/* Function: p7_anchors_SampleFromTrace()
* Synopsis: Make a reasonable anchor set from a trace.
*
* Purpose: Make a reasonable anchor set from trace <tr>, by
* randomly sampling a match state in each domain.
* Return the anchor set in <anch>, which will be
* reallocated if needed.
*
* <tr> must be indexed by the caller with <p7_trace_Index()>.
*
* Returns: <eslOK> on success.
*
* Throws: <eslEMEM> on reallocation failure.
*/
int
p7_anchors_SampleFromTrace(P7_ANCHORS *anch, ESL_RANDOMNESS *rng, const P7_TRACE *tr)
{
int D = tr->ndom;
int d,z,w;
int nM;
int status;
if ((status = p7_anchors_Resize(anch, D)) != eslOK) goto ERROR;
for (d = 1; d <= D; d++)
{
for (nM = 0, z = tr->tfrom[d-1]; z <= tr->tto[d-1]; z++) // P7_TRACE numbers domains 0..D-1, off by one from P7_ANCHORS
if (p7_trace_IsM(tr->st[z])) nM++;
ESL_DASSERT1(( nM ));
w = 1+esl_rnd_Roll(rng, nM); // w = 1..nM : choice of which M state to make the anchor
for ( z = tr->tfrom[d-1]; w; z++) // when w reaches 0, tr->st[z] is the M state we want to make the anchor, and we break out; there's a final z++, so the state we want ends up being z-1
if (p7_trace_IsM(tr->st[z])) w--;
ESL_DASSERT1(( p7_trace_IsM(tr->st[z-1]) )); // since the logic above is overly elegant... better doublecheck.
anch->a[d].i0 = tr->i[z-1];
anch->a[d].k0 = tr->k[z-1];
}
p7_anchor_SetSentinels(anch->a, D, tr->L, tr->M);
anch->D = D;
return eslOK;
ERROR:
return status;
}
/* Function: esl_msaweight_IDFilter()
* Synopsis: Filter by %ID.
* Incept: ER, Wed Oct 29 10:06:43 2008 [Janelia]
*
* Purpose: Constructs a new alignment by removing near-identical
* sequences from a given alignment (where identity is
* calculated *based on the alignment*).
* Does not affect the given alignment.
* Keeps earlier sequence, discards later one.
*
* Usually called as an ad hoc sequence "weighting" mechanism.
*
* Limitations:
* Unparsed Stockholm markup is not propagated into the
* new alignment.
*
* Return: <eslOK> on success, and the <newmsa>.
*
* Throws: <eslEMEM> on allocation error. <eslEINVAL> if a pairwise
* identity calculation fails because of corrupted sequence
* data. In either case, the <msa> is unmodified.
*
* Xref: squid::weight.c::FilterAlignment().
*/
int
esl_msaweight_IDFilter(const ESL_MSA *msa, double maxid, ESL_MSA **ret_newmsa)
{
int *list = NULL; /* array of seqs in new msa */
int *useme = NULL; /* TRUE if seq is kept in new msa */
int nnew; /* number of seqs in new alignment */
double ident; /* pairwise percentage id */
int i,j; /* seqs counters*/
int remove; /* TRUE if sq is to be removed */
int status;
/* Contract checks
*/
ESL_DASSERT1( (msa != NULL) );
ESL_DASSERT1( (msa->nseq >= 1) );
ESL_DASSERT1( (msa->alen >= 1) );
/* allocate */
ESL_ALLOC(list, sizeof(int) * msa->nseq);
ESL_ALLOC(useme, sizeof(int) * msa->nseq);
esl_vec_ISet(useme, msa->nseq, 0); /* initialize array */
/* find which seqs to keep (list) */
nnew = 0;
for (i = 0; i < msa->nseq; i++)
{
remove = FALSE;
for (j = 0; j < nnew; j++)
{
if (! (msa->flags & eslMSA_DIGITAL)) {
if ((status = esl_dst_CPairId(msa->aseq[i], msa->aseq[list[j]], &ident, NULL, NULL)) != eslOK) goto ERROR;
}
#ifdef eslAUGMENT_ALPHABET
else {
if ((status = esl_dst_XPairId(msa->abc, msa->ax[i], msa->ax[list[j]], &ident, NULL, NULL)) != eslOK) goto ERROR;
}
#endif
if (ident > maxid)
{
remove = TRUE;
break;
}
}
if (remove == FALSE) {
list[nnew++] = i;
useme[i] = TRUE;
}
}
if ((status = esl_msa_SequenceSubset(msa, useme, ret_newmsa)) != eslOK) goto ERROR;
free(list);
free(useme);
return eslOK;
ERROR:
if (list != NULL) free(list);
if (useme != NULL) free(useme);
return status;
}
static double
sxp_complete_binned_func(double *p, int np, void *dptr)
{
struct sxp_binned_data *data = (struct sxp_binned_data *) dptr;
ESL_HISTOGRAM *g = data->g;
double logL = 0.;
double ai, bi; /* lower, upper bounds on bin */
double lambda, tau;
int i;
double tmp;
lambda = exp(p[0]);
tau = exp(p[1]);
ESL_DASSERT1(( ! isnan(lambda) ));
ESL_DASSERT1(( ! isnan(tau) ));
for (i = g->cmin; i <= g->imax; i++) /* for each occupied bin */
{
if (g->obs[i] == 0) continue;
ai = esl_histogram_Bin2LBound(g, i);
bi = esl_histogram_Bin2UBound(g, i);
if (ai < data->mu) ai = data->mu; /* careful at leftmost bound */
tmp = esl_sxp_cdf(bi, data->mu, lambda, tau) -
esl_sxp_cdf(ai, data->mu, lambda, tau);
if (tmp == 0.) return eslINFINITY;
logL += g->obs[i] * log(tmp);
}
return -logL; /* minimizing NLL */
}
int
p7_mpas_stats_CompareAS2Trace(P7_MPAS_STATS *stats, const P7_ANCHORS *anch, const P7_TRACE *tr)
{
int ad;
int td = 0;
int anch_in_this_td = 0;
stats->anch_outside = 0;
stats->anch_unique = 0;
stats->anch_multiple = 0;
stats->dom_zero = 0;
stats->dom_one = 0;
stats->dom_multiple = 0;
/* For n domains in tr:
* they can either be hit 0 times, 1 time, or 2+ times by anchors.
