+/***** BEGIN: multi-threading *****/
+
+typedef struct {
+ BGZF *fp;
+ struct mtaux_t *mt;
+ void *buf;
+ int i, errcode, toproc;
+} worker_t;
+
+typedef struct mtaux_t {
+ int n_threads, n_blks, curr, done;
+ volatile int proc_cnt;
+ void **blk;
+ int *len;
+ worker_t *w;
+ pthread_t *tid;
+ pthread_mutex_t lock;
+ pthread_cond_t cv;
+} mtaux_t;
+
+static int worker_aux(worker_t *w)
+{
+ int i, tmp, stop = 0;
+ // wait for condition: to process or all done
+ pthread_mutex_lock(&w->mt->lock);
+ while (!w->toproc && !w->mt->done)
+ pthread_cond_wait(&w->mt->cv, &w->mt->lock);
+ if (w->mt->done) stop = 1;
+ w->toproc = 0;
+ pthread_mutex_unlock(&w->mt->lock);
+ if (stop) return 1; // to quit the thread
+ w->errcode = 0;
+ for (i = w->i; i < w->mt->curr; i += w->mt->n_threads) {
+ int clen = BGZF_MAX_BLOCK_SIZE;
+ if (bgzf_compress(w->buf, &clen, w->mt->blk[i], w->mt->len[i], w->fp->compress_level) != 0)
+ w->errcode |= BGZF_ERR_ZLIB;
+ memcpy(w->mt->blk[i], w->buf, clen);
+ w->mt->len[i] = clen;
+ }
+ tmp = __sync_fetch_and_add(&w->mt->proc_cnt, 1);
+ return 0;
+}
+
+static void *mt_worker(void *data)
+{
+ while (worker_aux(data) == 0);
+ return 0;
+}
+
+int bgzf_mt(BGZF *fp, int n_threads, int n_sub_blks)
+{
+ int i;
+ mtaux_t *mt;
+ pthread_attr_t attr;
+ if (!fp->is_write || fp->mt || n_threads <= 1) return -1;
+ mt = calloc(1, sizeof(mtaux_t));
+ mt->n_threads = n_threads;
+ mt->n_blks = n_threads * n_sub_blks;
+ mt->len = calloc(mt->n_blks, sizeof(int));
+ mt->blk = calloc(mt->n_blks, sizeof(void*));
+ for (i = 0; i < mt->n_blks; ++i)
+ mt->blk[i] = malloc(BGZF_MAX_BLOCK_SIZE);
+ mt->tid = calloc(mt->n_threads, sizeof(pthread_t)); // tid[0] is not used, as the worker 0 is launched by the master
+ mt->w = calloc(mt->n_threads, sizeof(worker_t));
+ for (i = 0; i < mt->n_threads; ++i) {
+ mt->w[i].i = i;
+ mt->w[i].mt = mt;
+ mt->w[i].fp = fp;
+ mt->w[i].buf = malloc(BGZF_MAX_BLOCK_SIZE);
+ }
+ pthread_attr_init(&attr);
+ pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_JOINABLE);
+ pthread_mutex_init(&mt->lock, 0);
+ pthread_cond_init(&mt->cv, 0);
+ for (i = 1; i < mt->n_threads; ++i) // worker 0 is effectively launched by the master thread
+ pthread_create(&mt->tid[i], &attr, mt_worker, &mt->w[i]);
+ fp->mt = mt;
+ return 0;
+}
+
+static void mt_destroy(mtaux_t *mt)
+{
+ int i;
+ // signal all workers to quit
+ pthread_mutex_lock(&mt->lock);
+ mt->done = 1; mt->proc_cnt = 0;
+ pthread_cond_broadcast(&mt->cv);
+ pthread_mutex_unlock(&mt->lock);
+ for (i = 1; i < mt->n_threads; ++i) pthread_join(mt->tid[i], 0); // worker 0 is effectively launched by the master thread
+ // free other data allocated on heap
+ for (i = 0; i < mt->n_blks; ++i) free(mt->blk[i]);
+ for (i = 0; i < mt->n_threads; ++i) free(mt->w[i].buf);
+ free(mt->blk); free(mt->len); free(mt->w); free(mt->tid);
+ pthread_cond_destroy(&mt->cv);
+ pthread_mutex_destroy(&mt->lock);
+ free(mt);
+}
+
+static void mt_queue(BGZF *fp)
+{
+ mtaux_t *mt = (mtaux_t*)fp->mt;
+ assert(mt->curr < mt->n_blks); // guaranteed by the caller
+ memcpy(mt->blk[mt->curr], fp->uncompressed_block, fp->block_offset);
+ mt->len[mt->curr] = fp->block_offset;
+ fp->block_offset = 0;
+ ++mt->curr;
+}
+
+static int mt_flush(BGZF *fp)
+{
+ int i;
+ mtaux_t *mt = (mtaux_t*)fp->mt;
+ if (fp->block_offset) mt_queue(fp); // guaranteed that assertion does not fail
+ // signal all the workers to compress
+ pthread_mutex_lock(&mt->lock);
+ for (i = 0; i < mt->n_threads; ++i) mt->w[i].toproc = 1;
+ mt->proc_cnt = 0;
+ pthread_cond_broadcast(&mt->cv);
+ pthread_mutex_unlock(&mt->lock);
+ // worker 0 is doing things here
+ worker_aux(&mt->w[0]);
+ // wait for all the threads to complete
+ while (mt->proc_cnt < mt->n_threads);
+ // dump data to disk
+ for (i = 0; i < mt->n_threads; ++i) fp->errcode |= mt->w[i].errcode;
+ for (i = 0; i < mt->curr; ++i)
+ if (fwrite(mt->blk[i], 1, mt->len[i], fp->fp) != mt->len[i])
+ fp->errcode |= BGZF_ERR_IO;
+ mt->curr = 0;
+ return 0;
+}
+
+static int mt_lazy_flush(BGZF *fp)
+{
+ mtaux_t *mt = (mtaux_t*)fp->mt;
+ mt_queue(fp);
+ if (mt->curr == mt->n_blks)
+ return mt_flush(fp);
+ return -1;
+}
+
+static ssize_t mt_write(BGZF *fp, const void *data, ssize_t length)
+{
+ const uint8_t *input = data;
+ ssize_t rest = length;
+ while (rest) {
+ int copy_length = BGZF_BLOCK_SIZE - fp->block_offset < rest? BGZF_BLOCK_SIZE - fp->block_offset : rest;
+ memcpy(fp->uncompressed_block + fp->block_offset, input, copy_length);
+ fp->block_offset += copy_length; input += copy_length; rest -= copy_length;
+ if (fp->block_offset == BGZF_BLOCK_SIZE) mt_lazy_flush(fp);
+ }
+ return length - rest;
+}
+
+/***** END: multi-threading *****/
+