|
0
|
1 #include <stdio.h>
|
|
|
2 #include <stdlib.h>
|
|
|
3 #include <sys/time.h>
|
|
|
4 #include <inttypes.h>
|
|
|
5 #include <ctype.h>
|
|
|
6 #include <string.h>
|
|
|
7 #include <pthread.h>
|
|
|
8 #include <emmintrin.h>
|
|
|
9 #include <math.h>
|
|
|
10 #include "structs.h"
|
|
|
11
|
|
|
12 #define ALIGNED_BYTES 16
|
|
|
13
|
|
|
14 void terror(char *s) {
|
|
|
15 fprintf(stdout, "ERR**** %s ****\n", s);
|
|
|
16 exit(-1);
|
|
|
17 }
|
|
|
18
|
|
|
19 char buffered_fgetc(char *buffer, uint64_t *pos, uint64_t *read, FILE *f) {
|
|
|
20 if (*pos >= READBUF) {
|
|
|
21 *pos = 0;
|
|
|
22 memset(buffer, 0, READBUF);
|
|
|
23 *read = fread(buffer, 1, READBUF, f);
|
|
|
24 }
|
|
|
25 *pos = *pos + 1;
|
|
|
26 return buffer[*pos-1];
|
|
|
27 }
|
|
|
28
|
|
|
29 void get_num_seqs_and_length(char * seq_buffer, uint64_t * n_seqs, uint64_t * t_len, LoadingDBArgs * ldbargs){
|
|
|
30 char check[16] = ">>>>>>>>>>>>>>>>";
|
|
|
31 uint64_t a_fourth = *t_len / 4;
|
|
|
32
|
|
|
33 uint64_t i=0;
|
|
|
34
|
|
|
35 __m128i * ptr1;
|
|
|
36 __m128i * ptr2;
|
|
|
37 __m128i res = _mm_setzero_si128();
|
|
|
38 ptr2 = (__m128i *) check;
|
|
|
39
|
|
|
40 ldbargs[0].read_from = 0;
|
|
|
41 for(i=0; i<a_fourth - 16; i+=16){
|
|
|
42 ptr1 = (__m128i *) &seq_buffer[i];
|
|
|
43 res = _mm_cmpeq_epi8(*ptr1, *ptr2);
|
|
|
44 // TODO parallelize this
|
|
|
45 *n_seqs += (uint64_t) (((unsigned char *) &res)[0]) >> 7;
|
|
|
46 *n_seqs += (uint64_t) (((unsigned char *) &res)[1]) >> 7;
|
|
|
47 *n_seqs += (uint64_t) (((unsigned char *) &res)[2]) >> 7;
|
|
|
48 *n_seqs += (uint64_t) (((unsigned char *) &res)[3]) >> 7;
|
|
|
49 *n_seqs += (uint64_t) (((unsigned char *) &res)[4]) >> 7;
|
|
|
50 *n_seqs += (uint64_t) (((unsigned char *) &res)[5]) >> 7;
|
|
|
51 *n_seqs += (uint64_t) (((unsigned char *) &res)[6]) >> 7;
|
|
|
52 *n_seqs += (uint64_t) (((unsigned char *) &res)[7]) >> 7;
|
|
|
53 *n_seqs += (uint64_t) (((unsigned char *) &res)[8]) >> 7;
|
|
|
54 *n_seqs += (uint64_t) (((unsigned char *) &res)[9]) >> 7;
|
|
|
55 *n_seqs += (uint64_t) (((unsigned char *) &res)[10]) >> 7;
|
|
|
56 *n_seqs += (uint64_t) (((unsigned char *) &res)[11]) >> 7;
|
|
|
57 *n_seqs += (uint64_t) (((unsigned char *) &res)[12]) >> 7;
|
|
|
58 *n_seqs += (uint64_t) (((unsigned char *) &res)[13]) >> 7;
|
|
|
59 *n_seqs += (uint64_t) (((unsigned char *) &res)[14]) >> 7;
|
|
|
60 *n_seqs += (uint64_t) (((unsigned char *) &res)[15]) >> 7;
|
|
|
61 }
|
|
|
62 while(seq_buffer[i] != '>'){ ++i; }
|
|
|
63 ldbargs[0].read_to = i;
|
|
|
64 ldbargs[0].data_database->n_seqs = *n_seqs;
|
|
|
65 ldbargs[1].read_from = i;
|
|
|
66 while(i % 16 != 0){ ++i; if(seq_buffer[i] == '>') terror("Sequence of length < 16"); }
|
|
|
67 if(i > ldbargs[1].read_from) ++(*n_seqs); // For the advance to align bytes
|
|
|
68
|
|
|
