/* process_workunit()
*
* This is the routine that actually does the work.
*
* A work unit consists of one HMM, <hmm>.
* The result is the <scores> array, which contains an array of N scores;
* caller provides this memory.
* How those scores are generated is controlled by the application configuration in <cfg>.
*/
static int
process_workunit(ESL_GETOPTS *go, struct cfg_s *cfg, char *errbuf, P7_HMM *hmm, double *scores, int *alilens)
{
int L = esl_opt_GetInteger(go, "-L");
P7_PROFILE *gm = NULL;
P7_OPROFILE *om = NULL;
P7_REFMX *rmx = NULL;
P7_CHECKPTMX *cx = NULL;
P7_FILTERMX *fx = NULL;
P7_TRACE *tr = NULL;
ESL_DSQ *dsq = NULL;
int i;
int scounts[p7T_NSTATETYPES]; /* state usage counts from a trace */
float sc;
float nullsc;
int status;
P7_HARDWARE *hw;
if ((hw = p7_hardware_Create ()) == NULL) p7_Fail("Couldn't get HW information data structure");
/* Optionally set a custom background, determined by model composition;
* an experimental hack.
*/
if (esl_opt_GetBoolean(go, "--bgcomp"))
{
float *p = NULL;
float KL;
p7_hmm_CompositionKLDist(hmm, cfg->bg, &KL, &p);
esl_vec_FCopy(p, cfg->abc->K, cfg->bg->f);
}
/* Create and configure our generic profile, as requested */
gm = p7_profile_Create(hmm->M, cfg->abc);
if (esl_opt_GetBoolean(go, "--multi"))
{
if (esl_opt_GetBoolean(go, "--dual")) { p7_profile_Config (gm, hmm, cfg->bg); }
else if (esl_opt_GetBoolean(go, "--local")) { p7_profile_ConfigLocal (gm, hmm, cfg->bg, L); }
else if (esl_opt_GetBoolean(go, "--glocal")) { p7_profile_ConfigGlocal(gm, hmm, cfg->bg, L); }
}
else if (esl_opt_GetBoolean(go, "--uni"))
{
if (esl_opt_GetBoolean(go, "--dual")) { p7_profile_ConfigCustom (gm, hmm, cfg->bg, L, 0.0, 0.5); }
else if (esl_opt_GetBoolean(go, "--local")) { p7_profile_ConfigUnilocal (gm, hmm, cfg->bg, L); }
else if (esl_opt_GetBoolean(go, "--glocal")) { p7_profile_ConfigUniglocal(gm, hmm, cfg->bg, L); }
}
p7_profile_SetLength(gm, L);
p7_bg_SetLength(cfg->bg, L);
if (esl_opt_GetBoolean(go, "--x-no-lengthmodel")) elide_length_model(gm, cfg->bg);
/* Allocate DP matrix for <gm>.
*/
rmx = p7_refmx_Create(gm->M, L);
/* Create and configure the vectorized profile, if needed;
* and allocate its DP matrix
*/
if (esl_opt_GetBoolean(go, "--vector"))
{
om = p7_oprofile_Create(gm->M, cfg->abc, om->simd);
p7_oprofile_Convert(gm, om);
cx = p7_checkptmx_Create(gm->M, L, ESL_MBYTES(32), om->simd);
fx = p7_filtermx_Create(gm->M, om->simd);
}
/* Remaining allocation */
ESL_ALLOC(dsq, sizeof(ESL_DSQ) * (L+2));
tr = p7_trace_Create();
/* Collect scores from N random sequences of length L */
for (i = 0; i < cfg->N; i++)
{
esl_rsq_xfIID(cfg->r, cfg->bg->f, cfg->abc->K, L, dsq);
sc = eslINFINITY;
/* Vectorized implementations of Viterbi, MSV may overflow.
* In this case, they'll leave sc=eslINFINITY.
* Then we fail over to the nonvector "generic" implementation.
* That's why this next block isn't an if/else.
