void UHMMCalibrate::calibrateParallel(WorkPool_s *wpool, TaskStateInfo& si) {
HMMERTaskLocalData *tls = getHMMERTaskLocalData();
struct alphabet_s *al = &tls->al;
struct plan7_s *hmm = wpool->hmm;
struct dpmatrix_s *mx = CreatePlan7Matrix(1, hmm->M, 25, 0);
int len;
float sc;
char *seq;
unsigned char *dsq;
while(true) {
/* generate a sequence */
{
QMutexLocker locker(&wpool->lockInput);
wpool->nseq++;
if (wpool->nseq > wpool->nsample) { /* we're done; release input lock, break loop */
break;
}
if (wpool->fixedlen) {
len = wpool->fixedlen;
} else {
do {
len = (int) Gaussrandom(wpool->lenmean, wpool->lensd);
} while (len < 1);
}
seq = RandomSequence(al->Alphabet, wpool->randomseq.data(), al->Alphabet_size, len);
}
/* compute score */
dsq = DigitizeSequence(seq, len);
if (P7ViterbiSpaceOK(len, hmm->M, mx)) {
sc = P7Viterbi(dsq, len, hmm, mx, NULL);
} else {
int pStub;
sc = P7SmallViterbi(dsq, len, hmm, mx, NULL, pStub);
}
free(dsq);
free(seq);
/* save output */
QMutexLocker locker(&wpool->lockOutput);
AddToHistogram(wpool->hist, sc);
wpool->max_score = qMax(wpool->max_score, sc);
si.progress = int(100*wpool->nseq/float(wpool->nsample)); //TODO: update progress for all tasks?
if (wpool->progress!=NULL) {
*wpool->progress = si.progress;
}
}
FreePlan7Matrix(mx);
}
/* Function: main_loop_serial()
* Date: SRE, Tue Aug 18 16:18:28 1998 [St. Louis]
*
* Purpose: Given an HMM and parameters for synthesizing random
* sequences; return a histogram of scores.
* (Serial version)
*
* Args: hmm - an HMM to calibrate.
* seed - random number seed
* nsample - number of seqs to synthesize
* lenmean - mean length of random sequence
* lensd - std dev of random seq length
* fixedlen - if nonzero, override lenmean, always this len
* ret_hist - RETURN: the score histogram
* ret_max - RETURN: highest score seen in simulation
*
* Returns: (void)
* hist is alloc'ed here, and must be free'd by caller.
*/
static void
main_loop_serial(struct plan7_s *hmm, int seed, int nsample,
float lenmean, float lensd, int fixedlen,
struct histogram_s **ret_hist, float *ret_max)
{
struct histogram_s *hist;
float randomseq[MAXABET];
float p1;
float max;
char *seq;
char *dsq;
float score;
int sqlen;
int idx;
/* Initialize.
* We assume we've already set the alphabet (safe, because
* HMM input sets the alphabet).