69 //printf("seqs at 0: %"PRIu64" readfrom: %"PRIu64", readto: %"PRIu64", \n", ldbargs[0].data_database->n_seqs, ldbargs[0].read_from, ldbargs[0].read_to);
|
|
|
70
|
|
|
71
|
|
|
72 for(i=i; i<2*(a_fourth) - 16; i+=16){
|
|
|
73 ptr1 = (__m128i *) &seq_buffer[i];
|
|
|
74 res = _mm_cmpeq_epi8(*ptr1, *ptr2);
|
|
|
75 // TODO parallelize this
|
|
|
76 *n_seqs += (uint64_t) (((unsigned char *) &res)[0]) >> 7;
|
|
|
77 *n_seqs += (uint64_t) (((unsigned char *) &res)[1]) >> 7;
|
|
|
78 *n_seqs += (uint64_t) (((unsigned char *) &res)[2]) >> 7;
|
|
|
79 *n_seqs += (uint64_t) (((unsigned char *) &res)[3]) >> 7;
|
|
|
80 *n_seqs += (uint64_t) (((unsigned char *) &res)[4]) >> 7;
|
|
|
81 *n_seqs += (uint64_t) (((unsigned char *) &res)[5]) >> 7;
|
|
|
82 *n_seqs += (uint64_t) (((unsigned char *) &res)[6]) >> 7;
|
|
|
83 *n_seqs += (uint64_t) (((unsigned char *) &res)[7]) >> 7;
|
|
|
84 *n_seqs += (uint64_t) (((unsigned char *) &res)[8]) >> 7;
|
|
|
85 *n_seqs += (uint64_t) (((unsigned char *) &res)[9]) >> 7;
|
|
|
86 *n_seqs += (uint64_t) (((unsigned char *) &res)[10]) >> 7;
|
|
|
87 *n_seqs += (uint64_t) (((unsigned char *) &res)[11]) >> 7;
|
|
|
88 *n_seqs += (uint64_t) (((unsigned char *) &res)[12]) >> 7;
|
|
|
89 *n_seqs += (uint64_t) (((unsigned char *) &res)[13]) >> 7;
|
|
|
90 *n_seqs += (uint64_t) (((unsigned char *) &res)[14]) >> 7;
|
|
|
91 *n_seqs += (uint64_t) (((unsigned char *) &res)[15]) >> 7;
|
|
|
92 }
|
|
|
93 while(seq_buffer[i] != '>') ++i;
|
|
|
94 ldbargs[1].read_to = i;
|
|
|
95 ldbargs[1].data_database->n_seqs = *n_seqs - ldbargs[0].data_database->n_seqs;
|
|
|
96 ldbargs[2].read_from = i;
|
|
|
97 while(i % 16 != 0){ ++i; if(seq_buffer[i] == '>') terror("Sequence of length < 16"); }
|
|
|
98 if(i > ldbargs[2].read_from) ++(*n_seqs); // For the advance to align bytes
|
|
|
99
|
|
|
100 //printf("seqs at 1: %"PRIu64" readfrom: %"PRIu64", readto: %"PRIu64", \n", ldbargs[1].data_database->n_seqs, ldbargs[1].read_from, ldbargs[1].read_to);
|
|
|
101
|
|
|
102
|
|
|
103 for(i=i; i<3*(a_fourth) - 16; i+=16){
|
|
|
104 ptr1 = (__m128i *) &seq_buffer[i];
|
|
|
105 res = _mm_cmpeq_epi8(*ptr1, *ptr2);
|
|
|
106 // TODO parallelize this
|
|
|
107 *n_seqs += (uint64_t) (((unsigned char *) &res)[0]) >> 7;
|
|
|
108 *n_seqs += (uint64_t) (((unsigned char *) &res)[1]) >> 7;
|
|
|
109 *n_seqs += (uint64_t) (((unsigned char *) &res)[2]) >> 7;
|
|
|
110 *n_seqs += (uint64_t) (((unsigned char *) &res)[3]) >> 7;
|
|
|
111 *n_seqs += (uint64_t) (((unsigned char *) &res)[4]) >> 7;
|
|
|
112 *n_seqs += (uint64_t) (((unsigned char *) &res)[5]) >> 7;
|
|
|
113 *n_seqs += (uint64_t) (((unsigned char *) &res)[6]) >> 7;
|
|
|
114 *n_seqs += (uint64_t) (((unsigned char *) &res)[7]) >> 7;
|
|
|
115 *n_seqs += (uint64_t) (((unsigned char *) &res)[8]) >> 7;
|
|
|
