 -1
- metadata +2 -1
data/ext/edlib/edlib.cpp
ADDED
@@ -0,0 +1,1482 @@
|
|
1
|
+
#include "edlib.h"
|
2
|
+
|
3
|
+
#include <stdint.h>
|
4
|
+
#include <cstdlib>
|
5
|
+
#include <algorithm>
|
6
|
+
#include <vector>
|
7
|
+
#include <cstring>
|
8
|
+
#include <string>
|
9
|
+
|
10
|
+
using namespace std;
|
11
|
+
|
12
|
+
typedef uint64_t Word;
|
13
|
+
static const int WORD_SIZE = sizeof(Word) * 8; // Size of Word in bits
|
14
|
+
static const Word WORD_1 = static_cast<Word>(1);
|
15
|
+
static const Word HIGH_BIT_MASK = WORD_1 << (WORD_SIZE - 1); // 100..00
|
16
|
+
static const int MAX_UCHAR = 255;
|
17
|
+
|
18
|
+
// Data needed to find alignment.
|
19
|
+
struct AlignmentData {
|
20
|
+
Word* Ps;
|
21
|
+
Word* Ms;
|
22
|
+
int* scores;
|
23
|
+
int* firstBlocks;
|
24
|
+
int* lastBlocks;
|
25
|
+
|
26
|
+
AlignmentData(int maxNumBlocks, int targetLength) {
|
27
|
+
// We build a complete table and mark first and last block for each column
|
28
|
+
// (because algorithm is banded so only part of each columns is used).
|
29
|
+
// TODO: do not build a whole table, but just enough blocks for each column.
|
30
|
+
Ps = new Word[maxNumBlocks * targetLength];
|
31
|
+
Ms = new Word[maxNumBlocks * targetLength];
|
32
|
+
scores = new int[maxNumBlocks * targetLength];
|
33
|
+
firstBlocks = new int[targetLength];
|
34
|
+
lastBlocks = new int[targetLength];
|
35
|
+
}
|
36
|
+
|
37
|
+
~AlignmentData() {
|
38
|
+
delete[] Ps;
|
39
|
+
delete[] Ms;
|
40
|
+
delete[] scores;
|
41
|
+
delete[] firstBlocks;
|
42
|
+
delete[] lastBlocks;
|
43
|
+
}
|
44
|
+
};
|
45
|
+
|
46
|
+
struct Block {
|
47
|
+
Word P; // Pvin
|
48
|
+
Word M; // Mvin
|
49
|
+
int score; // score of last cell in block;
|
50
|
+
|
51
|
+
Block() {}
|
52
|
+
Block(Word p, Word m, int s) :P(p), M(m), score(s) {}
|
53
|
+
};
|
54
|
+
|
55
|
+
|
56
|
+
/**
|
57
|
+
* Defines equality relation on alphabet characters.
|
58
|
+
* By default each character is always equal only to itself, but you can also provide additional equalities.
|
59
|
+
*/
|
60
|
+
class EqualityDefinition {
|
61
|
+
private:
|
62
|
+
bool matrix[MAX_UCHAR + 1][MAX_UCHAR + 1];
|
63
|
+
public:
|
64
|
+
EqualityDefinition(const string& alphabet,
|
65
|
+
const EdlibEqualityPair* additionalEqualities = NULL,
|
66
|
+
const int additionalEqualitiesLength = 0) {
|
67
|
+
for (int i = 0; i < static_cast<int>(alphabet.size()); i++) {
|
68
|
+
for (int j = 0; j < static_cast<int>(alphabet.size()); j++) {
|
69
|
+
matrix[i][j] = (i == j);
|
70
|
+
}
|
71
|
+
}
|
72
|
+
if (additionalEqualities != NULL) {
|
73
|
+
for (int i = 0; i < additionalEqualitiesLength; i++) {
|
74
|
+
size_t firstTransformed = alphabet.find(additionalEqualities[i].first);
|
75
|
+
size_t secondTransformed = alphabet.find(additionalEqualities[i].second);
|
76
|
+
if (firstTransformed != string::npos && secondTransformed != string::npos) {
|
77
|
+
matrix[firstTransformed][secondTransformed] = matrix[secondTransformed][firstTransformed] = true;
|
78
|
+
}
|
79
|
+
}
|
80
|
+
}
|
81
|
+
}
|
82
|
+
|
83
|
+
/**
|
84
|
+
* @param a Element from transformed sequence.
|
85
|
+
* @param b Element from transformed sequence.
|
86
|
+
* @return True if a and b are defined as equal, false otherwise.
|
87
|
+
*/
|
88
|
+
bool areEqual(unsigned char a, unsigned char b) const {
|
89
|
+
return matrix[a][b];
|
90
|
+
}
|
91
|
+
};
|
92
|
+
|
93
|
+
static int myersCalcEditDistanceSemiGlobal(const Word* Peq, int W, int maxNumBlocks,
|
94
|
+
int queryLength,
|
95
|
+
const unsigned char* target, int targetLength,
|
96
|
+
int k, EdlibAlignMode mode,
|
97
|
+
int* bestScore_, int** positions_, int* numPositions_);
|
98
|
+
|
99
|
+
static int myersCalcEditDistanceNW(const Word* Peq, int W, int maxNumBlocks,
|
100
|
+
int queryLength,
|
101
|
+
const unsigned char* target, int targetLength,
|
102
|
+
int k, int* bestScore_,
|
103
|
+
int* position_, bool findAlignment,
|
104
|
+
AlignmentData** alignData, int targetStopPosition);
|
105
|
+
|
106
|
+
|
107
|
+
static int obtainAlignment(
|
108
|
+
const unsigned char* query, const unsigned char* rQuery, int queryLength,
|
109
|
+
const unsigned char* target, const unsigned char* rTarget, int targetLength,
|
110
|
+
const EqualityDefinition& equalityDefinition, int alphabetLength, int bestScore,
|
111
|
+
unsigned char** alignment, int* alignmentLength);
|
112
|
+
|
113
|
+
static int obtainAlignmentHirschberg(
|
114
|
+
const unsigned char* query, const unsigned char* rQuery, int queryLength,
|
115
|
+
const unsigned char* target, const unsigned char* rTarget, int targetLength,
|
116
|
+
const EqualityDefinition& equalityDefinition, int alphabetLength, int bestScore,
|
117
|
+
unsigned char** alignment, int* alignmentLength);
|
118
|
+
|
119
|
+
static int obtainAlignmentTraceback(int queryLength, int targetLength,
|
120
|
+
int bestScore, const AlignmentData* alignData,
|
121
|
+
unsigned char** alignment, int* alignmentLength);
|
122
|
+
|
123
|
+
static string transformSequences(const char* queryOriginal, int queryLength,
|
124
|
+
const char* targetOriginal, int targetLength,
|
125
|
+
unsigned char** queryTransformed,
|
126
|
+
unsigned char** targetTransformed);
|
127
|
+
|
128
|
+
static inline int ceilDiv(int x, int y);
|
129
|
+
|
130
|
+
static inline unsigned char* createReverseCopy(const unsigned char* seq, int length);
|
131
|
+
|
132
|
+
static inline Word* buildPeq(const int alphabetLength,
|
133
|
+
const unsigned char* query,
|
134
|
+
const int queryLength,
|
135
|
+
const EqualityDefinition& equalityDefinition);
|
136
|
+
|
137
|
+
|
138
|
+
/**
|
139
|
+
* Main edlib method.
|
140
|
+
*/
|
141
|
+
extern "C" EdlibAlignResult edlibAlign(const char* const queryOriginal, const int queryLength,
|
142
|
+
const char* const targetOriginal, const int targetLength,
|
143
|
+
const EdlibAlignConfig config) {
|
144
|
+
EdlibAlignResult result;
|
145
|
+
result.status = EDLIB_STATUS_OK;
|
146
|
+
result.editDistance = -1;
|
147
|
+
result.endLocations = result.startLocations = NULL;
|
148
|
+
result.numLocations = 0;
|
149
|
+
result.alignment = NULL;
|
150
|
+
result.alignmentLength = 0;
|
151
|
+
result.alphabetLength = 0;
|
152
|
+
|
153
|
+
/*------------ TRANSFORM SEQUENCES AND RECOGNIZE ALPHABET -----------*/
|
154
|
+
unsigned char* query, * target;
|
155
|
+
string alphabet = transformSequences(queryOriginal, queryLength, targetOriginal, targetLength,
|
156
|
+
&query, &target);
|
157
|
+
result.alphabetLength = static_cast<int>(alphabet.size());
|
158
|
+
/*-------------------------------------------------------*/
|
159
|
+
|
160
|
+
// Handle special situation when at least one of the sequences has length 0.
|
161
|
+
if (queryLength == 0 || targetLength == 0) {
|
162
|
+
if (config.mode == EDLIB_MODE_NW) {
|
163
|
+
result.editDistance = std::max(queryLength, targetLength);
|
164
|
+
result.endLocations = static_cast<int *>(malloc(sizeof(int) * 1));
|
165
|
+
result.endLocations[0] = targetLength - 1;
|
166
|
+
result.numLocations = 1;
|
167
|
+
} else if (config.mode == EDLIB_MODE_SHW || config.mode == EDLIB_MODE_HW) {
|
168
|
+
result.editDistance = queryLength;
|
169
|
+
result.endLocations = static_cast<int *>(malloc(sizeof(int) * 1));
|
170
|
+
result.endLocations[0] = -1;
|
171
|
+
result.numLocations = 1;
|
172
|
+
} else {
|
173
|
+
result.status = EDLIB_STATUS_ERROR;
|
174
|
+
}
|
175
|
+
|
176
|
+
free(query);
|
177
|
+
free(target);
|
178
|
+
return result;
|
179
|
+
}
|
180
|
+
|
181
|
+
/*--------------------- INITIALIZATION ------------------*/
|
182
|
+
int maxNumBlocks = ceilDiv(queryLength, WORD_SIZE); // bmax in Myers
|
183
|
+
int W = maxNumBlocks * WORD_SIZE - queryLength; // number of redundant cells in last level blocks
|
184
|
+
EqualityDefinition equalityDefinition(alphabet, config.additionalEqualities, config.additionalEqualitiesLength);
|
185
|
+
Word* Peq = buildPeq(static_cast<int>(alphabet.size()), query, queryLength, equalityDefinition);
|
186
|
+
/*-------------------------------------------------------*/
|
187
|
+
|
188
|
+
/*------------------ MAIN CALCULATION -------------------*/
|
189
|
+
// TODO: Store alignment data only after k is determined? That could make things faster.
|
190
|
+
int positionNW; // Used only when mode is NW.
|
191
|
+
AlignmentData* alignData = NULL;
|
192
|
+
bool dynamicK = false;
|
193
|
+
int k = config.k;
|
194
|
+
if (k < 0) { // If valid k is not given, auto-adjust k until solution is found.
|
195
|
+
dynamicK = true;
|
196
|
+
k = WORD_SIZE; // Gives better results than smaller k.