* For m anchors in anch:
* they can either fall outside any domain, uniquely in a domain, or multiply in a domain.
*
* Watch out: ad (in anchor set) is 1..D; td (in trace) is 0..D-1.
*/
for (ad = 1; ad <= anch->D; ad++)
{
if (anch->a[ad].i0 < tr->sqfrom[td] || td == tr->ndom)
stats->anch_outside++;
else if (anch->a[ad].i0 >= tr->sqfrom[td] && anch->a[ad].i0 <= tr->sqto[td])
anch_in_this_td++;
else
{
/* we have to advance <td>, and try again */
if (anch_in_this_td == 0) { stats->dom_zero++; }
else if (anch_in_this_td == 1) { stats->anch_unique++; stats->dom_one++; }
else if (anch_in_this_td > 1) { stats->anch_multiple += anch_in_this_td; stats->dom_multiple++; }
anch_in_this_td = 0;
td++;
ad--; /* forces reevaluation of <ad> when we go back around; a bit hacky! */
}
}
/* we're out of anchors. If td == tr->ndom, we also know we
* handled what happened with anchors in the last <td>. But if
* td == tr->ndom-1, we haven't yet resolved what happened with final <td> yet,
* and if td is even smaller, we have some dom_zero's to count.
*/
for (; td < tr->ndom; td++)
{
if (anch_in_this_td == 0) { stats->dom_zero++; }
else if (anch_in_this_td == 1) { stats->anch_unique++; stats->dom_one++; }
else if (anch_in_this_td > 1) { stats->anch_multiple += anch_in_this_td; stats->dom_multiple++; }
anch_in_this_td = 0;
}
ESL_DASSERT1(( stats->dom_zero + stats->dom_one + stats->dom_multiple == tr->ndom ));
ESL_DASSERT1(( stats->anch_outside + stats->anch_unique + stats->anch_multiple == anch->D ));
stats->has_part2 = TRUE;
return eslOK;
}
/* Function: p7_masstrace_CountTrace()
* Synopsis: Count domain endpoints into endpoint distributions.
*
* Purpose: Given a traceback <tr>, determine if it contains a domain
* specified by the anchor point <i0>, <k0>, <st0>. If it
* doesn't, return without doing anything. If it does,
* count that domain's start/end points into the <mt>
* structure, and update the <*ntr> count by one.
*
* This function is useful in unit tests that approximate
* the mass trace calculation using a large ensemble of
* stochastic tracebacks.
*
* Before the first <_CountTrace()> call on an <mt>, you
* call <p7_masstrace_Zero()> on it to initialize it.
*
* The counts in <mt> are collected as a histogram; after
* an entire ensemble has been collected, <mt> needs to be
* converted to a cumulative distribution.
*
* A special case arises exactly at the 'midpoints' in the
* <mt> vectors <kmass> and <imass>. <kmass[k]> will be the
* start point cumulative distribution P(ka <= k) for
* k<=k0, and the end point cumulative distribution P(kb >=
* k) for k>=k0. In a cumulative distributoin, it's ok that
* kmass[k0] is defined as both the start and end value,
* because it's 1.0 in both cases. But in a histogram, we
* would have to distinguish whether kmass[k0] has seen a
* start ka versus an end kb. Instead of doing something
* special to handle this, instead we don't count kmass[k0]
* (or imass[i0]) at all; and when we convert to a
* cumulative distribution, we'll set these to 1.0.
*
* Because this is counting in single-precision floating
* point arithmetic, it can't accurately count an ensemble
* of more than about $10^7$ traces.