116 *n_seqs += (uint64_t) (((unsigned char *) &res)[9]) >> 7;
|
|
|
117 *n_seqs += (uint64_t) (((unsigned char *) &res)[10]) >> 7;
|
|
|
118 *n_seqs += (uint64_t) (((unsigned char *) &res)[11]) >> 7;
|
|
|
119 *n_seqs += (uint64_t) (((unsigned char *) &res)[12]) >> 7;
|
|
|
120 *n_seqs += (uint64_t) (((unsigned char *) &res)[13]) >> 7;
|
|
|
121 *n_seqs += (uint64_t) (((unsigned char *) &res)[14]) >> 7;
|
|
|
122 *n_seqs += (uint64_t) (((unsigned char *) &res)[15]) >> 7;
|
|
|
123 }
|
|
|
124 while(seq_buffer[i] != '>') ++i;
|
|
|
125 ldbargs[2].read_to = i;
|
|
|
126 ldbargs[2].data_database->n_seqs = *n_seqs - (ldbargs[1].data_database->n_seqs + ldbargs[0].data_database->n_seqs);
|
|
|
127 ldbargs[3].read_from = i;
|
|
|
128 while(i % 16 != 0){ ++i; if(seq_buffer[i] == '>') terror("Sequence of length < 16"); }
|
|
|
129 if(i > ldbargs[3].read_from) ++(*n_seqs); // For the advance to align bytes
|
|
|
130
|
|
|
131 //printf("seqs at 2: %"PRIu64" readfrom: %"PRIu64", readto: %"PRIu64", \n", ldbargs[2].data_database->n_seqs, ldbargs[2].read_from, ldbargs[2].read_to);
|
|
|
132
|
|
|
133
|
|
|
134 for(i=i; i<(*t_len) - 16; i+=16){
|
|
|
135 ptr1 = (__m128i *) &seq_buffer[i];
|
|
|
136 res = _mm_cmpeq_epi8(*ptr1, *ptr2);
|
|
|
137 // TODO parallelize this
|
|
|
138 *n_seqs += (uint64_t) (((unsigned char *) &res)[0]) >> 7;
|
|
|
139 *n_seqs += (uint64_t) (((unsigned char *) &res)[1]) >> 7;
|
|
|
140 *n_seqs += (uint64_t) (((unsigned char *) &res)[2]) >> 7;
|
|
|
141 *n_seqs += (uint64_t) (((unsigned char *) &res)[3]) >> 7;
|
|
|
142 *n_seqs += (uint64_t) (((unsigned char *) &res)[4]) >> 7;
|
|
|
143 *n_seqs += (uint64_t) (((unsigned char *) &res)[5]) >> 7;
|
|
|
144 *n_seqs += (uint64_t) (((unsigned char *) &res)[6]) >> 7;
|
|
|
145 *n_seqs += (uint64_t) (((unsigned char *) &res)[7]) >> 7;
|
|
|
146 *n_seqs += (uint64_t) (((unsigned char *) &res)[8]) >> 7;
|
|
|
147 *n_seqs += (uint64_t) (((unsigned char *) &res)[9]) >> 7;
|
|
|
148 *n_seqs += (uint64_t) (((unsigned char *) &res)[10]) >> 7;
|
|
|
149 *n_seqs += (uint64_t) (((unsigned char *) &res)[11]) >> 7;
|
|
|
150 *n_seqs += (uint64_t) (((unsigned char *) &res)[12]) >> 7;
|
|
|
151 *n_seqs += (uint64_t) (((unsigned char *) &res)[13]) >> 7;
|
|
|
152 *n_seqs += (uint64_t) (((unsigned char *) &res)[14]) >> 7;
|
|
|
153 *n_seqs += (uint64_t) (((unsigned char *) &res)[15]) >> 7;
|
|
|
154 }
|
|
|
155 while(i < *t_len){
|
|
|
156 if(seq_buffer[i] == '>') ++(*n_seqs);
|
|
|
157 ++i;
|
|
|
158 }
|
|
|
159 ldbargs[3].read_to = *t_len;
|
|
|
160 ldbargs[3].data_database->n_seqs = *n_seqs - (ldbargs[2].data_database->n_seqs + ldbargs[1].data_database->n_seqs + ldbargs[0].data_database->n_seqs);
|
|
|
161
|
|
|
162 //printf("seqs at 3: %"PRIu64" readfrom: %"PRIu64", readto: %"PRIu64", \n", ldbargs[3].data_database->n_seqs, ldbargs[3].read_from, ldbargs[3].read_to);