|
197
|
+
}
|
198
|
+
|
199
|
+
do {
|
200
|
+
if (config.mode == EDLIB_MODE_HW || config.mode == EDLIB_MODE_SHW) {
|
201
|
+
myersCalcEditDistanceSemiGlobal(Peq, W, maxNumBlocks,
|
202
|
+
queryLength, target, targetLength,
|
203
|
+
k, config.mode, &(result.editDistance),
|
204
|
+
&(result.endLocations), &(result.numLocations));
|
205
|
+
} else { // mode == EDLIB_MODE_NW
|
206
|
+
myersCalcEditDistanceNW(Peq, W, maxNumBlocks,
|
207
|
+
queryLength, target, targetLength,
|
208
|
+
k, &(result.editDistance), &positionNW,
|
209
|
+
false, &alignData, -1);
|
210
|
+
}
|
211
|
+
k *= 2;
|
212
|
+
} while(dynamicK && result.editDistance == -1);
|
213
|
+
|
214
|
+
if (result.editDistance >= 0) { // If there is solution.
|
215
|
+
// If NW mode, set end location explicitly.
|
216
|
+
if (config.mode == EDLIB_MODE_NW) {
|
217
|
+
result.endLocations = static_cast<int *>(malloc(sizeof(int) * 1));
|
218
|
+
result.endLocations[0] = targetLength - 1;
|
219
|
+
result.numLocations = 1;
|
220
|
+
}
|
221
|
+
|
222
|
+
// Find starting locations.
|
223
|
+
if (config.task == EDLIB_TASK_LOC || config.task == EDLIB_TASK_PATH) {
|
224
|
+
result.startLocations = static_cast<int *>(malloc(result.numLocations * sizeof(int)));
|
225
|
+
if (config.mode == EDLIB_MODE_HW) { // If HW, I need to calculate start locations.
|
226
|
+
const unsigned char* rTarget = createReverseCopy(target, targetLength);
|
227
|
+
const unsigned char* rQuery = createReverseCopy(query, queryLength);
|
228
|
+
// Peq for reversed query.
|
229
|
+
Word* rPeq = buildPeq(static_cast<int>(alphabet.size()), rQuery, queryLength, equalityDefinition);
|
230
|
+
for (int i = 0; i < result.numLocations; i++) {
|
231
|
+
int endLocation = result.endLocations[i];
|
232
|
+
if (endLocation == -1) {
|
233
|
+
// NOTE: Sometimes one of optimal solutions is that query starts before target, like this:
|
234
|
+
// AAGG <- target
|
235
|
+
// CCTT <- query
|
236
|
+
// It will never be only optimal solution and it does not happen often, however it is
|
237
|
+
// possible and in that case end location will be -1. What should we do with that?
|
238
|
+
// Should we just skip reporting such end location, although it is a solution?
|
239
|
+
// If we do report it, what is the start location? -4? -1? Nothing?
|
240
|
+
// TODO: Figure this out. This has to do in general with how we think about start
|
241
|
+
// and end locations.
|
242
|
+
// Also, we have alignment later relying on this locations to limit the space of it's
|
243
|
+
// search -> how can it do it right if these locations are negative or incorrect?
|
244
|
+
result.startLocations[i] = 0; // I put 0 for now, but it does not make much sense.
|
245
|
+
} else {
|
246
|
+
int bestScoreSHW, numPositionsSHW;
|
247
|
+
int* positionsSHW;
|
248
|
+
myersCalcEditDistanceSemiGlobal(
|
249
|
+
rPeq, W, maxNumBlocks,
|
250
|
+
queryLength, rTarget + targetLength - endLocation - 1, endLocation + 1,
|
251
|
+
result.editDistance, EDLIB_MODE_SHW,
|
252
|
+
&bestScoreSHW, &positionsSHW, &numPositionsSHW);
|
253
|
+
// Taking last location as start ensures that alignment will not start with insertions
|
254
|
+
// if it can start with mismatches instead.
|
255
|
+
result.startLocations[i] = endLocation - positionsSHW[numPositionsSHW - 1];
|
256
|
+
free(positionsSHW);
|
257
|
+
}
|
258
|
+
}
|
259
|
+
delete[] rTarget;
|
260
|
+
delete[] rQuery;
|
261
|
+
delete[] rPeq;
|
262
|
+
} else { // If mode is SHW or NW
|
263
|
+
for (int i = 0; i < result.numLocations; i++) {
|
264
|
+
result.startLocations[i] = 0;
|
265
|
+
}
|
266
|
+
}
|
267
|
+
}
|
268
|
+
|
269
|
+
// Find alignment -> all comes down to finding alignment for NW.
|
270
|
+
// Currently we return alignment only for first pair of locations.
|
271
|
+
if (config.task == EDLIB_TASK_PATH) {
|
272
|
+
int alnStartLocation = result.startLocations[0];
|
273
|
+
int alnEndLocation = result.endLocations[0];
|
274
|
+
const unsigned char* alnTarget = target + alnStartLocation;
|
275
|
+
const int alnTargetLength = alnEndLocation - alnStartLocation + 1;
|
276
|
+
const unsigned char* rAlnTarget = createReverseCopy(alnTarget, alnTargetLength);
|
277
|
+
const unsigned char* rQuery = createReverseCopy(query, queryLength);
|
278
|
+
obtainAlignment(query, rQuery, queryLength,
|
279
|
+
alnTarget, rAlnTarget, alnTargetLength,
|
280
|
+
equalityDefinition, static_cast<int>(alphabet.size()), result.editDistance,
|
281
|
+
&(result.alignment), &(result.alignmentLength));
|
282
|
+
delete[] rAlnTarget;
|
283
|
+
delete[] rQuery;
|
284
|
+
}
|
285
|
+
}
|
286
|
+
/*-------------------------------------------------------*/
|
287
|
+
|
288
|
+
//--- Free memory ---//
|
289
|
+
delete[] Peq;
|
290
|
+
free(query);
|
291
|
+
free(target);
|
292
|
+
if (alignData) delete alignData;
|
293
|
+
//-------------------//
|
294
|
+
|
295
|
+
return result;
|
296
|
+
}
|
297
|
+
|
298
|
+
extern "C" char* edlibAlignmentToCigar(const unsigned char* const alignment, const int alignmentLength,
|
299
|
+
const EdlibCigarFormat cigarFormat) {
|
300
|
+
if (cigarFormat != EDLIB_CIGAR_EXTENDED && cigarFormat != EDLIB_CIGAR_STANDARD) {
|
301
|
+
return 0;
|
302
|
+
}
|
303
|
+
|
304
|
+
// Maps move code from alignment to char in cigar.
|
305
|
+
// 0 1 2 3
|
306
|
+
char moveCodeToChar[] = {'=', 'I', 'D', 'X'};
|
307
|
+
if (cigarFormat == EDLIB_CIGAR_STANDARD) {
|
308
|
+
moveCodeToChar[0] = moveCodeToChar[3] = 'M';
|
309
|
+
}
|
310
|
+
|
311
|
+
vector<char>* cigar = new vector<char>();
|
312
|
+
char lastMove = 0; // Char of last move. 0 if there was no previous move.
|
313
|
+
int numOfSameMoves = 0;
|
314
|
+
for (int i = 0; i <= alignmentLength; i++) {
|
315
|
+
// if new sequence of same moves started
|
316
|
+
if (i == alignmentLength || (moveCodeToChar[alignment[i]] != lastMove && lastMove != 0)) {
|
317
|
+
// Write number of moves to cigar string.
|
318
|
+
int numDigits = 0;
|
319
|
+
for (; numOfSameMoves; numOfSameMoves /= 10) {
|
320
|
+
cigar->push_back('0' + numOfSameMoves % 10);
|
321
|
+
numDigits++;
|
322
|
+
}
|
323
|
+
reverse(cigar->end() - numDigits, cigar->end());
|
324
|
+
// Write code of move to cigar string.
|
325
|
+
cigar->push_back(lastMove);
|
326
|
+
// If not at the end, start new sequence of moves.
|
327
|
+
if (i < alignmentLength) {
|
328
|
+
// Check if alignment has valid values.
|
329
|
+
if (alignment[i] > 3) {
|
330
|
+
delete cigar;
|
331
|
+
return 0;
|
332
|
+
}
|
333
|
+
numOfSameMoves = 0;
|
334
|
+
}
|
335
|
+
}
|
336
|
+
if (i < alignmentLength) {
|
337
|
+
lastMove = moveCodeToChar[alignment[i]];
|
338
|
+
numOfSameMoves++;
|
339
|
+
}
|
340
|
+
}
|
341
|
+
cigar->push_back(0); // Null character termination.
|
342
|
+
char* cigar_ = static_cast<char *>(malloc(cigar->size() * sizeof(char)));
|
343
|
+
memcpy(cigar_, &(*cigar)[0], cigar->size() * sizeof(char));
|
344
|
+
delete cigar;
|
345
|
+
|
346
|
+
return cigar_;
|
347
|
+
}
|
348
|
+
|
349
|
+
/**
|
350
|
+
* Build Peq table for given query and alphabet.
|
351
|
+
* Peq is table of dimensions alphabetLength+1 x maxNumBlocks.
|
352
|
+
* Bit i of Peq[s * maxNumBlocks + b] is 1 if i-th symbol from block b of query equals symbol s, otherwise it is 0.
|
353
|
+
* NOTICE: free returned array with delete[]!
|
354
|
+
*/
|
355
|
+
static inline Word* buildPeq(const int alphabetLength,
|
356
|
+
const unsigned char* const query,
|
357
|
+
const int queryLength,
|
358
|
+
const EqualityDefinition& equalityDefinition) {
|
359
|
+
int maxNumBlocks = ceilDiv(queryLength, WORD_SIZE);
|
360
|
+
// table of dimensions alphabetLength+1 x maxNumBlocks. Last symbol is wildcard.
|
361
|
+
Word* Peq = new Word[(alphabetLength + 1) * maxNumBlocks];
|
362
|
+
|
363
|
+
// Build Peq (1 is match, 0 is mismatch). NOTE: last column is wildcard(symbol that matches anything) with just 1s
|
364
|
+
for (int symbol = 0; symbol <= alphabetLength; symbol++) {
|
365
|
+
for (int b = 0; b < maxNumBlocks; b++) {
|
366
|
+
if (symbol < alphabetLength) {
|
367
|
+
Peq[symbol * maxNumBlocks + b] = 0;
|
368
|
+
for (int r = (b+1) * WORD_SIZE - 1; r >= b * WORD_SIZE; r--) {
|
369
|
+
Peq[symbol * maxNumBlocks + b] <<= 1;
|
370
|
+
// NOTE: We pretend like query is padded at the end with W wildcard symbols
|
371
|
+
if (r >= queryLength || equalityDefinition.areEqual(query[r], symbol))
|
372
|
+
Peq[symbol * maxNumBlocks + b] += 1;
|
373
|
+
}
|
374
|
+
} else { // Last symbol is wildcard, so it is all 1s
|
375
|
+
Peq[symbol * maxNumBlocks + b] = static_cast<Word>(-1);
|
376
|
+
}
|
377
|
+
}
|
378
|
+
}
|
379
|
+
|
380
|
+
return Peq;
|
381
|
+
}
|
382
|
+
|
383
|
+
|
384
|
+
/**
|
385
|
+
* Returns new sequence that is reverse of given sequence.
|
386
|
+
* Free returned array with delete[].