*
* Args: tr - trace structure
* i0 - i sequence position coord of domain anchor
* k0 - k model position coord of domain anchor
* st0 - a main model {MID}{LG} state type of domain anchor
* mt - mass trace object to count endpoints in
* ntr - updated number of traces that contain the anchor
*
* Returns: <eslOK> on success. "Success" includes ignoring a trace
* that does not contain the anchor <i0>,<k0>,<st0>. If the
* trace does contain the anchor, start/endpoint counts in
* <mt> are incremented by one, and <*ntr> is incremented
* by one.
*/
int
p7_masstrace_CountTrace(const P7_TRACE *tr, int i0, int k0, int st0, P7_MASSTRACE *mt, int *ntr)
{
int i,z0,z;
int ia,ib, ka,kb;
int foundit = FALSE;
/* Contract checks on arguments */
ESL_DASSERT1( ( i0>=1 && i0 <= tr->L) );
ESL_DASSERT1( ( k0>=1 && k0 <= tr->M) );
ESL_DASSERT1( ( p7_trace_IsMain(st0)) );
ESL_DASSERT1( ( mt->i0 == 0 || mt->i0 == i0) );
ESL_DASSERT1( ( mt->k0 == 0 || mt->k0 == k0) );
ESL_DASSERT1( ( mt->st0 == 0 || mt->st0 == st0) );
ESL_DASSERT1( ( mt->L == tr->L) );
ESL_DASSERT1( ( mt->M == tr->M) );
/* Find the anchor, if it's there. */
for (i=0, z0 = 0; z0 < tr->N; z0++)
{
if (tr->i[z0]) i = tr->i[z0]; /* update i. only emitting states have tr->i[z] set */
if (i > i0 ) break; /* failed to find anchor. */
if (i == i0 && tr->st[z0] == st0 && tr->k[z0] == k0) { foundit = TRUE; break; }
}
if (! foundit) return eslOK; /* If no anchor: successful return, ignoring this trace. */
/* Find leftmost bounds of domain */
for (ia = i0, ka = k0, z = z0; z >= 0 && tr->st[z] != p7T_B; z--)
{
if (tr->i[z]) ia = tr->i[z];
if (tr->k[z]) ka = tr->k[z];
}
ESL_DASSERT1( ( tr->st[z] == p7T_B) );
/* Find rightmost bounds of domain */
for (ib = i0, kb = k0, z = z0; z < tr->N && tr->st[z] != p7T_E; z++)
{
if (tr->i[z]) ib = tr->i[z];
if (tr->k[z]) kb = tr->k[z];
}
ESL_DASSERT1( ( tr->st[z] == p7T_E) );
/* Increment counters */
if (ka < k0) mt->kmass[ka] += 1.; /* note the guards against incrementing the overlapped start/end at k0,i0 */
if (kb > k0) mt->kmass[kb] += 1.;
if (mt->imass && ia < i0) mt->imass[ia] += 1.; /* also, guard for the optional <imass> data in <mt> */
if (mt->imass && ib > i0) mt->imass[ib] += 1.;
*ntr += 1;
/* Make sure i0,k0,st0 are set. */
mt->i0 = i0;
mt->k0 = k0;
mt->st0 = st0;
return eslOK;
}
/* guaranteed s1 >= -INFTY, p2 >= -INFTY */
int
ILogsumNI(int s1, int s2)
{
ESL_DASSERT1((s1 >= -INFTY));
ESL_DASSERT1((s2 >= -INFTY));
const int diff = s1-s2;
if (diff >= LOGSUM_TBL) return s1;
else if (diff <= -LOGSUM_TBL) return s2;
else if (diff > 0) return s1 + ilogsum_lookup[diff];
else return s2 + ilogsum_lookup[-diff];
}
static int
a2m_padding_digital(ESL_MSA *msa, char **csflag, int *nins, int ncons)
{
ESL_DSQ *ax = NULL; /* new aligned sequence - will be swapped into msa->ax[] */
ESL_DSQ gapsym = esl_abc_XGetGap(msa->abc);
int apos, cpos, spos; /* position counters for alignment 0..alen, consensus cols 0..cpos-1, sequence position 0..slen-1 */
int alen;
int icount;
int idx;
int status;
alen = ncons;
for (cpos = 0; cpos <= ncons; cpos++)
alen += nins[cpos];
ESL_ALLOC(msa->rf, sizeof(char) * (alen+1));
for (apos = 0, cpos = 0; cpos <= ncons; cpos++)
{
for (icount = 0; icount < nins[cpos]; icount++) msa->rf[apos++] = '.';
if (cpos < ncons) msa->rf[apos++] = 'x';
}
msa->rf[apos] = '\0';
for (idx = 0; idx < msa->nseq; idx++)
{
ESL_ALLOC(ax, sizeof(ESL_DSQ) * (alen + 2));
ax[0] = eslDSQ_SENTINEL;
apos = spos = 0;
for (cpos = 0; cpos <= ncons; cpos++)
{
icount = 0;
while (csflag[idx][spos] == FALSE) { ax[apos+1] = msa->ax[idx][spos+1]; apos++; spos++; icount++; }
while (icount < nins[cpos]) { ax[apos+1] = gapsym; apos++; icount++; }
if (cpos < ncons) { ax[apos+1] = msa->ax[idx][spos+1]; apos++; spos++; }
}
ESL_DASSERT1( (msa->ax[idx][spos+1] == eslDSQ_SENTINEL) );
ESL_DASSERT1( (apos == alen) );
ax[alen+1] = eslDSQ_SENTINEL;
free(msa->ax[idx]);
msa->ax[idx] = ax;
ax = NULL;
}
msa->alen = alen;
return eslOK;
ERROR:
if (ax) free(ax);
return status;
}
/* Function: p7_masstrace_Zero()
* Synopsis: Initialize cumulative endpoint distributions to zeros.