|
|
|
163
|
|
|
164 }
|
|
|
165
|
|
|
166 Queue * generate_queue(Head * queue_head, uint64_t t_reads, uint64_t n_threads, uint64_t levels){
|
|
|
167 uint64_t i, j;
|
|
|
168 uint64_t reads_per_thread;
|
|
|
169 if(levels > t_reads) levels = t_reads;
|
|
|
170 uint64_t pieces = t_reads/levels;
|
|
|
171 uint64_t from, to, t_queues = 0, current_queue = 0;
|
|
|
172 for(i=0;i<levels;i++) t_queues += ((i+1)*n_threads);
|
|
|
173 if(t_queues > t_reads) t_queues = t_reads;
|
|
|
174 Queue * queues = (Queue *) malloc(t_queues * sizeof(Queue));
|
|
|
175 if(queues == NULL) terror("Could not allocate queue tasks");
|
|
|
176 queue_head->head = &queues[0];
|
|
|
177
|
|
|
178
|
|
|
179 for(i=0;i<levels;i++){
|
|
|
180
|
|
|
181 //reads_per_thread = (uint64_t) (floorl((long double) pieces / (long double) ((i+1)*n_threads)));
|
|
|
182 reads_per_thread = (uint64_t) (ceill((long double) pieces / (long double) ((i+1)*n_threads)));
|
|
|
183
|
|
|
184
|
|
|
185 for(j=0;j<(i+1)*n_threads;j++){
|
|
|
186 from = j * reads_per_thread + (pieces*i);
|
|
|
187 to = (j + 1) * reads_per_thread + (pieces*i);
|
|
|
188
|
|
|
189 if(j == (i+1)*n_threads - 1) to = pieces*(i+1);
|
|
|
190
|
|
|
191 if(from >= t_reads){
|
|
|
192 queues[current_queue-1].next = NULL;
|
|
|
193 return &queues[0];
|
|
|
194 }
|
|
|
195
|
|
|
196
|
|
|
197 if(i == levels - 1 && j == (i+1)*n_threads - 1){
|
|
|
198 //If its the last
|
|
|
199 queues[current_queue].next = NULL;
|
|
|
200 }else{
|
|
|
201 //Else add it to the queue
|
|
|
202 queues[current_queue].next = &queues[current_queue+1];
|
|
|
203 }
|
|
|
204
|
|
|
205
|
|
|
206 queues[current_queue].r1 = from;
|
|
|
207 queues[current_queue].r2 = (to <= t_reads) ? (to) : (t_reads);
|
|
|
208 current_queue++;
|
|
|
209 //printf("current_piece: %"PRIu64"-%"PRIu64" diff: %"PRIu64"\n", from, to, to - from);
|
|
|
210
|
|
|
211 }
|
|
|
212
|
|
|
213 }
|
|
|
214 //printf("TREADS was %"PRIu64"\n", t_reads);
|
|
|
215 queues[current_queue-1].r2 = t_reads;
|
|
|
216 return &queues[0];
|
|
|
217 }
|
|
|
218
|
|
|
219 void print_queue(Queue * q){
|
|
|
220 fprintf(stdout, "Task: %"PRIu64"-%"PRIu64"\n", q->r1, q->r2);
|
|
|
221 }
|
|
|
222
|
|
|
223 Queue * get_task_from_queue(Head * queue_head, pthread_mutex_t * lock){
|
|
|
224 pthread_mutex_lock(lock);
|
|
|
225
|
|
|
226 Queue * ptr = queue_head->head;
|
|
|
227 if(queue_head->head != NULL) queue_head->head = queue_head->head->next;
|
|
|
228 //if(ptr != NULL){ printf("Taking "); /*print_queue(ptr);*/ }
|
|
|
229
|
|
|
230 pthread_mutex_unlock(lock);
|
|
|
231
|
|
|
232
|
|
|
233 return ptr;
|
|
|
234 }
|
|
|
235
|
|
|
236 uint64_t quick_pow4(uint64_t n){
|
|
|
237 static uint64_t pow4[33]={1L, 4L, 16L, 64L, 256L, 1024L, 4096L, 16384L, 65536L,
|
|
|