*/
if (esl_opt_GetBoolean(go, "--vector"))
{
if (esl_opt_GetBoolean(go, "--vit")) p7_ViterbiFilter(dsq, L, om, fx, &sc);
else if (esl_opt_GetBoolean(go, "--fwd")) p7_ForwardFilter(dsq, L, om, cx, &sc);
else if (esl_opt_GetBoolean(go, "--msv")) p7_MSVFilter (dsq, L, om, fx, &sc);
}
/* If we tried a vector calculation above but it overflowed,
* or if we're to do --generic DP calculations, sc==eslINFINITY now;
* hence the if condition here:
*/
if (sc == eslINFINITY)
{
if (esl_opt_GetBoolean(go, "--fwd")) p7_ReferenceForward(dsq, L, gm, rmx, &sc); /* any mode: dual,local,glocal; gm's config takes care of this */
else if (esl_opt_GetBoolean(go, "--vit")) p7_ReferenceViterbi(dsq, L, gm, rmx, tr, &sc); /* local-only mode. cmdline opts processing has already assured that --local set */
else if (esl_opt_GetBoolean(go, "--msv")) p7_Die("We used to be able to do a generic MSV algorithm - but no longer");
}
/* Optional: get Viterbi alignment length too. */
if (esl_opt_GetBoolean(go, "-a")) /* -a only works with Viterbi; getopts has checked this already; <tr> must be valid */
{
p7_trace_GetStateUseCounts(tr, scounts);
/* there's various ways we could counts "alignment length".
* Here we'll use the total length of model used, in nodes: M+D states.
* score vs al would gives us relative entropy / model position.
*/
/* alilens[i] = scounts[p7T_D] + scounts[p7T_I]; SRE: temporarily testing this instead */
alilens[i] = scounts[p7T_ML] + scounts[p7T_DL] + scounts[p7T_IL] +
scounts[p7T_MG] + scounts[p7T_DG] + scounts[p7T_IG];
p7_trace_Reuse(tr);
}
p7_bg_NullOne(cfg->bg, dsq, L, &nullsc);
scores[i] = (sc - nullsc) / eslCONST_LOG2;
if (cx) p7_checkptmx_Reuse(cx);
if (fx) p7_filtermx_Reuse(fx);
p7_refmx_Reuse(rmx);
}
status = eslOK;
/* deliberate flowthru */
ERROR:
if (dsq != NULL) free(dsq);
p7_checkptmx_Destroy(cx);
p7_filtermx_Destroy(fx);
p7_oprofile_Destroy(om);
p7_profile_Destroy(gm);
p7_refmx_Destroy(rmx);
p7_trace_Destroy(tr);
if (status == eslEMEM) sprintf(errbuf, "allocation failure");
return status;
}
int
main(int argc, char **argv)
{
ESL_GETOPTS *go = p7_CreateDefaultApp(options, 2, argc, argv, banner, usage);
char *hmmfile = esl_opt_GetArg(go, 1);
char *seqfile = esl_opt_GetArg(go, 2);
ESL_ALPHABET *abc = NULL;
P7_HMMFILE *hfp = NULL;
P7_HMM *hmm = NULL;
P7_BG *bg = NULL;
P7_PROFILE *gm = NULL;
P7_OPROFILE *om = NULL;
P7_OMX *ox = NULL;
P7_GMX *gx = NULL;
ESL_SQ *sq = NULL;
ESL_SQFILE *sqfp = NULL;
int format = eslSQFILE_UNKNOWN;
float msvraw, nullsc, msvscore;
float graw, gscore;
double P, gP;
int status;
/* Read in one HMM */
if (p7_hmmfile_OpenE(hmmfile, NULL, &hfp, NULL) != eslOK) p7_Fail("Failed to open HMM file %s", hmmfile);
if (p7_hmmfile_Read(hfp, &abc, &hmm) != eslOK) p7_Fail("Failed to read HMM");
/* Open sequence file for reading */
sq = esl_sq_CreateDigital(abc);