*/
sre_srandom(seed);
P7Logoddsify(hmm, TRUE);
P7DefaultNullModel(randomseq, &p1);
hist = AllocHistogram(-200, 200, 100);
max = -FLT_MAX;
for (idx = 0; idx < nsample; idx++)
{
/* choose length of random sequence */
if (fixedlen) sqlen = fixedlen;
else do sqlen = (int) Gaussrandom(lenmean, lensd); while (sqlen < 1);
/* generate it */
seq = RandomSequence(Alphabet, randomseq, Alphabet_size, sqlen);
dsq = DigitizeSequence(seq, sqlen);
if (P7ViterbiSize(sqlen, hmm->M) <= RAMLIMIT)
score = P7Viterbi(dsq, sqlen, hmm, NULL);
else
score = P7SmallViterbi(dsq, sqlen, hmm, NULL);
AddToHistogram(hist, score);
if (score > max) max = score;
free(dsq);
free(seq);
}
*ret_hist = hist;
*ret_max = max;
return;
}
void
JMMArrayTable::StreamAllocationSizeHistogram
(
ostream& output
)
const
{
JSize histo[ JMemoryManager::kHistogramSlotCount ];
bzero(histo, sizeof(histo));
const JSize count = itsAllocatedTable->GetElementCount();
for (JIndex i=1;i<=count;i++)
{
AddToHistogram(itsAllocatedTable->GetElement(i), histo);
}
StreamHistogram(output, histo);
}
void
JMMHashTable::StreamAllocationSizeHistogram
(
ostream& output
)
const
{
JSize histo[ JMemoryManager::kHistogramSlotCount ];
bzero(histo, sizeof(histo));
JConstHashCursor<JMMRecord> cursor(itsAllocatedTable);
while ( cursor.NextFull() )
{
AddToHistogram(cursor.GetValue(), histo);
}
StreamHistogram(output, histo);
}
boolean should_store_Hscore(Hscore * hs,int score)
{
hs->total++;
if( hs->report_level != -1 && (hs->total % hs->report_level == 0)) {
if( hs->len > 0) {
info("Done %d comparisons: last stored comparison was %s to %s",hs->total,hs->ds[hs->len-1]->query->name,hs->ds[hs->len-1]->target->name);
} else {
info("Done %d comparisons: No stored comparisons",hs->total);
}
}
if( hs->his != NULL && hs->score_to_his != NULL ) {
AddToHistogram(hs->his,(*hs->score_to_his)(score));
}
if( hs->should_store == NULL ) {
return TRUE;
}
return (*hs->should_store)(score,hs->score_level);
}
static void
main_loop_serial(struct plan7_s *hmm, const char* seq, int seqLen, struct threshold_s *thresh, int do_forward,
int do_null2, int do_xnu, struct histogram_s *histogram, struct tophit_s *ghit, struct tophit_s *dhit,
int *ret_nseq, TaskStateInfo& ti)
{
//get HMMERTaskLocalData
HMMERTaskLocalData *tld = getHMMERTaskLocalData();
alphabet_s *al = &tld->al;
struct dpmatrix_s *mx; // DP matrix, growable
struct p7trace_s *tr; // traceback
unsigned char *dsq; // digitized target sequence
float sc; // score of an HMM search
double pvalue; // pvalue of an HMM score
double evalue; // evalue of an HMM score
// Create a DP matrix; initially only two rows big, but growable;
// we overalloc by 25 rows (L dimension) when we grow; not growable
// in model dimension, since we know the hmm size
mx = CreatePlan7Matrix(1, hmm->M, 25, 0);
assert(seqLen > 0);
dsq = DigitizeSequence(seq, seqLen);
if (do_xnu && al->Alphabet_type == hmmAMINO) {
XNU(dsq, seqLen);
}
// 1. Recover a trace by Viterbi.
// In extreme cases, the alignment may be literally impossible;
// in which case, the score comes out ridiculously small (but not
// necessarily <= -INFTY, because we're not terribly careful
// about underflow issues), and tr will be returned as NULL.
if (P7ViterbiSpaceOK(seqLen, hmm->M, mx)) {
sc = P7Viterbi(dsq, seqLen, hmm, mx, &tr);
} else {
sc = P7SmallViterbi(dsq, seqLen, hmm, mx, &tr, ti.progress);
}
// 2. If we're using Forward scores, calculate the
// whole sequence score; this overrides anything
// PostprocessSignificantHit() is going to do to the per-seq score.
if (do_forward) {
sc = P7Forward(dsq, seqLen, hmm, NULL);
if (do_null2) sc -= TraceScoreCorrection(hmm, tr, dsq);
}
// 2. Store score/pvalue for global alignment; will sort on score,
// which in hmmsearch is monotonic with E-value.
// Keep all domains in a significant sequence hit.
// We can only make a lower bound estimate of E-value since
// we don't know the final value of nseq yet, so the list
// of hits we keep in memory is >= the list we actually
// output.