*
* Purpose: Zero the cumulative distributions in <mt>, preparing to
* collect masstrace endpoint data for a sequence of length
* <L> and a profile of length <M>.
*
* Args: mt - mass trace object to collect endpoint data in
* M - profile length
* L - sequence length
*
* Returns: <eslOK> on success.
*/
int
p7_masstrace_Zero(P7_MASSTRACE *mt, int M, int L)
{
/* contract checks / argument validation */
ESL_DASSERT1( (mt->imass == NULL || L+2 <= mt->ialloc ) );
ESL_DASSERT1( (M+2 <= mt->kalloc) );
if (mt->imass) esl_vec_FSet(mt->imass, L+2, 0.0f);
esl_vec_FSet(mt->kmass, M+2, 0.0f);
mt->L = L;
mt->M = M;
return eslOK;
}
/* Return the negative gradient at a point, determined
* numerically.
*/
static void
numeric_derivative(double *x, double *u, int n,
double (*func)(double *, int, void*),
void *prm, double relstep,
double *dx)
{
int i;
double delta;
double f1, f2;
double tmp;
for (i = 0; i < n; i++)
{
delta = fabs(u[i] * relstep);
tmp = x[i];
x[i] = tmp + delta;
f1 = (*func)(x, n, prm);
x[i] = tmp - delta;
f2 = (*func)(x, n, prm);
x[i] = tmp;
dx[i] = (-0.5 * (f1-f2)) / delta;
ESL_DASSERT1((! isnan(dx[i])));
}
}
/* Using FChoose() here would mean allocating tmp space for 2M-1 paths;
* instead we use the fact that E(i) is itself the necessary normalization
* factor, and implement FChoose's algorithm here for an on-the-fly
* calculation.
* Note that that means double-precision calculation, to be sure 0.0 <= roll < 1.0
*/
static inline int
select_e(ESL_RANDOMNESS *rng, const P7_OPROFILE *om, const P7_OMX *ox, int i, int *ret_k)
{
int Q = p7O_NQF(ox->M);
double sum = 0.0;
double roll = esl_random(rng);
double norm = 1.0 / ox->xmx[i*p7X_NXCELLS+p7X_E];
__m128 xEv = _mm_set1_ps(norm); /* all M, D already scaled exactly the same */
union { __m128 v; float p[4]; } u;
int q,r;
while (1) {
for (q = 0; q < Q; q++)
{
u.v = _mm_mul_ps(ox->dpf[i][q*3 + p7X_M], xEv);
for (r = 0; r < 4; r++) {
sum += u.p[r];
if (roll < sum) { *ret_k = r*Q + q + 1; return p7T_M;}
}
u.v = _mm_mul_ps(ox->dpf[i][q*3 + p7X_D], xEv);
for (r = 0; r < 4; r++) {
sum += u.p[r];
if (roll < sum) { *ret_k = r*Q + q + 1; return p7T_D;}
}
}
ESL_DASSERT1((sum > 0.99));
}
/*UNREACHED*/
ESL_EXCEPTION(-1, "unreached code was reached. universe collapses.");
}
/* Function: p7_coords2_hash_Create()
* Synopsis: Create a <P7_COORDS2_HASH>
*
* Purpose: Allocate and initialize a <P7_COORDS2_HASH> hash table for storing
* lots of coord2 arrays (i.e. domain annotations).
*
* The <init_*> arguments let you set non-default initial
* allocation sizes. To use the default for any of these,
* pass a 0 value. Defaults are 128 for the initial
* hashtable size <init_hashsize>; 128 for the initial
* allocation for number of keys to be stored <init_nkeyalloc>;
* and 2048 for the initial allocation for the number
* of integers to be stored in key data.
*
* In general the initialization defaults should be
* fine. All three are grown automatically as needed, as
* you add keys to the hash.
*
* "key data" means <n> <start>/<end> pairs, plus <n>
* itself: it takes 2n+1 integers to store a <P7_COORD2>
* array of length <n>.
*
* <hashsize> must be a power of 2; remember that if you
* pass a non-default value.
*
* Args: init_hashsize : initial hashtable size. Power of 2; >0.
* init_keyalloc : initial allocation for # keys. >0.
* init_calloc : initial allocation for key data. >0.
*
* Returns: pointer to the new <P7_COORDS2_HASH> object on success.
*
* Throws: <NULL> on allocation failure.