|
387
|
+
*/
|
388
|
+
static inline unsigned char* createReverseCopy(const unsigned char* const seq, const int length) {
|
389
|
+
unsigned char* rSeq = new unsigned char[length];
|
390
|
+
for (int i = 0; i < length; i++) {
|
391
|
+
rSeq[i] = seq[length - i - 1];
|
392
|
+
}
|
393
|
+
return rSeq;
|
394
|
+
}
|
395
|
+
|
396
|
+
/**
|
397
|
+
* Corresponds to Advance_Block function from Myers.
|
398
|
+
* Calculates one word(block), which is part of a column.
|
399
|
+
* Highest bit of word (one most to the left) is most bottom cell of block from column.
|
400
|
+
* Pv[i] and Mv[i] define vin of cell[i]: vin = cell[i] - cell[i-1].
|
401
|
+
* @param [in] Pv Bitset, Pv[i] == 1 if vin is +1, otherwise Pv[i] == 0.
|
402
|
+
* @param [in] Mv Bitset, Mv[i] == 1 if vin is -1, otherwise Mv[i] == 0.
|
403
|
+
* @param [in] Eq Bitset, Eq[i] == 1 if match, 0 if mismatch.
|
404
|
+
* @param [in] hin Will be +1, 0 or -1.
|
405
|
+
* @param [out] PvOut Bitset, PvOut[i] == 1 if vout is +1, otherwise PvOut[i] == 0.
|
406
|
+
* @param [out] MvOut Bitset, MvOut[i] == 1 if vout is -1, otherwise MvOut[i] == 0.
|
407
|
+
* @param [out] hout Will be +1, 0 or -1.
|
408
|
+
*/
|
409
|
+
static inline int calculateBlock(Word Pv, Word Mv, Word Eq, const int hin,
|
410
|
+
Word &PvOut, Word &MvOut) {
|
411
|
+
// hin can be 1, -1 or 0.
|
412
|
+
// 1 -> 00...01
|
413
|
+
// 0 -> 00...00
|
414
|
+
// -1 -> 11...11 (2-complement)
|
415
|
+
|
416
|
+
Word hinIsNeg = static_cast<Word>(hin >> 2) & WORD_1; // 00...001 if hin is -1, 00...000 if 0 or 1
|
417
|
+
|
418
|
+
Word Xv = Eq | Mv;
|
419
|
+
// This is instruction below written using 'if': if (hin < 0) Eq |= (Word)1;
|
420
|
+
Eq |= hinIsNeg;
|
421
|
+
Word Xh = (((Eq & Pv) + Pv) ^ Pv) | Eq;
|
422
|
+
|
423
|
+
Word Ph = Mv | ~(Xh | Pv);
|
424
|
+
Word Mh = Pv & Xh;
|
425
|
+
|
426
|
+
int hout = 0;
|
427
|
+
// This is instruction below written using 'if': if (Ph & HIGH_BIT_MASK) hout = 1;
|
428
|
+
hout = (Ph & HIGH_BIT_MASK) >> (WORD_SIZE - 1);
|
429
|
+
// This is instruction below written using 'if': if (Mh & HIGH_BIT_MASK) hout = -1;
|
430
|
+
hout -= (Mh & HIGH_BIT_MASK) >> (WORD_SIZE - 1);
|
431
|
+
|
432
|
+
Ph <<= 1;
|
433
|
+
Mh <<= 1;
|
434
|
+
|
435
|
+
// This is instruction below written using 'if': if (hin < 0) Mh |= (Word)1;
|
436
|
+
Mh |= hinIsNeg;
|
437
|
+
// This is instruction below written using 'if': if (hin > 0) Ph |= (Word)1;
|
438
|
+
Ph |= static_cast<Word>((hin + 1) >> 1);
|
439
|
+
|
440
|
+
PvOut = Mh | ~(Xv | Ph);
|
441
|
+
MvOut = Ph & Xv;
|
442
|
+
|
443
|
+
return hout;
|
444
|
+
}
|
445
|
+
|
446
|
+
/**
|
447
|
+
* Does ceiling division x / y.
|
448
|
+
* Note: x and y must be non-negative and x + y must not overflow.
|
449
|
+
*/
|
450
|
+
static inline int ceilDiv(const int x, const int y) {
|
451
|
+
return x % y ? x / y + 1 : x / y;
|
452
|
+
}
|
453
|
+
|
454
|
+
static inline int min(const int x, const int y) {
|
455
|
+
return x < y ? x : y;
|
456
|
+
}
|
457
|
+
|
458
|
+
static inline int max(const int x, const int y) {
|
459
|
+
return x > y ? x : y;
|
460
|
+
}
|
461
|
+
|
462
|
+
|
463
|
+
/**
|
464
|
+
* @param [in] block
|
465
|
+
* @return Values of cells in block, starting with bottom cell in block.
|
466
|
+
*/
|
467
|
+
static inline vector<int> getBlockCellValues(const Block block) {
|
468
|
+
vector<int> scores(WORD_SIZE);
|
469
|
+
int score = block.score;
|
470
|
+
Word mask = HIGH_BIT_MASK;
|
471
|
+
for (int i = 0; i < WORD_SIZE - 1; i++) {
|
472
|
+
scores[i] = score;
|
473
|
+
if (block.P & mask) score--;
|
474
|
+
if (block.M & mask) score++;
|
475
|
+
mask >>= 1;
|
476
|
+
}
|
477
|
+
scores[WORD_SIZE - 1] = score;
|
478
|
+
return scores;
|
479
|
+
}
|
480
|
+
|
481
|
+
/**
|
482
|
+
* Writes values of cells in block into given array, starting with first/top cell.
|
483
|
+
* @param [in] block
|
484
|
+
* @param [out] dest Array into which cell values are written. Must have size of at least WORD_SIZE.
|
485
|
+
*/
|
486
|
+
static inline void readBlock(const Block block, int* const dest) {
|
487
|
+
int score = block.score;
|
488
|
+
Word mask = HIGH_BIT_MASK;
|
489
|
+
for (int i = 0; i < WORD_SIZE - 1; i++) {
|
490
|
+
dest[WORD_SIZE - 1 - i] = score;
|
491
|
+
if (block.P & mask) score--;
|
492
|
+
if (block.M & mask) score++;
|
493
|
+
mask >>= 1;
|
494
|
+
}
|
495
|
+
dest[0] = score;
|
496
|
+
}
|
497
|
+
|
498
|
+
/**
|
499
|
+
* Writes values of cells in block into given array, starting with last/bottom cell.
|
500
|
+
* @param [in] block
|
501
|
+
* @param [out] dest Array into which cell values are written. Must have size of at least WORD_SIZE.
|
502
|
+
*/
|
503
|
+
static inline void readBlockReverse(const Block block, int* const dest) {
|
504
|
+
int score = block.score;
|
505
|
+
Word mask = HIGH_BIT_MASK;
|
506
|
+
for (int i = 0; i < WORD_SIZE - 1; i++) {
|
507
|
+
dest[i] = score;
|
508
|
+
if (block.P & mask) score--;
|
509
|
+
if (block.M & mask) score++;
|
510
|
+
mask >>= 1;
|
511
|
+
}
|
512
|
+
dest[WORD_SIZE - 1] = score;
|
513
|
+
}
|
514
|
+
|
515
|
+
/**
|
516
|
+
* @param [in] block
|
517
|
+
* @param [in] k
|
518
|
+
* @return True if all cells in block have value larger than k, otherwise false.
|
519
|
+
*/
|
520
|
+
static inline bool allBlockCellsLarger(const Block block, const int k) {
|
521
|
+
vector<int> scores = getBlockCellValues(block);
|
522
|
+
for (int i = 0; i < WORD_SIZE; i++) {
|
523
|
+
if (scores[i] <= k) return false;
|
524
|
+
}
|
525
|
+
return true;
|
526
|
+
}
|
527
|
+
|
528
|
+
|
529
|
+
/**
|
530
|
+
* Uses Myers' bit-vector algorithm to find edit distance for one of semi-global alignment methods.
|
531
|
+
* @param [in] Peq Query profile.
|
532
|
+
* @param [in] W Size of padding in last block.
|
533
|
+
* TODO: Calculate this directly from query, instead of passing it.
|
534
|
+
* @param [in] maxNumBlocks Number of blocks needed to cover the whole query.
|
535
|
+
* TODO: Calculate this directly from query, instead of passing it.
|
536
|
+
* @param [in] queryLength
|
537
|
+
* @param [in] target
|
538
|
+
* @param [in] targetLength
|
539
|
+
* @param [in] k
|
540
|
+
* @param [in] mode EDLIB_MODE_HW or EDLIB_MODE_SHW
|
541
|
+
* @param [out] bestScore_ Edit distance.
|
542
|
+
* @param [out] positions_ Array of 0-indexed positions in target at which best score was found.
|
543
|
+
Make sure to free this array with free().
|
544
|
+
* @param [out] numPositions_ Number of positions in the positions_ array.
|
545
|
+
* @return Status.
|
546
|
+
*/
|
547
|
+
static int myersCalcEditDistanceSemiGlobal(
|
548
|
+
const Word* const Peq, const int W, const int maxNumBlocks,
|
549
|
+
const int queryLength,
|
550
|
+
const unsigned char* const target, const int targetLength,
|
551
|
+
int k, const EdlibAlignMode mode,
|
552
|
+
int* const bestScore_, int** const positions_, int* const numPositions_) {
|
553
|
+
*positions_ = NULL;
|
554
|
+
*numPositions_ = 0;
|
555
|
+
|
556
|
+
// firstBlock is 0-based index of first block in Ukkonen band.
|
557
|
+
// lastBlock is 0-based index of last block in Ukkonen band.
|
558
|
+
int firstBlock = 0;
|
559
|
+
int lastBlock = min(ceilDiv(k + 1, WORD_SIZE), maxNumBlocks) - 1; // y in Myers
|
560
|
+
Block *bl; // Current block
|
561
|
+
|
562
|
+
Block* blocks = new Block[maxNumBlocks];
|
563
|
+
|
564
|
+
// For HW, solution will never be larger then queryLength.
|
565
|
+
if (mode == EDLIB_MODE_HW) {
|
566
|
+
k = min(queryLength, k);
|
567
|
+
}
|
568
|
+
|
569
|
+
// Each STRONG_REDUCE_NUM column is reduced in more expensive way.
|
570
|
+
// This gives speed up of about 2 times for small k.
|
571
|
+
const int STRONG_REDUCE_NUM = 2048;
|
572
|
+
|
573
|
+
// Initialize P, M and score
|
574
|
+
bl = blocks;
|
575
|
+
for (int b = 0; b <= lastBlock; b++) {
|
576
|
+
bl->score = (b + 1) * WORD_SIZE;
|
577
|
+
bl->P = static_cast<Word>(-1); // All 1s
|
578
|
+
bl->M = static_cast<Word>(0);
|
579
|
+
bl++;
|
580
|
+
}
|
581
|
+
|
582
|
+
int bestScore = -1;
|
583
|
+
vector<int> positions; // TODO: Maybe put this on heap?