238 262144L, 1048576L, 4194304L, 16777216L, 67108864L, 268435456L, 1073741824L, 4294967296L,
|
|
|
239 17179869184L, 68719476736L, 274877906944L, 1099511627776L, 4398046511104L, 17592186044416L,
|
|
|
240 70368744177664L, 281474976710656L, 1125899906842624L, 4503599627370496L, 18014398509481984L,
|
|
|
241 72057594037927936L, 288230376151711744L, 1152921504606846976L, 4611686018427387904L};
|
|
|
242 return pow4[n];
|
|
|
243 }
|
|
|
244
|
|
|
245 uint64_t quick_pow4byLetter(uint64_t n, const char c){
|
|
|
246 static uint64_t pow4_G[33]={2*1L, 2*4L, 2*16L, 2*64L, 2*256L, 2*1024L, 2*4096L, 2*16384L, 2*65536L,
|
|
|
247 (uint64_t)2*262144L, (uint64_t)2*1048576L,(uint64_t)2*4194304L, (uint64_t)2*16777216L, (uint64_t)2*67108864L, (uint64_t)2*268435456L, (uint64_t)2*1073741824L, (uint64_t)2*4294967296L,
|
|
|
248 (uint64_t)2*17179869184L, (uint64_t)2*68719476736L, (uint64_t)2*274877906944L, (uint64_t)2*1099511627776L, (uint64_t)2*4398046511104L, (uint64_t)2*17592186044416L,
|
|
|
249 (uint64_t)2*70368744177664L, (uint64_t)2*281474976710656L, (uint64_t)2*1125899906842624L, (uint64_t)2*4503599627370496L, (uint64_t)2*18014398509481984L,
|
|
|
250 (uint64_t)2*72057594037927936L, (uint64_t) 2*288230376151711744L, (uint64_t) 2*1152921504606846976L, (uint64_t) 2*4611686018427387904L};
|
|
|
251
|
|
|
252 static uint64_t pow4_T[33]={3*1L, 3*4L, 3*16L, 3*64L, 3*256L, 3*1024L, 3*4096L, 3*16384L, 3*65536L,
|
|
|
253 (uint64_t)3*262144L, (uint64_t) 3*1048576L, (uint64_t)3*4194304L, (uint64_t)3*16777216L, (uint64_t)3*67108864L, (uint64_t)3*268435456L, (uint64_t)3*1073741824L, (uint64_t)3*4294967296L,
|
|
|
254 (uint64_t)3*17179869184L, (uint64_t)3*68719476736L, (uint64_t)3*274877906944L, (uint64_t)3*1099511627776L, (uint64_t)3*4398046511104L, (uint64_t)3*17592186044416L,
|
|
|
255 (uint64_t)3*70368744177664L, (uint64_t)3*281474976710656L, (uint64_t)3*1125899906842624L, (uint64_t)3*4503599627370496L, (uint64_t)3*18014398509481984L,
|
|
|
256 (uint64_t)3*72057594037927936L, (uint64_t) 3*288230376151711744L, (uint64_t) 3*1152921504606846976L, (uint64_t) 3*4611686018427387904L};
|
|
|
257
|
|
|
258 if(c == 'A') return 0;
|
|
|
259 if(c == 'C') return quick_pow4(n);
|
|
|
260 if(c == 'G') return pow4_G[n];
|
|
|
261 if(c == 'T') return pow4_T[n];
|
|
|
262 return 0;
|
|
|
263 }
|
|
|
264
|
|
|
265 uint64_t hashOfWord(const unsigned char * word, uint32_t k){
|
|
|
266
|
|
|
267 uint64_t value = 0, jIdx;
|
|
|
268 for(jIdx=0;jIdx<k;jIdx++){
|
|
|
269 value += quick_pow4byLetter(k-(jIdx+1), (char) word[jIdx]);
|
|
|
270 }
|
|
|
271 return value;
|
|
|
272 }
|
|
|
273
|
|
|
274
|
|
|
275 uint64_t asciiToUint64(const char *text){
|
|
|
276 uint64_t number=0;
|
|
|
277
|
|
|
278 for(;*text;text++){
|
|
|
279 char digit=*text-'0';
|
|
|
280 number=(number*10)+digit;
|
|
|
281 }
|
|
|
282 return number;
|
|
|
283 }
|