*/
P7_COORDS2_HASH *
p7_coords2_hash_Create(int32_t init_hashsize, int32_t init_nkeyalloc, int32_t init_calloc)
{
P7_COORDS2_HASH *ch = NULL;
int32_t i;
int status;
ESL_DASSERT1(( init_hashsize == 0 || (init_hashsize && ((init_hashsize & (init_hashsize-1)) == 0)))); /* hashsize is a power of 2 (bitshifting trickery) */
ESL_ALLOC(ch, sizeof(P7_COORDS2_HASH));
ch->hashtable = NULL;
ch->key_offset = NULL;
ch->nxt = NULL;
ch->cmem = NULL;
ch->nkeys = 0;
ch->cn = 0;
ch->hashsize = (init_hashsize > 0 ? init_hashsize : 128);
ch->kalloc = (init_nkeyalloc > 0 ? init_nkeyalloc : 128);
ch->calloc = (init_calloc > 0 ? init_calloc : 2048);
ESL_ALLOC(ch->hashtable, sizeof(int32_t) * ch->hashsize);
for (i = 0; i < ch->hashsize; i++) ch->hashtable[i] = -1;
ESL_ALLOC(ch->key_offset, sizeof(int32_t) * ch->kalloc);
ESL_ALLOC(ch->nxt, sizeof(int32_t) * ch->kalloc);
ESL_ALLOC(ch->cmem, sizeof(int32_t) * ch->calloc);
return ch;
ERROR:
p7_coords2_hash_Destroy(ch);
return NULL;
}
/* Function: p7_anchors_Resize()
* Synopsis: Reallocate a P7_ANCHORS object, if necessary
*
* Purpose: Make sure that <anch> can hold an array of
* at least <D> anchors.
*
* Does not alter any data that are already stored
* in <anch>, so it's safe to resize an anchor
* array that we're growing incrementally (as in
* segmental divide and conquer MPAS algorithm).
*
* D=0 is a valid argument and may occur in normal use; it
* results in a no-op, because the structure is always big
* enough to hold zero anchors.
*
* Xref: First example of a new pattern for how we
* can handle reallocation/reuse strategy,
* replacing _Reinit() and _Grow() interfaces.
* [SRE:J14/1]
*/
int
p7_anchors_Resize(P7_ANCHORS *anch, int D)
{
int nalloc;
int status;
/* Contract checks, argument validation */
ESL_DASSERT1(( anch->nalloc > 0 ));
if (D+2 <= anch->nalloc) return eslOK; // If we're big enough already, do nothing;
else if (D+2 < anch->nredline || anch->D > 0) // If we're under the redline max, or if it looks like
{ // we're building the anchor array incrementally,
nalloc = anch->nalloc; // we reallocate by doubling, trying to minimize
while (nalloc < D+2) nalloc *= 2; // the need for more reallocations soon.
} // If we're over redline AND it looks like we're
else nalloc = D+2; // starting an empty object, allocate exactly.
// Now nalloc will probably not be a multiple of two --
// but the next _Reuse() call will pull it back
// to the redline, which is.
ESL_REALLOC(anch->a, sizeof(P7_ANCHOR) * nalloc);
anch->nalloc = nalloc;
return eslOK;
ERROR:
return status;
}
/* Function: p7_filtermx_GrowTo()
* Synopsis: Resize filter DP matrix for new profile size.
*
* Purpose: Given an existing filter matrix structure <fx>,
* and the dimension <M> of a new profile that
* we're going to use (in consensus positions),
* assure that <fx> is large enough for such a
* profile; reallocate and reinitialize as needed.
*
* <p7_filtermx_Reuse(fx); p7_filtermx_GrowTo(fx, M)>
* is essentially equivalent to <p7_filtermx_Create(M)>,
* while minimizing reallocation.
*
* Returns: <eslOK> on success.
*
* Throws: <eslEMEM> on allocation failure. The state of
* <fx> is now undefined, and it should not be used.
*/
int
p7_filtermx_GrowTo_avx(P7_FILTERMX *fx, int allocM)
{
int status;
/* Contract checks / argument validation */
ESL_DASSERT1( (allocM >= 1 && allocM <= 100000) );
#ifdef HAVE_AVX2
/* is it already big enough? */
if (allocM <= fx->allocM_AVX) return eslOK;
/* if not, grow it */
ESL_REALLOC(fx->dp_mem_AVX, (sizeof(__m256i) * (p7F_NSCELLS * P7_NVW_AVX(allocM))) + (p7_VALIGN_AVX-1));
fx->allocM_AVX = allocM;
fx->dp_AVX = (__m256i *) ( (unsigned long int) ( (char *) fx->dp_mem_AVX + (p7_VALIGN_AVX-1)) & p7_VALIMASK_AVX);
return eslOK;
ERROR:
return status;
#endif //HAVE_AVX2
#ifndef HAVE_AVX2
return eslENORESULT;
#endif
}
/* Function: p7_filtermx_GrowTo()
* Synopsis: Resize filter DP matrix for new profile size.