//
pvalue = PValue(hmm, sc);
evalue = thresh->Z ? (double) thresh->Z * pvalue : (double) pvalue;
if (sc >= thresh->globT && evalue <= thresh->globE) {
sc = PostprocessSignificantHit(ghit, dhit,
tr, hmm, dsq, seqLen,
(char *)"sequence", //todo: sqinfo.name,
NULL,
NULL,
do_forward, sc,
do_null2,
thresh,
FALSE); // FALSE-> not hmmpfam mode, hmmsearch mode
}
AddToHistogram(histogram, sc);
P7FreeTrace(tr);
free(dsq);
FreePlan7Matrix(mx);
return;
}
static void main_loop_serial(struct plan7_s *hmm, int seed, int nsample,
float lenmean, float lensd, int fixedlen,
struct histogram_s **ret_hist, float *ret_max, int& cancelFlag, int& progress)
{
struct histogram_s *hist;
struct dpmatrix_s *mx;
float randomseq[MAXABET];
float p1;
float max;
char *seq;
unsigned char *dsq;
float score;
int sqlen;
int idx;
// Initialize.
// We assume we've already set the alphabet (safe, because
// HMM input sets the alphabet).
sre_srandom(seed);
//get HMMERTaskLocalData
HMMERTaskLocalData *tls = getHMMERTaskLocalData();
alphabet_s &al = tls->al;
SetAlphabet(hmm->atype);
P7Logoddsify(hmm, TRUE);
P7DefaultNullModel(randomseq, &p1);
hist = AllocHistogram(-200, 200, 100);
mx = CreatePlan7Matrix(1, hmm->M, 25, 0);
max = -FLT_MAX;
progress = 0;
int pStub;
for (idx = 0; idx < nsample && !cancelFlag; idx++) {
// choose length of random sequence
if (fixedlen) {
sqlen = fixedlen;
} else {
do sqlen = (int) Gaussrandom(lenmean, lensd); while (sqlen < 1);
}
// generate it
seq = RandomSequence(al.Alphabet, randomseq, al.Alphabet_size, sqlen);
dsq = DigitizeSequence(seq, sqlen);
if (P7ViterbiSpaceOK(sqlen, hmm->M, mx)) {
score = P7Viterbi(dsq, sqlen, hmm, mx, NULL);
} else {
score = P7SmallViterbi(dsq, sqlen, hmm, mx, NULL, pStub);
}
AddToHistogram(hist, score);
max = qMax(score, max);
progress = int(100*idx/float(nsample));
free(dsq);
free(seq);
}
FreePlan7Matrix(mx);
*ret_hist = hist;
*ret_max = max;
}
int
main(void)
{
int master_tid; /* PVM TID of our master */
int slaveidx; /* my slave index (0..nslaves-1) */
struct plan7_s *hmm; /* HMM to calibrate, sent from master */
struct histogram_s *hist; /* score histogram */
int hmmidx; /* index of this HMM */
char *seq; /* synthetic random sequence */
char *dsq; /* digitized seq */
int len; /* length of seq */
float sc; /* score of seq aligned to HMM */
float max; /* maximum score seen in sample */
int seed; /* random number seed */
int nsample; /* number of seqs to sample */
int fixedlen; /* if nonzero, fixed length of seq */
float lenmean; /* Gaussian mean length of seq */
float lensd; /* Gaussian length std. dev. for seq */
int fitok; /* TRUE if EVD fit was OK */
float randomseq[MAXABET]; /* iid frequencies of residues */
float p1;
int alphatype; /* alphabet type, hmmAMINO or hmmNUCLEIC */
int idx;
int code;
/* Register leave_pvm() cleanup function so any exit() call
* first calls pvm_exit().
*/
if (atexit(leave_pvm) != 0) { pvm_exit(); Die("slave couldn't register leave_pvm()"); }
/*****************************************************************
* initialization.
* Master broadcasts the problem to us: parameters of the
* HMM calibration.
******************************************************************/
master_tid = pvm_parent(); /* who's our master? */
pvm_recv(master_tid, HMMPVM_INIT);
pvm_upkint(&nsample, 1, 1);
pvm_upkint(&fixedlen, 1, 1);
pvm_upkfloat(&lenmean, 1, 1);
pvm_upkfloat(&lensd, 1, 1);
/* tell the master we're OK and ready to go (or not)
*/
code = HMMPVM_OK;
pvm_initsend(PvmDataDefault);
pvm_pkint(&code, 1, 1);
pvm_send(master_tid, HMMPVM_RESULTS);
/*****************************************************************
* Main loop.