|
584
|
+
const int startHout = mode == EDLIB_MODE_HW ? 0 : 1; // If 0 then gap before query is not penalized;
|
585
|
+
const unsigned char* targetChar = target;
|
586
|
+
for (int c = 0; c < targetLength; c++) { // for each column
|
587
|
+
const Word* Peq_c = Peq + (*targetChar) * maxNumBlocks;
|
588
|
+
|
589
|
+
//----------------------- Calculate column -------------------------//
|
590
|
+
int hout = startHout;
|
591
|
+
bl = blocks + firstBlock;
|
592
|
+
Peq_c += firstBlock;
|
593
|
+
for (int b = firstBlock; b <= lastBlock; b++) {
|
594
|
+
hout = calculateBlock(bl->P, bl->M, *Peq_c, hout, bl->P, bl->M);
|
595
|
+
bl->score += hout;
|
596
|
+
bl++; Peq_c++;
|
597
|
+
}
|
598
|
+
bl--; Peq_c--;
|
599
|
+
//------------------------------------------------------------------//
|
600
|
+
|
601
|
+
//---------- Adjust number of blocks according to Ukkonen ----------//
|
602
|
+
if ((lastBlock < maxNumBlocks - 1) && (bl->score - hout <= k) // bl is pointing to last block
|
603
|
+
&& ((*(Peq_c + 1) & WORD_1) || hout < 0)) { // Peq_c is pointing to last block
|
604
|
+
// If score of left block is not too big, calculate one more block
|
605
|
+
lastBlock++; bl++; Peq_c++;
|
606
|
+
bl->P = static_cast<Word>(-1); // All 1s
|
607
|
+
bl->M = static_cast<Word>(0);
|
608
|
+
bl->score = (bl - 1)->score - hout + WORD_SIZE + calculateBlock(bl->P, bl->M, *Peq_c, hout, bl->P, bl->M);
|
609
|
+
} else {
|
610
|
+
while (lastBlock >= firstBlock && bl->score >= k + WORD_SIZE) {
|
611
|
+
lastBlock--; bl--; Peq_c--;
|
612
|
+
}
|
613
|
+
}
|
614
|
+
|
615
|
+
// Every some columns, do some expensive but also more efficient block reducing.
|
616
|
+
// This is important!
|
617
|
+
//
|
618
|
+
// Reduce the band by decreasing last block if possible.
|
619
|
+
if (c % STRONG_REDUCE_NUM == 0) {
|
620
|
+
while (lastBlock >= 0 && lastBlock >= firstBlock && allBlockCellsLarger(*bl, k)) {
|
621
|
+
lastBlock--; bl--; Peq_c--;
|
622
|
+
}
|
623
|
+
}
|
624
|
+
// For HW, even if all cells are > k, there still may be solution in next
|
625
|
+
// column because starting conditions at upper boundary are 0.
|
626
|
+
// That means that first block is always candidate for solution,
|
627
|
+
// and we can never end calculation before last column.
|
628
|
+
if (mode == EDLIB_MODE_HW && lastBlock == -1) {
|
629
|
+
lastBlock++; bl++; Peq_c++;
|
630
|
+
}
|
631
|
+
|
632
|
+
// Reduce band by increasing first block if possible. Not applicable to HW.
|
633
|
+
if (mode != EDLIB_MODE_HW) {
|
634
|
+
while (firstBlock <= lastBlock && blocks[firstBlock].score >= k + WORD_SIZE) {
|
635
|
+
firstBlock++;
|
636
|
+
}
|
637
|
+
if (c % STRONG_REDUCE_NUM == 0) { // Do strong reduction every some blocks
|
638
|
+
while (firstBlock <= lastBlock && allBlockCellsLarger(blocks[firstBlock], k)) {
|
639
|
+
firstBlock++;
|
640
|
+
}
|
641
|
+
}
|
642
|
+
}
|
643
|
+
|
644
|
+
// If band stops to exist finish
|
645
|
+
if (lastBlock < firstBlock) {
|
646
|
+
*bestScore_ = bestScore;
|
647
|
+
if (bestScore != -1) {
|
648
|
+
*positions_ = static_cast<int *>(malloc(sizeof(int) * static_cast<int>(positions.size())));
|
649
|
+
*numPositions_ = static_cast<int>(positions.size());
|
650
|
+
copy(positions.begin(), positions.end(), *positions_);
|
651
|
+
}
|
652
|
+
delete[] blocks;
|
653
|
+
return EDLIB_STATUS_OK;
|
654
|
+
}
|
655
|
+
//------------------------------------------------------------------//
|
656
|
+
|
657
|
+
//------------------------- Update best score ----------------------//
|
658
|
+
if (lastBlock == maxNumBlocks - 1) {
|
659
|
+
int colScore = bl->score;
|
660
|
+
if (colScore <= k) { // Scores > k dont have correct values (so we cannot use them), but are certainly > k.
|
661
|
+
// NOTE: Score that I find in column c is actually score from column c-W
|
662
|
+
if (bestScore == -1 || colScore <= bestScore) {
|
663
|
+
if (colScore != bestScore) {
|
664
|
+
positions.clear();
|
665
|
+
bestScore = colScore;
|
666
|
+
// Change k so we will look only for equal or better
|
667
|
+
// scores then the best found so far.
|
668
|
+
k = bestScore;
|
669
|
+
}
|
670
|
+
positions.push_back(c - W);
|
671
|
+
}
|
672
|
+
}
|
673
|
+
}
|
674
|
+
//------------------------------------------------------------------//
|
675
|
+
|
676
|
+
targetChar++;
|
677
|
+
}
|
678
|
+
|
679
|
+
|
680
|
+
// Obtain results for last W columns from last column.
|
681
|
+
if (lastBlock == maxNumBlocks - 1) {
|
682
|
+
vector<int> blockScores = getBlockCellValues(*bl);
|
683
|
+
for (int i = 0; i < W; i++) {
|
684
|
+
int colScore = blockScores[i + 1];
|
685
|
+
if (colScore <= k && (bestScore == -1 || colScore <= bestScore)) {
|
686
|
+
if (colScore != bestScore) {
|
687
|
+
positions.clear();
|
688
|
+
k = bestScore = colScore;
|
689
|
+
}
|
690
|
+
positions.push_back(targetLength - W + i);
|
691
|
+
}
|
692
|
+
}
|
693
|
+
}
|
694
|
+
|
695
|
+
*bestScore_ = bestScore;
|
696
|
+
if (bestScore != -1) {
|
697
|
+
*positions_ = static_cast<int *>(malloc(sizeof(int) * static_cast<int>(positions.size())));
|
698
|
+
*numPositions_ = static_cast<int>(positions.size());
|
699
|
+
copy(positions.begin(), positions.end(), *positions_);
|
700
|
+
}
|
701
|
+
|
702
|
+
delete[] blocks;
|
703
|
+
return EDLIB_STATUS_OK;
|
704
|
+
}
|
705
|
+
|
706
|
+
|
707
|
+
/**
|
708
|
+
* Uses Myers' bit-vector algorithm to find edit distance for global(NW) alignment method.
|
709
|
+
* @param [in] Peq Query profile.
|
710
|
+
* @param [in] W Size of padding in last block.
|
711
|
+
* TODO: Calculate this directly from query, instead of passing it.
|
712
|
+
* @param [in] maxNumBlocks Number of blocks needed to cover the whole query.
|
713
|
+
* TODO: Calculate this directly from query, instead of passing it.
|
714
|
+
* @param [in] queryLength
|
715
|
+
* @param [in] target
|
716
|
+
* @param [in] targetLength
|
717
|
+
* @param [in] k
|
718
|
+
* @param [out] bestScore_ Edit distance.
|
719
|
+
* @param [out] position_ 0-indexed position in target at which best score was found.
|
720
|
+
* @param [in] findAlignment If true, whole matrix is remembered and alignment data is returned.
|
721
|
+
* Quadratic amount of memory is consumed.
|
722
|
+
* @param [out] alignData Data needed for alignment traceback (for reconstruction of alignment).
|
723
|
+
* Set only if findAlignment is set to true, otherwise it is NULL.
|
724
|
+
* Make sure to free this array using delete[].
|
725
|
+
* @param [out] targetStopPosition If set to -1, whole calculation is performed normally, as expected.
|
726
|
+
* If set to p, calculation is performed up to position p in target (inclusive)
|
727
|
+
* and column p is returned as the only column in alignData.
|
728
|
+
* @return Status.
|
729
|
+
*/
|
730
|
+
static int myersCalcEditDistanceNW(const Word* const Peq, const int W, const int maxNumBlocks,
|
731
|
+
const int queryLength,
|
732
|
+
const unsigned char* const target, const int targetLength,
|
733
|
+
int k, int* const bestScore_,
|
734
|
+
int* const position_, const bool findAlignment,
|
735
|
+
AlignmentData** const alignData, const int targetStopPosition) {
|
736
|
+
if (targetStopPosition > -1 && findAlignment) {
|
737
|
+
// They can not be both set at the same time!
|
738
|
+
return EDLIB_STATUS_ERROR;
|
739
|
+
}
|
740
|
+
|
741
|
+
// Each STRONG_REDUCE_NUM column is reduced in more expensive way.
|
742
|
+
const int STRONG_REDUCE_NUM = 2048; // TODO: Choose this number dinamically (based on query and target lengths?), so it does not affect speed of computation
|
743
|
+
|
744
|
+
if (k < abs(targetLength - queryLength)) {
|
745
|
+
*bestScore_ = *position_ = -1;
|
746
|
+
return EDLIB_STATUS_OK;
|
747
|
+
}
|
748
|
+
|
749
|
+
k = min(k, max(queryLength, targetLength)); // Upper bound for k
|
750
|
+
|
751
|
+
// firstBlock is 0-based index of first block in Ukkonen band.
|
752
|
+
// lastBlock is 0-based index of last block in Ukkonen band.
|
753
|
+
int firstBlock = 0;
|
754
|
+
// This is optimal now, by my formula.
|
755
|
+
int lastBlock = min(maxNumBlocks, ceilDiv(min(k, (k + queryLength - targetLength) / 2) + 1, WORD_SIZE)) - 1;
|
756
|
+
Block* bl; // Current block
|
757
|
+
|
758
|
+
Block* blocks = new Block[maxNumBlocks];
|
759
|
+
|
760
|
+
// Initialize P, M and score
|
761
|
+
bl = blocks;
|
762
|
+
for (int b = 0; b <= lastBlock; b++) {
|
763
|
+
bl->score = (b + 1) * WORD_SIZE;
|
764
|
+
bl->P = static_cast<Word>(-1); // All 1s
|
765
|
+
bl->M = static_cast<Word>(0);
|
766
|
+
bl++;
|
767
|
+
}
|
768
|
+
|
769
|
+
// If we want to find alignment, we have to store needed data.
|
770
|
+
if (findAlignment)
|
771
|
+
*alignData = new AlignmentData(maxNumBlocks, targetLength);
|
772
|
+
else if (targetStopPosition > -1)
|
773
|
+
*alignData = new AlignmentData(maxNumBlocks, 1);
|
774
|
+
else
|
775
|
+
*alignData = NULL;
|
776
|
+
|
777
|
+
const unsigned char* targetChar = target;
|
778
|
+
for (int c = 0; c < targetLength; c++) { // for each column
|
779
|
+
const Word* Peq_c = Peq + *targetChar * maxNumBlocks;
|
780
|
+
|
781
|
+
//----------------------- Calculate column -------------------------//
|
782
|
+
int hout = 1;
|
783
|
+
bl = blocks + firstBlock;
|
784
|
+
for (int b = firstBlock; b <= lastBlock; b++) {
|
785
|
+
hout = calculateBlock(bl->P, bl->M, Peq_c[b], hout, bl->P, bl->M);
|
786
|
+
bl->score += hout;
|
787
|
+
bl++;
|
788
|
+
}
|
789
|
+
bl--;
|
790
|
+
//------------------------------------------------------------------//
|
791
|
+
// bl now points to last block
|
792
|
+
|
793
|
+
// Update k. I do it only on end of column because it would slow calculation too much otherwise.