*
* Purpose: Given an existing filter matrix structure <fx>,
* and the dimension <M> of a new profile that
* we're going to use (in consensus positions),
* assure that <fx> is large enough for such a
* profile; reallocate and reinitialize as needed.
*
* <p7_filtermx_Reuse(fx); p7_filtermx_GrowTo(fx, M)>
* is essentially equivalent to <p7_filtermx_Create(M)>,
* while minimizing reallocation.
*
* Returns: <eslOK> on success.
*
* Throws: <eslEMEM> on allocation failure. The state of
* <fx> is now undefined, and it should not be used.
*/
int
p7_filtermx_GrowTo_neon64(P7_FILTERMX *fx, int allocM)
{
#ifdef HAVE_NEON64
int status;
/* Contract checks / argument validation */
ESL_DASSERT1( (allocM >= 1 && allocM <= 100000) );
if (allocM <= fx->allocM) return eslOK;
/* if not, grow it */
ESL_REALLOC(fx->dp_mem, (sizeof(esl_neon_128i_t) * (p7F_NSCELLS * P7_NVW(allocM))) + (p7_VALIGN-1));
fx->allocM = allocM;
fx->dp = (esl_neon_128i_t *) ( (unsigned long int) ( (char *) fx->dp_mem + (p7_VALIGN-1)) & p7_VALIMASK);
return eslOK;
ERROR:
return status;
#endif //HAVE_NEON64
#ifndef HAVE_NEON64
return eslENORESULT;
#endif
}
static int
a2m_padding_text(ESL_MSA *msa, char **csflag, int *nins, int ncons)
{
char *aseq = NULL; /* new aligned sequence - will be swapped into msa->aseq[] */
int apos, cpos, spos; /* position counters for alignment 0..alen, consensus cols 0..cpos-1, sequence position 0..slen-1 */
int alen;
int icount;
int idx;
int status;
alen = ncons;
for (cpos = 0; cpos <= ncons; cpos++)
alen += nins[cpos];
ESL_ALLOC(msa->rf, sizeof(char) * (alen+1));
for (apos = 0, cpos = 0; cpos <= ncons; cpos++)
{
for (icount = 0; icount < nins[cpos]; icount++) msa->rf[apos++] = '.';
if (cpos < ncons) msa->rf[apos++] = 'x';
}
msa->rf[apos] = '\0';
for (idx = 0; idx < msa->nseq; idx++)
{
ESL_ALLOC(aseq, sizeof(char) * (alen + 1));
apos = spos = 0;
for (cpos = 0; cpos <= ncons; cpos++)
{
icount = 0;
while (csflag[idx][spos] == FALSE) { aseq[apos] = msa->aseq[idx][spos]; apos++; spos++; icount++; }
while (icount < nins[cpos]) { aseq[apos] = '.'; apos++; icount++; }
if (cpos < ncons) { aseq[apos] = msa->aseq[idx][spos]; apos++; spos++; }
}
ESL_DASSERT1( (msa->aseq[idx][spos] == '\0') );
ESL_DASSERT1( (apos == alen) );
aseq[alen] = '\0';
free(msa->aseq[idx]);
msa->aseq[idx] = aseq;
aseq = NULL;
}
msa->alen = alen;
return eslOK;
ERROR:
if (aseq) free(aseq);
return status;
}
/* Function: p7_filtermx_DumpMFRow()
* Synopsis: Dump one row from MSV version of a DP matrix.
*
* Purpose: Dump current row of MSV calculations from DP matrix <fx>
* for diagnostics, and include the values of specials
* <xE>, etc. The index <rowi> for the current row is used
* as a row label. This routine has to be specialized for
* the layout of the MSVFilter() row, because it's all
* match scores dp[0..q..Q-1], rather than triplets of
* M,D,I.
*
* If <rowi> is 0, print a header first too.
*
* The output format is coordinated with <p7_refmx_Dump()> to
* facilitate comparison to a known answer.
*
* This also works for an SSV filter row, for SSV implementations
* that use a single row of DP memory (like <_longtarget>).
* The Knudsen assembly code SSV does not use any RAM.
*
* Returns: <eslOK> on success.
*
* Throws: <eslEMEM> on allocation failure.