* Receive a random number seed, then an HMM to search against.
* If we receive a -1 seed, we shut down.
*****************************************************************/
slaveidx = -1;
for (;;)
{
pvm_recv(master_tid, HMMPVM_WORK);
pvm_upkint(&seed, 1, 1);
if (seed == -1) break; /* shutdown signal */
pvm_upkint(&hmmidx, 1, 1);
pvm_upkint(&alphatype,1, 1);
SetAlphabet(alphatype);
hmm = PVMUnpackHMM();
if (hmm == NULL) Die("oh no, the HMM never arrived");
if (slaveidx == -1) slaveidx = hmmidx;
P7DefaultNullModel(randomseq, &p1);
sre_srandom(seed);
P7Logoddsify(hmm, TRUE);
hist = AllocHistogram(-200, 200, 100);
max = -FLT_MAX;
for (idx = 0; idx < nsample; idx++)
{
/* choose length of random sequence */
if (fixedlen) len = fixedlen;
else do len = (int) Gaussrandom(lenmean, lensd); while (len < 1);
/* generate it */
seq = RandomSequence(Alphabet, randomseq, Alphabet_size, len);
dsq = DigitizeSequence(seq, len);
if (P7ViterbiSize(len, hmm->M) <= RAMLIMIT)
sc = P7Viterbi(dsq, len, hmm, NULL);
else
sc = P7SmallViterbi(dsq, len, hmm, NULL);
AddToHistogram(hist, sc);
if (sc > max) max = sc;
free(seq);
free(dsq);
}
/* Fit an EVD to the observed histogram.
* The TRUE left-censors and fits only the right slope of the histogram.
* The 9999. is an arbitrary high number that means we won't trim outliers
* on the right.
*/
fitok = ExtremeValueFitHistogram(hist, TRUE, 9999.);
/* Return output to master.
* Currently we don't send the histogram back, but we could.
*/
pvm_initsend(PvmDataDefault);
pvm_pkint(&slaveidx, 1, 1);
pvm_pkint(&hmmidx, 1, 1);
PVMPackString(hmm->name);
pvm_pkint(&fitok, 1, 1);
pvm_pkfloat(&(hist->param[EVD_MU]), 1, 1);
pvm_pkfloat(&(hist->param[EVD_LAMBDA]), 1, 1);
pvm_pkfloat(&max, 1, 1);
pvm_send(master_tid, HMMPVM_RESULTS);
/* cleanup
*/
FreeHistogram(hist);
FreePlan7(hmm);
}
/***********************************************
* Cleanup, return.
***********************************************/
return 0; /* pvm_exit() is called by atexit() registration. */
}
/* Function: main_loop_pvm()
* Date: SRE, Wed Aug 19 13:59:54 1998 [St. Louis]
*
* Purpose: Given an HMM and parameters for synthesizing random
* sequences; return a histogram of scores.
* (PVM version)
*
* Args: hmm - an HMM to calibrate.
* seed - random number seed
* nsample - number of seqs to synthesize
* lumpsize- # of seqs per slave exchange; controls granularity
* lenmean - mean length of random sequence
* lensd - std dev of random seq length
* fixedlen- if nonzero, override lenmean, always this len
* hist - RETURN: the score histogram
* ret_max - RETURN: highest score seen in simulation
* extrawatch - RETURN: total CPU time spend in slaves.
* ret_nslaves- RETURN: number of PVM slaves run.
*
* Returns: (void)
* hist is alloc'ed here, and must be free'd by caller.