|
794
|
+
// NOTICE: I add W when in last block because it is actually result from W cells to the left and W cells up.
|
795
|
+
k = min(k, bl->score
|
796
|
+
+ max(targetLength - c - 1, queryLength - ((1 + lastBlock) * WORD_SIZE - 1) - 1)
|
797
|
+
+ (lastBlock == maxNumBlocks - 1 ? W : 0));
|
798
|
+
|
799
|
+
//---------- Adjust number of blocks according to Ukkonen ----------//
|
800
|
+
//--- Adjust last block ---//
|
801
|
+
// If block is not beneath band, calculate next block. Only next because others are certainly beneath band.
|
802
|
+
if (lastBlock + 1 < maxNumBlocks
|
803
|
+
&& !(//score[lastBlock] >= k + WORD_SIZE || // NOTICE: this condition could be satisfied if above block also!
|
804
|
+
((lastBlock + 1) * WORD_SIZE - 1
|
805
|
+
> k - bl->score + 2 * WORD_SIZE - 2 - targetLength + c + queryLength))) {
|
806
|
+
lastBlock++; bl++;
|
807
|
+
bl->P = static_cast<Word>(-1); // All 1s
|
808
|
+
bl->M = static_cast<Word>(0);
|
809
|
+
int newHout = calculateBlock(bl->P, bl->M, Peq_c[lastBlock], hout, bl->P, bl->M);
|
810
|
+
bl->score = (bl - 1)->score - hout + WORD_SIZE + newHout;
|
811
|
+
hout = newHout;
|
812
|
+
}
|
813
|
+
|
814
|
+
// While block is out of band, move one block up.
|
815
|
+
// NOTE: Condition used here is more loose than the one from the article, since I simplified the max() part of it.
|
816
|
+
// I could consider adding that max part, for optimal performance.
|
817
|
+
while (lastBlock >= firstBlock
|
818
|
+
&& (bl->score >= k + WORD_SIZE
|
819
|
+
|| ((lastBlock + 1) * WORD_SIZE - 1 >
|
820
|
+
// TODO: Does not work if do not put +1! Why???
|
821
|
+
k - bl->score + 2 * WORD_SIZE - 2 - targetLength + c + queryLength + 1))) {
|
822
|
+
lastBlock--; bl--;
|
823
|
+
}
|
824
|
+
//-------------------------//
|
825
|
+
|
826
|
+
//--- Adjust first block ---//
|
827
|
+
// While outside of band, advance block
|
828
|
+
while (firstBlock <= lastBlock
|
829
|
+
&& (blocks[firstBlock].score >= k + WORD_SIZE
|
830
|
+
|| ((firstBlock + 1) * WORD_SIZE - 1 <
|
831
|
+
blocks[firstBlock].score - k - targetLength + queryLength + c))) {
|
832
|
+
firstBlock++;
|
833
|
+
}
|
834
|
+
//--------------------------/
|
835
|
+
|
836
|
+
|
837
|
+
// TODO: consider if this part is useful, it does not seem to help much
|
838
|
+
if (c % STRONG_REDUCE_NUM == 0) { // Every some columns do more expensive but more efficient reduction
|
839
|
+
while (lastBlock >= firstBlock) {
|
840
|
+
// If all cells outside of band, remove block
|
841
|
+
vector<int> scores = getBlockCellValues(*bl);
|
842
|
+
int numCells = lastBlock == maxNumBlocks - 1 ? WORD_SIZE - W : WORD_SIZE;
|
843
|
+
int r = lastBlock * WORD_SIZE + numCells - 1;
|
844
|
+
bool reduce = true;
|
845
|
+
for (int i = WORD_SIZE - numCells; i < WORD_SIZE; i++) {
|
846
|
+
// TODO: Does not work if do not put +1! Why???
|
847
|
+
if (scores[i] <= k && r <= k - scores[i] - targetLength + c + queryLength + 1) {
|
848
|
+
reduce = false;
|
849
|
+
break;
|
850
|
+
}
|
851
|
+
r--;
|
852
|
+
}
|
853
|
+
if (!reduce) break;
|
854
|
+
lastBlock--; bl--;
|
855
|
+
}
|
856
|
+
|
857
|
+
while (firstBlock <= lastBlock) {
|
858
|
+
// If all cells outside of band, remove block
|
859
|
+
vector<int> scores = getBlockCellValues(blocks[firstBlock]);
|
860
|
+
int numCells = firstBlock == maxNumBlocks - 1 ? WORD_SIZE - W : WORD_SIZE;
|
861
|
+
int r = firstBlock * WORD_SIZE + numCells - 1;
|
862
|
+
bool reduce = true;
|
863
|
+
for (int i = WORD_SIZE - numCells; i < WORD_SIZE; i++) {
|
864
|
+
if (scores[i] <= k && r >= scores[i] - k - targetLength + c + queryLength) {
|
865
|
+
reduce = false;
|
866
|
+
break;
|
867
|
+
}
|
868
|
+
r--;
|
869
|
+
}
|
870
|
+
if (!reduce) break;
|
871
|
+
firstBlock++;
|
872
|
+
}
|
873
|
+
}
|
874
|
+
|
875
|
+
|
876
|
+
// If band stops to exist finish
|
877
|
+
if (lastBlock < firstBlock) {
|
878
|
+
*bestScore_ = *position_ = -1;
|
879
|
+
delete[] blocks;
|
880
|
+
return EDLIB_STATUS_OK;
|
881
|
+
}
|
882
|
+
//------------------------------------------------------------------//
|
883
|
+
|
884
|
+
|
885
|
+
//---- Save column so it can be used for reconstruction ----//
|
886
|
+
if (findAlignment && c < targetLength) {
|
887
|
+
bl = blocks + firstBlock;
|
888
|
+
for (int b = firstBlock; b <= lastBlock; b++) {
|
889
|
+
(*alignData)->Ps[maxNumBlocks * c + b] = bl->P;
|
890
|
+
(*alignData)->Ms[maxNumBlocks * c + b] = bl->M;
|
891
|
+
(*alignData)->scores[maxNumBlocks * c + b] = bl->score;
|
892
|
+
(*alignData)->firstBlocks[c] = firstBlock;
|
893
|
+
(*alignData)->lastBlocks[c] = lastBlock;
|
894
|
+
bl++;
|
895
|
+
}
|
896
|
+
}
|
897
|
+
//----------------------------------------------------------//
|
898
|
+
//---- If this is stop column, save it and finish ----//
|
899
|
+
if (c == targetStopPosition) {
|
900
|
+
for (int b = firstBlock; b <= lastBlock; b++) {
|
901
|
+
(*alignData)->Ps[b] = (blocks + b)->P;
|
902
|
+
(*alignData)->Ms[b] = (blocks + b)->M;
|
903
|
+
(*alignData)->scores[b] = (blocks + b)->score;
|
904
|
+
(*alignData)->firstBlocks[0] = firstBlock;
|
905
|
+
(*alignData)->lastBlocks[0] = lastBlock;
|
906
|
+
}
|
907
|
+
*bestScore_ = -1;
|
908
|
+
*position_ = targetStopPosition;
|
909
|
+
delete[] blocks;
|
910
|
+
return EDLIB_STATUS_OK;
|
911
|
+
}
|
912
|
+
//----------------------------------------------------//
|
913
|
+
|
914
|
+
targetChar++;
|
915
|
+
}
|
916
|
+
|
917
|
+
if (lastBlock == maxNumBlocks - 1) { // If last block of last column was calculated
|
918
|
+
// Obtain best score from block -> it is complicated because query is padded with W cells
|
919
|
+
int bestScore = getBlockCellValues(blocks[lastBlock])[W];
|
920
|
+
if (bestScore <= k) {
|
921
|
+
*bestScore_ = bestScore;
|
922
|
+
*position_ = targetLength - 1;
|
923
|
+
delete[] blocks;
|
924
|
+
return EDLIB_STATUS_OK;
|
925
|
+
}
|
926
|
+
}
|
927
|
+
|
928
|
+
*bestScore_ = *position_ = -1;
|
929
|
+
delete[] blocks;
|
930
|
+
return EDLIB_STATUS_OK;
|
931
|
+
}
|
932
|
+
|
933
|
+
|
934
|
+
/**
|
935
|
+
* Finds one possible alignment that gives optimal score by moving back through the dynamic programming matrix,
|
936
|
+
* that is stored in alignData. Consumes large amount of memory: O(queryLength * targetLength).
|
937
|
+
* @param [in] queryLength Normal length, without W.
|
938
|
+
* @param [in] targetLength Normal length, without W.
|
939
|
+
* @param [in] bestScore Best score.
|
940
|
+
* @param [in] alignData Data obtained during finding best score that is useful for finding alignment.
|
941
|
+
* @param [out] alignment Alignment.
|
942
|
+
* @param [out] alignmentLength Length of alignment.
|
943
|
+
* @return Status code.
|
944
|
+
*/
|
945
|
+
static int obtainAlignmentTraceback(const int queryLength, const int targetLength,
|
946
|
+
const int bestScore, const AlignmentData* const alignData,
|
947
|
+
unsigned char** const alignment, int* const alignmentLength) {
|
948
|
+
const int maxNumBlocks = ceilDiv(queryLength, WORD_SIZE);
|
949
|
+
const int W = maxNumBlocks * WORD_SIZE - queryLength;
|
950
|
+
|
951
|
+
*alignment = static_cast<unsigned char*>(malloc((queryLength + targetLength - 1) * sizeof(unsigned char)));
|
952
|
+
*alignmentLength = 0;
|
953
|
+
int c = targetLength - 1; // index of column
|
954
|
+
int b = maxNumBlocks - 1; // index of block in column
|
955
|
+
int currScore = bestScore; // Score of current cell
|
956
|
+
int lScore = -1; // Score of left cell
|
957
|
+
int uScore = -1; // Score of upper cell
|
958
|
+
int ulScore = -1; // Score of upper left cell
|
959
|
+
Word currP = alignData->Ps[c * maxNumBlocks + b]; // P of current block
|
960
|
+
Word currM = alignData->Ms[c * maxNumBlocks + b]; // M of current block
|
961
|
+
// True if block to left exists and is in band
|
962
|
+
bool thereIsLeftBlock = c > 0 && b >= alignData->firstBlocks[c-1] && b <= alignData->lastBlocks[c-1];
|
963
|
+
// We set initial values of lP and lM to 0 only to avoid compiler warnings, they should not affect the
|
964
|
+
// calculation as both lP and lM should be initialized at some moment later (but compiler can not
|
965
|
+
// detect it since this initialization is guaranteed by "business" logic).