*/
int
p7_filtermx_DumpMFRow_neon64(const P7_FILTERMX *fx, int rowi, uint8_t xE, uint8_t xN, uint8_t xJ, uint8_t xB, uint8_t xC)
{
#ifdef HAVE_NEON64
int Q = P7_NVB(fx->M); /* number of vectors in the MSV row */
uint8_t *v = NULL; /* array of scores after unstriping them */
int q,z,k;
union { esl_neon_128i_t v; uint8_t i[16]; } tmp;
int status;
ESL_DASSERT1( (fx->type == p7F_MSVFILTER || fx->type == p7F_SSVFILTER) );
/* We'll unstripe the whole row; then print it in its normal order. */
ESL_ALLOC(v, sizeof(unsigned char) * ((Q*16)+1));
v[0] = 0;
/* Header (if we're on the 0th row) */
if (rowi == 0)
{
fprintf(fx->dfp, " ");
for (k = 0; k <= fx->M; k++) fprintf(fx->dfp, "%3d ", k);
fprintf(fx->dfp, "%3s %3s %3s %3s %3s\n", "E", "N", "J", "B", "C");
fprintf(fx->dfp, " ");
for (k = 0; k <= fx->M+5; k++) fprintf(fx->dfp, "%3s ", "---");
fprintf(fx->dfp, "\n");
}
/* Unpack and unstripe, then print M's. */
for (q = 0; q < Q; q++) {
tmp.v = fx->dp[q];
for (z = 0; z < 16; z++) v[q+Q*z+1] = tmp.i[z];
}
fprintf(fx->dfp, "%4d M ", rowi);
for (k = 0; k <= fx->M; k++) fprintf(fx->dfp, "%3d ", v[k]);
/* The specials */
fprintf(fx->dfp, "%3d %3d %3d %3d %3d\n", xE, xN, xJ, xB, xC);
/* I's are all 0's; print just to facilitate comparison to refmx. */
fprintf(fx->dfp, "%4d I ", rowi);
for (k = 0; k <= fx->M; k++) fprintf(fx->dfp, "%3d ", 0);
fprintf(fx->dfp, "\n");
/* D's are all 0's too */
fprintf(fx->dfp, "%4d D ", rowi);
for (k = 0; k <= fx->M; k++) fprintf(fx->dfp, "%3d ", 0);
fprintf(fx->dfp, "\n\n");
free(v);
return eslOK;
ERROR:
free(v);
return status;
#endif //HAVE_NEON64
#ifndef HAVE_NEON64
return eslENORESULT;
#endif
}
/* guaranteed s1 >= -INFTY, s2 >= -INFTY */
int
ILogsumNI(int s1, int s2)
{
ESL_DASSERT1((s1 > -INFTY));
ESL_DASSERT1((s2 > -INFTY));
/*assert(s1 > -INFTY);
assert(s2 > -INFTY);*/
const int max = ESL_MAX(s1, s2);
const int min = ESL_MIN(s1, s2);
return ((max-min) >= LOGSUM_TBL) ? max : max + ilogsum_lookup[max-min];
/* about 10% slower
if(s1 > s2)
return ((s1-s2) >= LOGSUM_TBL) ? s1 : s1 + ilogsum_lookup[s1-s2];
else
return ((s2-s1) >= LOGSUM_TBL) ? s2 : s2 + ilogsum_lookup[s2-s1];
*/
}
/* guaranteed s1 >= -INFTY, s2 >= -INFTY */
int
ILogsumNI_diff(int s1a, int s1b, int s2a, int s2b, int db)
{
/* db = s1b - s2b */
ESL_DASSERT1((s1a > -INFTY));
ESL_DASSERT1((s1b > -INFTY));
ESL_DASSERT1((s2a > -INFTY));
ESL_DASSERT1((s2b > -INFTY));
/*const int d = s1a-s2a+db;
if (d >= LOGSUM_TBL) return s1a + s1b;
else if (d > 0) return s1a + s1b + ilogsum_lookup[d];
else if (d <= -LOGSUM_TBL) return s2a + s2b;
else return s2a + s2b + ilogsum_lookup[-d];*/
const int d = s1a-s2a+db;
if(d > 0)
return (d >= LOGSUM_TBL) ? s1a + s1b : s1a + s1b + ilogsum_lookup[d];
else
return (d <= LOGSUM_TBL) ? s2a + s2b : s2a + s2b + ilogsum_lookup[-d];
}
/* Function: esl_msaweight_PB()
* Synopsis: PB (position-based) weights.
* Incept: SRE, Sun Nov 5 08:59:28 2006 [Janelia]
*
* Purpose: Given a multiple alignment <msa>, calculate sequence
* weights according to the position-based weighting
* algorithm (Henikoff and Henikoff, JMB 243:574-578,
* 1994). These weights are stored internally in the <msa>
* object, replacing any weights that may have already been
* there. Weights are $\geq 0$ and they sum to <msa->nseq>.
*
* The <msa> may be in either digitized or text mode.
* Digital mode is preferred, so that the algorithm
* deals with degenerate residue symbols properly.
*
* The Henikoffs' algorithm does not give rules for dealing
* with gaps or degenerate residue symbols. The rule here
* is to ignore them. This means that longer sequences
* initially get more weight; hence a "double
* normalization" in which the weights are first divided by
* sequence length in canonical residues (to compensate for
* that effect), then normalized to sum to nseq.
*
* An advantage of the PB method is efficiency.
* It is $O(1)$ in memory and $O(NL)$ time, for an alignment of
* N sequences and L columns. This makes it a good method
* for ad hoc weighting of very deep alignments.
*
* When the alignment is in simple text mode, IUPAC
* degenerate symbols are not dealt with correctly; instead,
* the algorithm simply uses the 26 letters as "residues"
* (case-insensitively), and treats all other residues as
* gaps.
*
* Returns: <eslOK> on success, and the weights inside <msa> have been
* modified.