*/
static void
main_loop_pvm(struct plan7_s *hmm, int seed, int nsample, int lumpsize,
float lenmean, float lensd, int fixedlen,
struct histogram_s **ret_hist, float *ret_max,
Stopwatch_t *extrawatch, int *ret_nslaves)
{
struct histogram_s *hist;
int master_tid;
int *slave_tid;
int nslaves;
int nsent; /* # of seqs we've asked for so far */
int ndone; /* # of seqs we've got results for so far */
int packet; /* # of seqs to have a slave do */
float max;
int slaveidx; /* id of a slave */
float *sc; /* scores returned by a slave */
Stopwatch_t slavewatch;
int i;
StopwatchZero(extrawatch);
hist = AllocHistogram(-200, 200, 100);
max = -FLT_MAX;
/* Initialize PVM
*/
if ((master_tid = pvm_mytid()) < 0)
Die("pvmd not responding -- do you have PVM running?");
#if DEBUGLEVEL >= 1
pvm_catchout(stderr); /* catch output for debugging */
#endif
PVMSpawnSlaves("hmmcalibrate-pvm", &slave_tid, &nslaves);
/* Initialize the slaves
*/
pvm_initsend(PvmDataDefault);
pvm_pkfloat(&lenmean, 1, 1);
pvm_pkfloat(&lensd, 1, 1);
pvm_pkint( &fixedlen, 1, 1);
pvm_pkint( &Alphabet_type, 1, 1);
pvm_pkint( &seed, 1, 1);
if (! PVMPackHMM(hmm)) Die("Failed to pack the HMM");
pvm_mcast(slave_tid, nslaves, HMMPVM_INIT);
SQD_DPRINTF1(("Initialized %d slaves\n", nslaves));
/* Confirm slaves' OK status.
*/
PVMConfirmSlaves(slave_tid, nslaves);
SQD_DPRINTF1(("Slaves confirm that they're ok...\n"));
/* Load the slaves
*/
nsent = ndone = 0;
for (slaveidx = 0; slaveidx < nslaves; slaveidx++)
{
packet = (nsample - nsent > lumpsize ? lumpsize : nsample - nsent);
pvm_initsend(PvmDataDefault);
pvm_pkint(&packet, 1, 1);
pvm_pkint(&slaveidx, 1, 1);
pvm_send(slave_tid[slaveidx], HMMPVM_WORK);
nsent += packet;
}
SQD_DPRINTF1(("Loaded %d slaves\n", nslaves));
/* Receive/send loop
*/
sc = MallocOrDie(sizeof(float) * lumpsize);
while (nsent < nsample)
{
/* integrity check of slaves */
PVMCheckSlaves(slave_tid, nslaves);
/* receive results */
SQD_DPRINTF2(("Waiting for results...\n"));
pvm_recv(-1, HMMPVM_RESULTS);
pvm_upkint(&slaveidx, 1, 1);
pvm_upkint(&packet, 1, 1);
pvm_upkfloat(sc, packet, 1);
SQD_DPRINTF2(("Got results.\n"));
ndone += packet;
/* store results */
for (i = 0; i < packet; i++) {
AddToHistogram(hist, sc[i]);
if (sc[i] > max) max = sc[i];
}
/* send new work */
packet = (nsample - nsent > lumpsize ? lumpsize : nsample - nsent);
pvm_initsend(PvmDataDefault);
pvm_pkint(&packet, 1, 1);
pvm_pkint(&slaveidx, 1, 1);
pvm_send(slave_tid[slaveidx], HMMPVM_WORK);
SQD_DPRINTF2(("Told slave %d to do %d more seqs.\n", slaveidx, packet));
nsent += packet;
}
/* Wait for the last output to come in.
*/
while (ndone < nsample)
{
/* integrity check of slaves */
PVMCheckSlaves(slave_tid, nslaves);
/* receive results */
SQD_DPRINTF1(("Waiting for final results...\n"));
pvm_recv(-1, HMMPVM_RESULTS);
pvm_upkint(&slaveidx, 1, 1);
pvm_upkint(&packet, 1, 1);
pvm_upkfloat(sc, packet, 1);
SQD_DPRINTF2(("Got some final results.\n"));
ndone += packet;
/* store results */
for (i = 0; i < packet; i++) {
AddToHistogram(hist, sc[i]);
if (sc[i] > max) max = sc[i];
}
}
/* Shut down the slaves: send -1,-1,-1.
*/
pvm_initsend(PvmDataDefault);
packet = -1;
pvm_pkint(&packet, 1, 1);
pvm_pkint(&packet, 1, 1);
pvm_pkint(&packet, 1, 1);
pvm_mcast(slave_tid, nslaves, HMMPVM_WORK);
/* Collect stopwatch results; quit the VM; return.