|
966
|
+
Word lP = 0, lM = 0;
|
967
|
+
if (thereIsLeftBlock) {
|
968
|
+
lP = alignData->Ps[(c - 1) * maxNumBlocks + b]; // P of block to the left
|
969
|
+
lM = alignData->Ms[(c - 1) * maxNumBlocks + b]; // M of block to the left
|
970
|
+
}
|
971
|
+
currP <<= W;
|
972
|
+
currM <<= W;
|
973
|
+
int blockPos = WORD_SIZE - W - 1; // 0 based index of current cell in blockPos
|
974
|
+
|
975
|
+
// TODO(martin): refactor this whole piece of code. There are too many if-else statements,
|
976
|
+
// it is too easy for a bug to hide and to hard to effectively cover all the edge-cases.
|
977
|
+
// We need better separation of logic and responsibilities.
|
978
|
+
while (true) {
|
979
|
+
if (c == 0) {
|
980
|
+
thereIsLeftBlock = true;
|
981
|
+
lScore = b * WORD_SIZE + blockPos + 1;
|
982
|
+
ulScore = lScore - 1;
|
983
|
+
}
|
984
|
+
|
985
|
+
// TODO: improvement: calculate only those cells that are needed,
|
986
|
+
// for example if I calculate upper cell and can move up,
|
987
|
+
// there is no need to calculate left and upper left cell
|
988
|
+
//---------- Calculate scores ---------//
|
989
|
+
if (lScore == -1 && thereIsLeftBlock) {
|
990
|
+
lScore = alignData->scores[(c - 1) * maxNumBlocks + b]; // score of block to the left
|
991
|
+
for (int i = 0; i < WORD_SIZE - blockPos - 1; i++) {
|
992
|
+
if (lP & HIGH_BIT_MASK) lScore--;
|
993
|
+
if (lM & HIGH_BIT_MASK) lScore++;
|
994
|
+
lP <<= 1;
|
995
|
+
lM <<= 1;
|
996
|
+
}
|
997
|
+
}
|
998
|
+
if (ulScore == -1) {
|
999
|
+
if (lScore != -1) {
|
1000
|
+
ulScore = lScore;
|
1001
|
+
if (lP & HIGH_BIT_MASK) ulScore--;
|
1002
|
+
if (lM & HIGH_BIT_MASK) ulScore++;
|
1003
|
+
}
|
1004
|
+
else if (c > 0 && b-1 >= alignData->firstBlocks[c-1] && b-1 <= alignData->lastBlocks[c-1]) {
|
1005
|
+
// This is the case when upper left cell is last cell in block,
|
1006
|
+
// and block to left is not in band so lScore is -1.
|
1007
|
+
ulScore = alignData->scores[(c - 1) * maxNumBlocks + b - 1];
|
1008
|
+
}
|
1009
|
+
}
|
1010
|
+
if (uScore == -1) {
|
1011
|
+
uScore = currScore;
|
1012
|
+
if (currP & HIGH_BIT_MASK) uScore--;
|
1013
|
+
if (currM & HIGH_BIT_MASK) uScore++;
|
1014
|
+
currP <<= 1;
|
1015
|
+
currM <<= 1;
|
1016
|
+
}
|
1017
|
+
//-------------------------------------//
|
1018
|
+
|
1019
|
+
// TODO: should I check if there is upper block?
|
1020
|
+
|
1021
|
+
//-------------- Move --------------//
|
1022
|
+
// Move up - insertion to target - deletion from query
|
1023
|
+
if (uScore != -1 && uScore + 1 == currScore) {
|
1024
|
+
currScore = uScore;
|
1025
|
+
lScore = ulScore;
|
1026
|
+
uScore = ulScore = -1;
|
1027
|
+
if (blockPos == 0) { // If entering new (upper) block
|
1028
|
+
if (b == 0) { // If there are no cells above (only boundary cells)
|
1029
|
+
(*alignment)[(*alignmentLength)++] = EDLIB_EDOP_INSERT; // Move up
|
1030
|
+
for (int i = 0; i < c + 1; i++) // Move left until end
|
1031
|
+
(*alignment)[(*alignmentLength)++] = EDLIB_EDOP_DELETE;
|
1032
|
+
break;
|
1033
|
+
} else {
|
1034
|
+
blockPos = WORD_SIZE - 1;
|
1035
|
+
b--;
|
1036
|
+
currP = alignData->Ps[c * maxNumBlocks + b];
|
1037
|
+
currM = alignData->Ms[c * maxNumBlocks + b];
|
1038
|
+
if (c > 0 && b >= alignData->firstBlocks[c-1] && b <= alignData->lastBlocks[c-1]) {
|
1039
|
+
thereIsLeftBlock = true;
|
1040
|
+
lP = alignData->Ps[(c - 1) * maxNumBlocks + b]; // TODO: improve this, too many operations
|
1041
|
+
lM = alignData->Ms[(c - 1) * maxNumBlocks + b];
|
1042
|
+
} else {
|
1043
|
+
thereIsLeftBlock = false;
|
1044
|
+
// TODO(martin): There may not be left block, but there can be left boundary - do we
|
1045
|
+
// handle this correctly then? Are l and ul score set correctly? I should check that / refactor this.
|
1046
|
+
}
|
1047
|
+
}
|
1048
|
+
} else {
|
1049
|
+
blockPos--;
|
1050
|
+
lP <<= 1;
|
1051
|
+
lM <<= 1;
|
1052
|
+
}
|
1053
|
+
// Mark move
|
1054
|
+
(*alignment)[(*alignmentLength)++] = EDLIB_EDOP_INSERT;
|
1055
|
+
}
|
1056
|
+
// Move left - deletion from target - insertion to query
|
1057
|
+
else if (lScore != -1 && lScore + 1 == currScore) {
|
1058
|
+
currScore = lScore;
|
1059
|
+
uScore = ulScore;
|
1060
|
+
lScore = ulScore = -1;
|
1061
|
+
c--;
|
1062
|
+
if (c == -1) { // If there are no cells to the left (only boundary cells)
|
1063
|
+
(*alignment)[(*alignmentLength)++] = EDLIB_EDOP_DELETE; // Move left
|
1064
|
+
int numUp = b * WORD_SIZE + blockPos + 1;
|
1065
|
+
for (int i = 0; i < numUp; i++) // Move up until end
|
1066
|
+
(*alignment)[(*alignmentLength)++] = EDLIB_EDOP_INSERT;
|
1067
|
+
break;
|
1068
|
+
}
|
1069
|
+
currP = lP;
|
1070
|
+
currM = lM;
|
1071
|
+
if (c > 0 && b >= alignData->firstBlocks[c-1] && b <= alignData->lastBlocks[c-1]) {
|
1072
|
+
thereIsLeftBlock = true;
|
1073
|
+
lP = alignData->Ps[(c - 1) * maxNumBlocks + b];
|
1074
|
+
lM = alignData->Ms[(c - 1) * maxNumBlocks + b];
|
1075
|
+
} else {
|
1076
|
+
if (c == 0) { // If there are no cells to the left (only boundary cells)
|
1077
|
+
thereIsLeftBlock = true;
|
1078
|
+
lScore = b * WORD_SIZE + blockPos + 1;
|
1079
|
+
ulScore = lScore - 1;
|
1080
|
+
} else {
|
1081
|
+
thereIsLeftBlock = false;
|
1082
|
+
}
|
1083
|
+
}
|
1084
|
+
// Mark move
|
1085
|
+
(*alignment)[(*alignmentLength)++] = EDLIB_EDOP_DELETE;
|
1086
|
+
}
|
1087
|
+
// Move up left - (mis)match
|
1088
|
+
else if (ulScore != -1) {
|
1089
|
+
unsigned char moveCode = ulScore == currScore ? EDLIB_EDOP_MATCH : EDLIB_EDOP_MISMATCH;
|
1090
|
+
currScore = ulScore;
|
1091
|
+
uScore = lScore = ulScore = -1;
|
1092
|
+
c--;
|
1093
|
+
if (c == -1) { // If there are no cells to the left (only boundary cells)
|
1094
|
+
(*alignment)[(*alignmentLength)++] = moveCode; // Move left
|
1095
|
+
int numUp = b * WORD_SIZE + blockPos;
|
1096
|
+
for (int i = 0; i < numUp; i++) // Move up until end
|
1097
|
+
(*alignment)[(*alignmentLength)++] = EDLIB_EDOP_INSERT;
|
1098
|
+
break;
|
1099
|
+
}
|
1100
|
+
if (blockPos == 0) { // If entering upper left block
|
1101
|
+
if (b == 0) { // If there are no more cells above (only boundary cells)
|
1102
|
+
(*alignment)[(*alignmentLength)++] = moveCode; // Move up left
|
1103
|
+
for (int i = 0; i < c + 1; i++) // Move left until end
|
1104
|
+
(*alignment)[(*alignmentLength)++] = EDLIB_EDOP_DELETE;
|
1105
|
+
break;
|
1106
|
+
}
|
1107
|
+
blockPos = WORD_SIZE - 1;
|
1108
|
+
b--;
|
1109
|
+
currP = alignData->Ps[c * maxNumBlocks + b];
|
1110
|
+
currM = alignData->Ms[c * maxNumBlocks + b];
|
1111
|
+
} else { // If entering left block
|
1112
|
+
blockPos--;
|
1113
|
+
currP = lP;
|
1114
|
+
currM = lM;
|
1115
|
+
currP <<= 1;
|
1116
|
+
currM <<= 1;
|
1117
|
+
}
|
1118
|
+
// Set new left block
|
1119
|
+
if (c > 0 && b >= alignData->firstBlocks[c-1] && b <= alignData->lastBlocks[c-1]) {
|
1120
|
+
thereIsLeftBlock = true;
|
1121
|
+
lP = alignData->Ps[(c - 1) * maxNumBlocks + b];
|
1122
|
+
lM = alignData->Ms[(c - 1) * maxNumBlocks + b];
|
1123
|
+
} else {
|
1124
|
+
if (c == 0) { // If there are no cells to the left (only boundary cells)
|
1125
|
+
thereIsLeftBlock = true;
|
1126
|
+
lScore = b * WORD_SIZE + blockPos + 1;
|
1127
|
+
ulScore = lScore - 1;
|
1128
|
+
} else {
|
1129
|
+
thereIsLeftBlock = false;
|
1130
|
+
}
|
1131
|
+
}
|
1132
|
+
// Mark move
|
1133
|
+
(*alignment)[(*alignmentLength)++] = moveCode;
|
1134
|
+
} else {
|
1135
|
+
// Reached end - finished!
|
1136
|
+
break;
|
1137
|
+
}
|
1138
|
+
//----------------------------------//
|
1139
|
+
}
|
1140
|
+
|
1141
|
+
*alignment = static_cast<unsigned char*>(realloc(*alignment, (*alignmentLength) * sizeof(unsigned char)));
|
1142
|
+
reverse(*alignment, *alignment + (*alignmentLength));
|
1143
|
+
return EDLIB_STATUS_OK;
|
1144
|
+
}
|
1145
|
+
|
1146
|
+
|
1147
|
+
/**
|
1148
|
+
* Finds one possible alignment that gives optimal score (bestScore).
|
1149
|
+
* It will split problem into smaller problems using Hirschberg's algorithm and when they are small enough,
|
1150
|
+
* it will solve them using traceback algorithm.
|
1151
|
+
* @param [in] query
|
1152
|
+
* @param [in] rQuery Reversed query.
|
1153
|
+
* @param [in] queryLength
|
1154
|
+
* @param [in] target
|
1155
|
+
* @param [in] rTarget Reversed target.