status = esl_sqfile_Open(seqfile, format, NULL, &sqfp);
if (status == eslENOTFOUND) p7_Fail("No such file.");
else if (status == eslEFORMAT) p7_Fail("Format unrecognized.");
else if (status == eslEINVAL) p7_Fail("Can't autodetect stdin or .gz.");
else if (status != eslOK) p7_Fail("Open failed, code %d.", status);
/* create default null model, then create and optimize profile */
bg = p7_bg_Create(abc);
p7_bg_SetLength(bg, sq->n);
gm = p7_profile_Create(hmm->M, abc);
p7_ProfileConfig(hmm, bg, gm, sq->n, p7_LOCAL);
om = p7_oprofile_Create(gm->M, abc);
p7_oprofile_Convert(gm, om);
/* allocate DP matrices, both a generic and an optimized one */
ox = p7_omx_Create(gm->M, 0, 0); /* one row version */
gx = p7_gmx_Create(gm->M, sq->n);
/* Useful to place and compile in for debugging:
p7_oprofile_Dump(stdout, om); dumps the optimized profile
p7_omx_SetDumpMode(stdout, ox, TRUE); makes the fast DP algorithms dump their matrices
p7_gmx_Dump(stdout, gx, p7_DEFAULT); dumps a generic DP matrix
p7_oprofile_SameMSV(om, gm);
*/
//p7_oprofile_Dump(stdout, om);
//p7_omx_SetDumpMode(stdout, ox, TRUE);
while ((status = esl_sqio_Read(sqfp, sq)) == eslOK)
{
p7_oprofile_ReconfigLength(om, sq->n);
p7_ReconfigLength(gm, sq->n);
p7_bg_SetLength(bg, sq->n);
p7_omx_GrowTo(ox, om->M, 0, sq->n);
p7_gmx_GrowTo(gx, gm->M, sq->n);
p7_MSVFilter (sq->dsq, sq->n, om, ox, &msvraw);
p7_bg_NullOne (bg, sq->dsq, sq->n, &nullsc);
msvscore = (msvraw - nullsc) / eslCONST_LOG2;
P = esl_gumbel_surv(msvscore, om->evparam[p7_MMU], om->evparam[p7_MLAMBDA]);
p7_GMSV(sq->dsq, sq->n, gm, gx, 2.0, &graw);
gscore = (graw - nullsc) / eslCONST_LOG2;
gP = esl_gumbel_surv(gscore, gm->evparam[p7_MMU], gm->evparam[p7_MLAMBDA]);
if (esl_opt_GetBoolean(go, "-1"))
{
printf("%-30s %-20s %9.2g %7.2f %9.2g %7.2f\n", sq->name, hmm->name, P, msvscore, gP, gscore);
}
else if (esl_opt_GetBoolean(go, "-P"))
{ /* output suitable for direct use in profmark benchmark postprocessors: */
printf("%g %.2f %s %s\n", P, msvscore, sq->name, hmm->name);
}
else
{
printf("target sequence: %s\n", sq->name);
printf("msv filter raw score: %.2f nats\n", msvraw);
printf("null score: %.2f nats\n", nullsc);
printf("per-seq score: %.2f bits\n", msvscore);
printf("P-value: %g\n", P);
printf("GMSV raw score: %.2f nats\n", graw);
printf("GSMV per-seq score: %.2f bits\n", gscore);
printf("GSMV P-value: %g\n", gP);
}
esl_sq_Reuse(sq);
}
/* cleanup */
esl_sq_Destroy(sq);
esl_sqfile_Close(sqfp);
p7_omx_Destroy(ox);
p7_gmx_Destroy(gx);
p7_oprofile_Destroy(om);
p7_profile_Destroy(gm);
p7_bg_Destroy(bg);
p7_hmm_Destroy(hmm);
p7_hmmfile_Close(hfp);
esl_alphabet_Destroy(abc);
esl_getopts_Destroy(go);
return 0;
}
void run_hmmer_pipeline(const char* seq) {
int index, i, status;
ESL_SQ* sq = esl_sq_CreateFrom(NULL, seq, NULL, NULL, NULL);
P7_OPROFILE *om = NULL;