*/
for (i = 0; i < nslaves; i++)
{
pvm_recv(-1, HMMPVM_RESULTS);
pvm_upkint(&slaveidx, 1, 1);
StopwatchPVMUnpack(&slavewatch);
SQD_DPRINTF1(("Slave %d finished; says it used %.2f cpu, %.2f sys\n",
slaveidx, slavewatch.user, slavewatch.sys));
StopwatchInclude(extrawatch, &slavewatch);
}
free(slave_tid);
free(sc);
pvm_exit();
*ret_hist = hist;
*ret_max = max;
*ret_nslaves = nslaves;
return;
}
/* Function: worker_thread()
* Date: SRE, Thu Jul 16 10:41:02 1998 [St. Louis]
*
* Purpose: The procedure executed by the worker threads.
*
* Args: ptr - (void *) that is recast to a pointer to
* the workpool.
*
* Returns: (void *)
*/
void *
worker_thread(void *ptr)
{
struct plan7_s *hmm;
struct workpool_s *wpool;
char *seq;
char *dsq;
int len;
float sc;
int rtn;
Stopwatch_t thread_watch;
StopwatchStart(&thread_watch);
wpool = (struct workpool_s *) ptr;
hmm = wpool->hmm;
for (;;)
{
/* 1. Synthesize a random sequence.
* The input sequence number is a shared resource,
* and sre_random() isn't thread-safe, so protect
* the whole section with mutex.
*/
/* acquire a lock */
if ((rtn = pthread_mutex_lock(&(wpool->input_lock))) != 0)
Die("pthread_mutex_lock failure: %s\n", strerror(rtn));
/* generate a sequence */
wpool->nseq++;
if (wpool->nseq > wpool->nsample)
{ /* we're done; release input lock, break loop */
if ((rtn = pthread_mutex_unlock(&(wpool->input_lock))) != 0)
Die("pthread_mutex_unlock failure: %s\n", strerror(rtn));
break;
}
if (wpool->fixedlen) len = wpool->fixedlen;
else do len = (int) Gaussrandom(wpool->lenmean, wpool->lensd); while (len < 1);
seq = RandomSequence(Alphabet, wpool->randomseq, Alphabet_size, len);
/* release the lock */
if ((rtn = pthread_mutex_unlock(&(wpool->input_lock))) != 0)
Die("pthread_mutex_unlock failure: %s\n", strerror(rtn));
/* 2. Score the sequence against the model.
*/
dsq = DigitizeSequence(seq, len);
if (P7ViterbiSize(len, hmm->M) <= RAMLIMIT)
sc = P7Viterbi(dsq, len, hmm, NULL);
else
sc = P7SmallViterbi(dsq, len, hmm, NULL);
free(dsq);
free(seq);
/* 3. Save the output; hist and max_score are shared,
* so protect this section with the output mutex.
*/
/* acquire lock on the output queue */
if ((rtn = pthread_mutex_lock(&(wpool->output_lock))) != 0)
Die("pthread_mutex_lock failure: %s\n", strerror(rtn));
/* save output */
AddToHistogram(wpool->hist, sc);
if (sc > wpool->max_score) wpool->max_score = sc;
/* release our lock */
if ((rtn = pthread_mutex_unlock(&(wpool->output_lock))) != 0)
Die("pthread_mutex_unlock failure: %s\n", strerror(rtn));
}
StopwatchStop(&thread_watch);
/* acquire lock on the output queue */
if ((rtn = pthread_mutex_lock(&(wpool->output_lock))) != 0)
Die("pthread_mutex_lock failure: %s\n", strerror(rtn));
/* accumulate cpu time into main stopwatch */
StopwatchInclude(&(wpool->watch), &thread_watch);
/* release our lock */
if ((rtn = pthread_mutex_unlock(&(wpool->output_lock))) != 0)
Die("pthread_mutex_unlock failure: %s\n", strerror(rtn));
pthread_exit(NULL);
return NULL; /* solely to silence compiler warnings */
}