|
1156
|
+
* @param [in] targetLength
|
1157
|
+
* @param [in] equalityDefinition
|
1158
|
+
* @param [in] alphabetLength
|
1159
|
+
* @param [in] bestScore Best(optimal) score.
|
1160
|
+
* @param [out] alignment Sequence of edit operations that make target equal to query.
|
1161
|
+
* @param [out] alignmentLength Length of alignment.
|
1162
|
+
* @return Status code.
|
1163
|
+
*/
|
1164
|
+
static int obtainAlignment(
|
1165
|
+
const unsigned char* const query, const unsigned char* const rQuery, const int queryLength,
|
1166
|
+
const unsigned char* const target, const unsigned char* const rTarget, const int targetLength,
|
1167
|
+
const EqualityDefinition& equalityDefinition, const int alphabetLength, const int bestScore,
|
1168
|
+
unsigned char** const alignment, int* const alignmentLength) {
|
1169
|
+
|
1170
|
+
// Handle special case when one of sequences has length of 0.
|
1171
|
+
if (queryLength == 0 || targetLength == 0) {
|
1172
|
+
*alignmentLength = targetLength + queryLength;
|
1173
|
+
*alignment = static_cast<unsigned char*>(malloc((*alignmentLength) * sizeof(unsigned char)));
|
1174
|
+
for (int i = 0; i < *alignmentLength; i++) {
|
1175
|
+
(*alignment)[i] = queryLength == 0 ? EDLIB_EDOP_DELETE : EDLIB_EDOP_INSERT;
|
1176
|
+
}
|
1177
|
+
return EDLIB_STATUS_OK;
|
1178
|
+
}
|
1179
|
+
|
1180
|
+
const int maxNumBlocks = ceilDiv(queryLength, WORD_SIZE);
|
1181
|
+
const int W = maxNumBlocks * WORD_SIZE - queryLength;
|
1182
|
+
int statusCode;
|
1183
|
+
|
1184
|
+
// TODO: think about reducing number of memory allocations in alignment functions, probably
|
1185
|
+
// by sharing some memory that is allocated only once. That refers to: Peq, columns in Hirschberg,
|
1186
|
+
// and it could also be done for alignments - we could have one big array for alignment that would be
|
1187
|
+
// sparsely populated by each of steps in recursion, and at the end we would just consolidate those results.
|
1188
|
+
|
1189
|
+
// If estimated memory consumption for traceback algorithm is smaller than 1MB use it,
|
1190
|
+
// otherwise use Hirschberg's algorithm. By running few tests I choose boundary of 1MB as optimal.
|
1191
|
+
long long alignmentDataSize = (2ll * sizeof(Word) + sizeof(int)) * maxNumBlocks * targetLength
|
1192
|
+
+ 2ll * sizeof(int) * targetLength;
|
1193
|
+
if (alignmentDataSize < 1024 * 1024) {
|
1194
|
+
int score_, endLocation_; // Used only to call function.
|
1195
|
+
AlignmentData* alignData = NULL;
|
1196
|
+
Word* Peq = buildPeq(alphabetLength, query, queryLength, equalityDefinition);
|
1197
|
+
myersCalcEditDistanceNW(Peq, W, maxNumBlocks,
|
1198
|
+
queryLength,
|
1199
|
+
target, targetLength,
|
1200
|
+
bestScore,
|
1201
|
+
&score_, &endLocation_, true, &alignData, -1);
|
1202
|
+
//assert(score_ == bestScore);
|
1203
|
+
//assert(endLocation_ == targetLength - 1);
|
1204
|
+
|
1205
|
+
statusCode = obtainAlignmentTraceback(queryLength, targetLength,
|
1206
|
+
bestScore, alignData, alignment, alignmentLength);
|
1207
|
+
delete alignData;
|
1208
|
+
delete[] Peq;
|
1209
|
+
} else {
|
1210
|
+
statusCode = obtainAlignmentHirschberg(query, rQuery, queryLength,
|
1211
|
+
target, rTarget, targetLength,
|
1212
|
+
equalityDefinition, alphabetLength, bestScore,
|
1213
|
+
alignment, alignmentLength);
|
1214
|
+
}
|
1215
|
+
return statusCode;
|
1216
|
+
}
|
1217
|
+
|
1218
|
+
|
1219
|
+
/**
|
1220
|
+
* Finds one possible alignment that gives optimal score (bestScore).
|
1221
|
+
* Uses Hirschberg's algorithm to split problem into two sub-problems, solve them and combine them together.
|
1222
|
+
* @param [in] query
|
1223
|
+
* @param [in] rQuery Reversed query.
|
1224
|
+
* @param [in] queryLength
|
1225
|
+
* @param [in] target
|
1226
|
+
* @param [in] rTarget Reversed target.
|
1227
|
+
* @param [in] targetLength
|
1228
|
+
* @param [in] alphabetLength
|
1229
|
+
* @param [in] bestScore Best(optimal) score.
|
1230
|
+
* @param [out] alignment Sequence of edit operations that make target equal to query.
|
1231
|
+
* @param [out] alignmentLength Length of alignment.
|
1232
|
+
* @return Status code.
|
1233
|
+
*/
|
1234
|
+
static int obtainAlignmentHirschberg(
|
1235
|
+
const unsigned char* const query, const unsigned char* const rQuery, const int queryLength,
|
1236
|
+
const unsigned char* const target, const unsigned char* const rTarget, const int targetLength,
|
1237
|
+
const EqualityDefinition& equalityDefinition, const int alphabetLength, const int bestScore,
|
1238
|
+
unsigned char** const alignment, int* const alignmentLength) {
|
1239
|
+
|
1240
|
+
const int maxNumBlocks = ceilDiv(queryLength, WORD_SIZE);
|
1241
|
+
const int W = maxNumBlocks * WORD_SIZE - queryLength;
|
1242
|
+
|
1243
|
+
Word* Peq = buildPeq(alphabetLength, query, queryLength, equalityDefinition);
|
1244
|
+
Word* rPeq = buildPeq(alphabetLength, rQuery, queryLength, equalityDefinition);
|
1245
|
+
|
1246
|
+
// Used only to call functions.
|
1247
|
+
int score_, endLocation_;
|
1248
|
+
|
1249
|
+
// Divide dynamic matrix into two halfs, left and right.
|
1250
|
+
const int leftHalfWidth = targetLength / 2;
|
1251
|
+
const int rightHalfWidth = targetLength - leftHalfWidth;
|
1252
|
+
|
1253
|
+
// Calculate left half.
|
1254
|
+
AlignmentData* alignDataLeftHalf = NULL;
|
1255
|
+
int leftHalfCalcStatus = myersCalcEditDistanceNW(
|
1256
|
+
Peq, W, maxNumBlocks, queryLength, target, targetLength, bestScore,
|
1257
|
+
&score_, &endLocation_, false, &alignDataLeftHalf, leftHalfWidth - 1);
|
1258
|
+
|
1259
|
+
// Calculate right half.
|
1260
|
+
AlignmentData* alignDataRightHalf = NULL;
|
1261
|
+
int rightHalfCalcStatus = myersCalcEditDistanceNW(
|
1262
|
+
rPeq, W, maxNumBlocks, queryLength, rTarget, targetLength, bestScore,
|
1263
|
+
&score_, &endLocation_, false, &alignDataRightHalf, rightHalfWidth - 1);
|
1264
|
+
|
1265
|
+
delete[] Peq;
|
1266
|
+
delete[] rPeq;
|
1267
|
+
|
1268
|
+
if (leftHalfCalcStatus == EDLIB_STATUS_ERROR || rightHalfCalcStatus == EDLIB_STATUS_ERROR) {
|
1269
|
+
if (alignDataLeftHalf) delete alignDataLeftHalf;
|
1270
|
+
if (alignDataRightHalf) delete alignDataRightHalf;
|
1271
|
+
return EDLIB_STATUS_ERROR;
|
1272
|
+
}
|
1273
|
+
|
1274
|
+
// Unwrap the left half.
|
1275
|
+
int firstBlockIdxLeft = alignDataLeftHalf->firstBlocks[0];
|
1276
|
+
int lastBlockIdxLeft = alignDataLeftHalf->lastBlocks[0];
|
1277
|
+
// TODO: avoid this allocation by using some shared array?
|
1278
|
+
// scoresLeft contains scores from left column, starting with scoresLeftStartIdx row (query index)
|
1279
|
+
// and ending with scoresLeftEndIdx row (0-indexed).
|
1280
|
+
int scoresLeftLength = (lastBlockIdxLeft - firstBlockIdxLeft + 1) * WORD_SIZE;
|
1281
|
+
int* scoresLeft = new int[scoresLeftLength];
|
1282
|
+
for (int blockIdx = firstBlockIdxLeft; blockIdx <= lastBlockIdxLeft; blockIdx++) {
|
1283
|
+
Block block(alignDataLeftHalf->Ps[blockIdx], alignDataLeftHalf->Ms[blockIdx],
|
1284
|
+
alignDataLeftHalf->scores[blockIdx]);
|
1285
|
+
readBlock(block, scoresLeft + (blockIdx - firstBlockIdxLeft) * WORD_SIZE);
|
1286
|
+
}
|
1287
|
+
int scoresLeftStartIdx = firstBlockIdxLeft * WORD_SIZE;
|
1288
|
+
// If last block contains padding, shorten the length of scores for the length of padding.
|
1289
|
+
if (lastBlockIdxLeft == maxNumBlocks - 1) {
|
1290
|
+
scoresLeftLength -= W;
|
1291
|
+
}
|
1292
|
+
|
1293
|
+
// Unwrap the right half (I also reverse it while unwraping).
|
1294
|
+
int firstBlockIdxRight = alignDataRightHalf->firstBlocks[0];
|
1295
|
+
int lastBlockIdxRight = alignDataRightHalf->lastBlocks[0];
|
1296
|
+
int scoresRightLength = (lastBlockIdxRight - firstBlockIdxRight + 1) * WORD_SIZE;
|
1297
|
+
int* scoresRight = new int[scoresRightLength];
|
1298
|
+
int* scoresRightOriginalStart = scoresRight;
|
1299
|
+
for (int blockIdx = firstBlockIdxRight; blockIdx <= lastBlockIdxRight; blockIdx++) {
|
1300
|
+
Block block(alignDataRightHalf->Ps[blockIdx], alignDataRightHalf->Ms[blockIdx],
|
1301
|
+
alignDataRightHalf->scores[blockIdx]);
|
1302
|
+
readBlockReverse(block, scoresRight + (lastBlockIdxRight - blockIdx) * WORD_SIZE);
|
1303
|
+
}
|
1304
|
+
int scoresRightStartIdx = queryLength - (lastBlockIdxRight + 1) * WORD_SIZE;
|
1305
|
+
// If there is padding at the beginning of scoresRight (that can happen because of reversing that we do),
|
1306
|
+
// move pointer forward to remove the padding (that is why we remember originalStart).