P7_PROFILE *gm = NULL;
float usc, vfsc, fwdsc; /* filter scores */
float filtersc; /* HMM null filter score */
float nullsc; /* null model score */
float seqbias;
float seq_score; /* the corrected per-seq bit score */
double P;
WRAPPER_RESULT* result;
num_results = 0;
if(sq->n == 0) {
esl_sq_Destroy(sq);
return;
}
esl_sq_Digitize(abc, sq);
int n = 0;
float oasc;
for(index = 0;index < num_models;index++) {
om = models[index];
p7_omx_Reuse(oxf);
p7_omx_Reuse(oxb);
p7_omx_GrowTo(oxf, om->M, sq->n, sq->n);
p7_omx_GrowTo(oxb, om->M, sq->n, sq->n);
p7_oprofile_ReconfigLength(om, sq->n);
p7_bg_SetFilter(bg, om->M, om->compo);
p7_bg_SetLength(bg, sq->n);
//Calibrate null model
p7_bg_NullOne(bg, sq->dsq, sq->n, &nullsc);
//MSV Filter
p7_MSVFilter(sq->dsq, sq->n, om, oxf, &usc);
seq_score = (usc - nullsc) / eslCONST_LOG2;
P = esl_gumbel_surv(seq_score, om->evparam[p7_MMU], om->evparam[p7_MLAMBDA]);
if (P > f1) continue;
//Bias filter (model compo)
p7_bg_FilterScore(bg, sq->dsq, sq->n, &filtersc);
seq_score = (usc - filtersc) / eslCONST_LOG2;
P = esl_gumbel_surv(seq_score, om->evparam[p7_MMU], om->evparam[p7_MLAMBDA]);
if (P > f1) continue;
//Viterbi filter (Only do if P value from Bias is high)
if(P > f2) {
p7_ViterbiFilter(sq->dsq, sq->n, om, oxf, &vfsc);
seq_score = (vfsc - filtersc) / eslCONST_LOG2;
P = esl_gumbel_surv(seq_score, om->evparam[p7_VMU], om->evparam[p7_VLAMBDA]);
if (P > f2) continue;
}
//Get the real probability (forward)
p7_Forward(sq->dsq, sq->n, om, oxf, &fwdsc);
seq_score = (fwdsc - filtersc) / eslCONST_LOG2;
P = esl_exp_surv(seq_score, om->evparam[p7_FTAU], om->evparam[p7_FLAMBDA]);
if(hmmer_error) {
fprintf(stderr, "HMM: %s, seq: %s", om->name, seq);
hmmer_error = 0;
continue;
}
if (P > f3) continue;
//Real hit, go in to posterior decoding and alignment
p7_omx_Reuse(oxb);
p7_trace_Reuse(tr);
p7_Backward(sq->dsq, sq->n, om, oxf, oxb, NULL);
status = p7_Decoding(om, oxf, oxb, oxb);
if(status == eslOK) {
//And then trace the result
p7_OptimalAccuracy(om, oxb, oxf, &oasc);
p7_OATrace(om, oxb, oxf, tr);
} else if(status == eslERANGE) {
fprintf(stderr, "Decoding overflow on model %s\n", om->name);
gm = gmodels[index];
if(gxf == NULL) {
gxf = p7_gmx_Create(gm->M, sq->n);
gxb = p7_gmx_Create(gm->M, sq->n);
} else {
p7_gmx_GrowTo(gxf, gm->M, sq->n);
p7_gmx_GrowTo(gxb, gm->M, sq->n);
}
p7_ReconfigLength(gm, sq->n);
p7_GForward (sq->dsq, sq->n, gm, gxf, &fwdsc);
p7_GBackward(sq->dsq, sq->n, gm, gxb, NULL);
p7_GDecoding(gm, gxf, gxb, gxb);
p7_GOptimalAccuracy(gm, gxb, gxf, &oasc);
p7_GOATrace (gm, gxb, gxf, tr);
p7_gmx_Reuse(gxf);
p7_gmx_Reuse(gxb);
}
if(hmmer_error) {
fprintf(stderr, "HMM: %s, seq: %s", om->name, seq);
hmmer_error = 0;
continue;
}
result = wrapper_results[num_results];
reuse_result(result, tr->N + om->M, om->name); //We're way overallocating here, but it's hard to know at this point how much space we'll need for the alignment (plus leading and trailing gaps)
trace_into(tr, result, sq, abc, om->M);
result->bits = seq_score;
num_results++;
}
esl_sq_Destroy(sq);
}
int
main(int argc, char **argv)
{
ESL_GETOPTS *go = p7_CreateDefaultApp(options, 1, argc, argv, banner, usage);
char *hmmfile = esl_opt_GetArg(go, 1);
ESL_STOPWATCH *w = esl_stopwatch_Create();
ESL_RANDOMNESS *r = esl_randomness_CreateFast(esl_opt_GetInteger(go, "-s"));
ESL_ALPHABET *abc = NULL;
P7_HMMFILE *hfp = NULL;
P7_HMM *hmm = NULL;
P7_BG *bg = NULL;
P7_PROFILE *gm = NULL;
P7_OPROFILE *om = NULL;
P7_OMX *ox = NULL;
P7_GMX *gx = NULL;
int L = esl_opt_GetInteger(go, "-L");
int N = esl_opt_GetInteger(go, "-N");
ESL_DSQ *dsq = malloc(sizeof(ESL_DSQ) * (L+2));
int i;
float sc1, sc2;
double base_time, bench_time, Mcs;
if (p7_hmmfile_OpenE(hmmfile, NULL, &hfp, NULL) != eslOK) p7_Fail("Failed to open HMM file %s", hmmfile);
if (p7_hmmfile_Read(hfp, &abc, &hmm) != eslOK) p7_Fail("Failed to read HMM");
bg = p7_bg_Create(abc);
p7_bg_SetLength(bg, L);
gm = p7_profile_Create(hmm->M, abc);
p7_ProfileConfig(hmm, bg, gm, L, p7_LOCAL);
om = p7_oprofile_Create(gm->M, abc);
p7_oprofile_Convert(gm, om);
p7_oprofile_ReconfigLength(om, L);
if (esl_opt_GetBoolean(go, "-x")) p7_profile_SameAsMF(om, gm);
ox = p7_omx_Create(gm->M, 0, 0);
gx = p7_gmx_Create(gm->M, L);
/* Get a baseline time: how long it takes just to generate the sequences */
esl_stopwatch_Start(w);
for (i = 0; i < N; i++)
esl_rsq_xfIID(r, bg->f, abc->K, L, dsq);
esl_stopwatch_Stop(w);
base_time = w->user;
esl_stopwatch_Start(w);
for (i = 0; i < N; i++)
{
esl_rsq_xfIID(r, bg->f, abc->K, L, dsq);
p7_MSVFilter (dsq, L, om, ox, &sc1);
/* -c option: compare generic to fast score */
if (esl_opt_GetBoolean(go, "-c"))
{
p7_GMSV (dsq, L, gm, gx, 2.0, &sc2);
printf("%.4f %.4f\n", sc1, sc2);
}
/* -x option: compare generic to fast score in a way that should give exactly the same result */
if (esl_opt_GetBoolean(go, "-x"))
{
p7_GViterbi(dsq, L, gm, gx, &sc2);
sc2 /= om->scale_b;
if (om->mode == p7_UNILOCAL) sc2 -= 2.0; /* that's ~ L \log \frac{L}{L+2}, for our NN,CC,JJ */
else if (om->mode == p7_LOCAL) sc2 -= 3.0; /* that's ~ L \log \frac{L}{L+3}, for our NN,CC,JJ */
printf("%.4f %.4f\n", sc1, sc2);
}
}
esl_stopwatch_Stop(w);
bench_time = w->user - base_time;
Mcs = (double) N * (double) L * (double) gm->M * 1e-6 / (double) bench_time;
esl_stopwatch_Display(stdout, w, "# CPU time: ");
printf("# M = %d\n", gm->M);
printf("# %.1f Mc/s\n", Mcs);
free(dsq);
p7_omx_Destroy(ox);
p7_gmx_Destroy(gx);
p7_oprofile_Destroy(om);
p7_profile_Destroy(gm);
p7_bg_Destroy(bg);
p7_hmm_Destroy(hmm);
p7_hmmfile_Close(hfp);
esl_alphabet_Destroy(abc);
esl_stopwatch_Destroy(w);
esl_randomness_Destroy(r);
esl_getopts_Destroy(go);
return 0;
}