|
1307
|
+
if (scoresRightStartIdx < 0) {
|
1308
|
+
//assert(scoresRightStartIdx == -1 * W);
|
1309
|
+
scoresRight += W;
|
1310
|
+
scoresRightStartIdx += W;
|
1311
|
+
scoresRightLength -= W;
|
1312
|
+
}
|
1313
|
+
|
1314
|
+
delete alignDataLeftHalf;
|
1315
|
+
delete alignDataRightHalf;
|
1316
|
+
|
1317
|
+
//--------------------- Find the best move ----------------//
|
1318
|
+
// Find the query/row index of cell in left column which together with its lower right neighbour
|
1319
|
+
// from right column gives the best score (when summed). We also have to consider boundary cells
|
1320
|
+
// (those cells at -1 indexes).
|
1321
|
+
// x|
|
1322
|
+
// -+-
|
1323
|
+
// |x
|
1324
|
+
int queryIdxLeftStart = max(scoresLeftStartIdx, scoresRightStartIdx - 1);
|
1325
|
+
int queryIdxLeftEnd = min(scoresLeftStartIdx + scoresLeftLength - 1,
|
1326
|
+
scoresRightStartIdx + scoresRightLength - 2);
|
1327
|
+
int leftScore = -1, rightScore = -1;
|
1328
|
+
int queryIdxLeftAlignment = -1; // Query/row index of cell in left column where alignment is passing through.
|
1329
|
+
bool queryIdxLeftAlignmentFound = false;
|
1330
|
+
for (int queryIdx = queryIdxLeftStart; queryIdx <= queryIdxLeftEnd; queryIdx++) {
|
1331
|
+
leftScore = scoresLeft[queryIdx - scoresLeftStartIdx];
|
1332
|
+
rightScore = scoresRight[queryIdx + 1 - scoresRightStartIdx];
|
1333
|
+
if (leftScore + rightScore == bestScore) {
|
1334
|
+
queryIdxLeftAlignment = queryIdx;
|
1335
|
+
queryIdxLeftAlignmentFound = true;
|
1336
|
+
break;
|
1337
|
+
}
|
1338
|
+
}
|
1339
|
+
// Check boundary cells.
|
1340
|
+
if (!queryIdxLeftAlignmentFound && scoresLeftStartIdx == 0 && scoresRightStartIdx == 0) {
|
1341
|
+
leftScore = leftHalfWidth;
|
1342
|
+
rightScore = scoresRight[0];
|
1343
|
+
if (leftScore + rightScore == bestScore) {
|
1344
|
+
queryIdxLeftAlignment = -1;
|
1345
|
+
queryIdxLeftAlignmentFound = true;
|
1346
|
+
}
|
1347
|
+
}
|
1348
|
+
if (!queryIdxLeftAlignmentFound && scoresLeftStartIdx + scoresLeftLength == queryLength
|
1349
|
+
&& scoresRightStartIdx + scoresRightLength == queryLength) {
|
1350
|
+
leftScore = scoresLeft[scoresLeftLength - 1];
|
1351
|
+
rightScore = rightHalfWidth;
|
1352
|
+
if (leftScore + rightScore == bestScore) {
|
1353
|
+
queryIdxLeftAlignment = queryLength - 1;
|
1354
|
+
queryIdxLeftAlignmentFound = true;
|
1355
|
+
}
|
1356
|
+
}
|
1357
|
+
|
1358
|
+
delete[] scoresLeft;
|
1359
|
+
delete[] scoresRightOriginalStart;
|
1360
|
+
|
1361
|
+
if (queryIdxLeftAlignmentFound == false) {
|
1362
|
+
// If there was no move that is part of optimal alignment, then there is no such alignment
|
1363
|
+
// or given bestScore is not correct!
|
1364
|
+
return EDLIB_STATUS_ERROR;
|
1365
|
+
}
|
1366
|
+
//----------------------------------------------------------//
|
1367
|
+
|
1368
|
+
// Calculate alignments for upper half of left half (upper left - ul)
|
1369
|
+
// and lower half of right half (lower right - lr).
|
1370
|
+
const int ulHeight = queryIdxLeftAlignment + 1;
|
1371
|
+
const int lrHeight = queryLength - ulHeight;
|
1372
|
+
const int ulWidth = leftHalfWidth;
|
1373
|
+
const int lrWidth = rightHalfWidth;
|
1374
|
+
unsigned char* ulAlignment = NULL; int ulAlignmentLength;
|
1375
|
+
int ulStatusCode = obtainAlignment(query, rQuery + lrHeight, ulHeight,
|
1376
|
+
target, rTarget + lrWidth, ulWidth,
|
1377
|
+
equalityDefinition, alphabetLength, leftScore,
|
1378
|
+
&ulAlignment, &ulAlignmentLength);
|
1379
|
+
unsigned char* lrAlignment = NULL; int lrAlignmentLength;
|
1380
|
+
int lrStatusCode = obtainAlignment(query + ulHeight, rQuery, lrHeight,
|
1381
|
+
target + ulWidth, rTarget, lrWidth,
|
1382
|
+
equalityDefinition, alphabetLength, rightScore,
|
1383
|
+
&lrAlignment, &lrAlignmentLength);
|
1384
|
+
if (ulStatusCode == EDLIB_STATUS_ERROR || lrStatusCode == EDLIB_STATUS_ERROR) {
|
1385
|
+
if (ulAlignment) free(ulAlignment);
|
1386
|
+
if (lrAlignment) free(lrAlignment);
|
1387
|
+
return EDLIB_STATUS_ERROR;
|
1388
|
+
}
|
1389
|
+
|
1390
|
+
// Build alignment by concatenating upper left alignment with lower right alignment.
|
1391
|
+
*alignmentLength = ulAlignmentLength + lrAlignmentLength;
|
1392
|
+
*alignment = static_cast<unsigned char*>(malloc((*alignmentLength) * sizeof(unsigned char)));
|
1393
|
+
memcpy(*alignment, ulAlignment, ulAlignmentLength);
|
1394
|
+
memcpy(*alignment + ulAlignmentLength, lrAlignment, lrAlignmentLength);
|
1395
|
+
|
1396
|
+
free(ulAlignment);
|
1397
|
+
free(lrAlignment);
|
1398
|
+
return EDLIB_STATUS_OK;
|
1399
|
+
}
|
1400
|
+
|
1401
|
+
|
1402
|
+
/**
|
1403
|
+
* Takes char query and char target, recognizes alphabet and transforms them into unsigned char sequences
|
1404
|
+
* where elements in sequences are not any more letters of alphabet, but their index in alphabet.
|
1405
|
+
* Most of internal edlib functions expect such transformed sequences.
|
1406
|
+
* This function will allocate queryTransformed and targetTransformed, so make sure to free them when done.
|
1407
|
+
* Example:
|
1408
|
+
* Original sequences: "ACT" and "CGT".
|
1409
|
+
* Alphabet would be recognized as "ACTG". Alphabet length = 4.
|
1410
|
+
* Transformed sequences: [0, 1, 2] and [1, 3, 2].
|
1411
|
+
* @param [in] queryOriginal
|
1412
|
+
* @param [in] queryLength
|
1413
|
+
* @param [in] targetOriginal
|
1414
|
+
* @param [in] targetLength
|
1415
|
+
* @param [out] queryTransformed It will contain values in range [0, alphabet length - 1].
|
1416
|
+
* @param [out] targetTransformed It will contain values in range [0, alphabet length - 1].
|
1417
|
+
* @return Alphabet as a string of unique characters, where index of each character is its value in transformed
|
1418
|
+
* sequences.
|
1419
|
+
*/
|
1420
|
+
static string transformSequences(const char* const queryOriginal, const int queryLength,
|
1421
|
+
const char* const targetOriginal, const int targetLength,
|
1422
|
+
unsigned char** const queryTransformed,
|
1423
|
+
unsigned char** const targetTransformed) {
|
1424
|
+
// Alphabet is constructed from letters that are present in sequences.
|
1425
|
+
// Each letter is assigned an ordinal number, starting from 0 up to alphabetLength - 1,
|
1426
|
+
// and new query and target are created in which letters are replaced with their ordinal numbers.
|
1427
|
+
// This query and target are used in all the calculations later.
|
1428
|
+
*queryTransformed = static_cast<unsigned char *>(malloc(sizeof(unsigned char) * queryLength));
|
1429
|
+
*targetTransformed = static_cast<unsigned char *>(malloc(sizeof(unsigned char) * targetLength));
|
1430
|
+
|
1431
|
+
string alphabet = "";
|
1432
|
+
|
1433
|
+
// Alphabet information, it is constructed on fly while transforming sequences.
|
1434
|
+
// letterIdx[c] is index of letter c in alphabet.
|
1435
|
+
unsigned char letterIdx[MAX_UCHAR + 1];
|
1436
|
+
bool inAlphabet[MAX_UCHAR + 1]; // inAlphabet[c] is true if c is in alphabet
|
1437
|
+
for (int i = 0; i < MAX_UCHAR + 1; i++) inAlphabet[i] = false;
|
1438
|
+
|
1439
|
+
for (int i = 0; i < queryLength; i++) {
|
1440
|
+
unsigned char c = static_cast<unsigned char>(queryOriginal[i]);
|
1441
|
+
if (!inAlphabet[c]) {
|
1442
|
+
inAlphabet[c] = true;
|
1443
|
+
letterIdx[c] = static_cast<unsigned char>(alphabet.size());
|
1444
|
+
alphabet += queryOriginal[i];
|
1445
|
+
}
|
1446
|
+
(*queryTransformed)[i] = letterIdx[c];
|
1447
|
+
}
|
1448
|
+
for (int i = 0; i < targetLength; i++) {
|
1449
|
+
unsigned char c = static_cast<unsigned char>(targetOriginal[i]);
|
1450
|
+
if (!inAlphabet[c]) {
|
1451
|
+
inAlphabet[c] = true;
|
1452
|
+
letterIdx[c] = static_cast<unsigned char>(alphabet.size());
|
1453
|
+
alphabet += targetOriginal[i];
|
1454
|
+
}
|
1455
|
+
(*targetTransformed)[i] = letterIdx[c];
|
1456
|
+
}
|
1457
|
+
|
1458
|
+
return alphabet;
|
1459
|
+
}
|
1460
|
+
|
1461
|
+
|
1462
|
+
extern "C" EdlibAlignConfig edlibNewAlignConfig(int k, EdlibAlignMode mode, EdlibAlignTask task,
|
1463
|
+
const EdlibEqualityPair* additionalEqualities,
|
1464
|
+
int additionalEqualitiesLength) {
|
1465
|
+
EdlibAlignConfig config;
|
1466
|
+
config.k = k;
|
1467
|
+
config.mode = mode;
|
1468
|
+
config.task = task;
|
1469
|
+
config.additionalEqualities = additionalEqualities;
|
1470
|
+
config.additionalEqualitiesLength = additionalEqualitiesLength;
|
1471
|
+
return config;
|
1472
|
+
}
|
1473
|
+
|
1474
|
+
extern "C" EdlibAlignConfig edlibDefaultAlignConfig(void) {
|
1475
|
+
return edlibNewAlignConfig(-1, EDLIB_MODE_NW, EDLIB_TASK_DISTANCE, NULL, 0);
|
1476
|
+
}
|
1477
|
+
|
1478
|
+
extern "C" void edlibFreeAlignResult(EdlibAlignResult result) {
|
1479
|
+
if (result.endLocations) free(result.endLocations);
|
1480
|
+
if (result.startLocations) free(result.startLocations);
|
1481
|
+
if (result.alignment) free(result.alignment);
|
1482
|
+
}
|