*
* Throws: <eslEMEM> on allocation error, in which case <msa> is
* returned unmodified.
*
* Xref: [Henikoff94b]; squid::weight.c::PositionBasedWeights().
*/
int
esl_msaweight_PB(ESL_MSA *msa)
{
int *nres = NULL; /* counts of each residue observed in a column */
int ntotal; /* number of different symbols observed in a column */
int rlen; /* number of residues in a sequence */
int idx, pos, i;
int K; /* alphabet size */
int status;
/* Contract checks
*/
ESL_DASSERT1( (msa->nseq >= 1) );
ESL_DASSERT1( (msa->alen >= 1) );
if (msa->nseq == 1) { msa->wgt[0] = 1.0; return eslOK; }
/* Initialize
*/
if (! (msa->flags & eslMSA_DIGITAL))
{ ESL_ALLOC_WITH_TYPE(nres, int*, sizeof(int) * 26); K = 26; }
/* Function: esl_sxp_cdf()
*
* Purpose: Calculates the cumulative distribution function for the
* stretched exponential pdf, $P(X \leq x)$, given
* quantile <x>, offset <mu>, and parameters <lambda> and <tau>.
*/
double
esl_sxp_cdf(double x, double mu, double lambda, double tau)
{
double y = lambda * (x-mu);
double val;
if (x <= mu) return 0.;
esl_stats_IncompleteGamma(1/tau, exp(tau * log(y)), &val, NULL);
ESL_DASSERT1 (( !isnan(val)));
return val;
}
/* jukescantor()
*
* The generalized Jukes/Cantor distance calculation.
* Given <n1> identities and <n2> differences, for a
* base alphabet size of <alphabet_size> (4 or 20);
* calculate J/C distance in substitutions/site and
* return it in <ret_distance>; calculate large-sample
* variance and return it in <ret_variance>.
*
* Returns <eslEDIVZERO> if there are no data (<n1+n2=0>).
*/
static int
jukescantor(int n1, int n2, int alphabet_size, double *opt_distance, double *opt_variance)
{
int status;
double D, K, N;
double x;
double distance, variance;
ESL_DASSERT1( (n1 >= 0) );
ESL_DASSERT1( (n2 >= 0) );
ESL_DASSERT1( (alphabet_size >= 0) );
if (n1+n2 == 0) { status = eslEDIVZERO; goto ERROR; }
K = (double) alphabet_size;
D = (double) n2 / (double) (n1+n2);
N = (double) (n1+n2);
x = 1. - D * K/(K-1.);
if (x <= 0.)
{
distance = HUGE_VAL;
variance = HUGE_VAL;
}
else
{
distance = -log(x) * K/(K-1);
variance = exp( 2.*K*distance/(K-1) ) * D * (1.-D) / N;
}
if (opt_distance != NULL) *opt_distance = distance;
if (opt_variance != NULL) *opt_variance = variance;
return eslOK;
ERROR:
if (opt_distance != NULL) *opt_distance = HUGE_VAL;
if (opt_variance != NULL) *opt_variance = HUGE_VAL;
return status;
}
/* Function: p7_masstrace_FinishCount()
* Synopsis: Convert counted histograms to cumulative endpoint prob distributions.
*
* Purpose: We've finished collecting endpoints from traces with
* <_CountTrace()> in <mt>, <ntr> of which had the
* specified domain anchor; now convert the counts to
* <mt>'s cumulative probability distributions.
*
* Args: mt - mass trace object we've collected endpoint counts in
* ntr - number of traces we counted into <mt> that contained the domain anchor
*
* Returns: <eslOK> on success; <mt> is now a valid <P7_MASSTRACE> object
* containing envelope endpoint cumulative probability distributions.
*/
int
p7_masstrace_FinishCount(P7_MASSTRACE *mt, int ntr)
{
int i,k;
ESL_DASSERT1( (ntr > 0) );
ESL_DASSERT1( (mt->i0) );
ESL_DASSERT1( (mt->k0) );
ESL_DASSERT1( (p7_trace_IsMain(mt->st0)) );
if (mt->imass)
{
for (i = 1; i < mt->i0; i++) mt->imass[i] += mt->imass[i-1];
for (i = mt->L; i > mt->i0; i--) mt->imass[i] += mt->imass[i+1];
esl_vec_FScale(mt->imass+1, mt->L, 1./(float) ntr);
mt->imass[mt->i0] = 1.;
}
for (k = 1; k < mt->k0; k++) mt->kmass[k] += mt->kmass[k-1];
for (k = mt->M; k > mt->k0; k--) mt->kmass[k] += mt->kmass[k+1];
esl_vec_FScale(mt->kmass+1, mt->M, 1./(float) ntr);
mt->kmass[mt->k0] = 1.;
return eslOK;
}
/* Function: esl_msaweight_BLOSUM()
* Synopsis: BLOSUM weights.
* Incept: SRE, Sun Nov 5 09:52:41 2006 [Janelia]
*
* Purpose: Given a multiple sequence alignment <msa> and an identity
* threshold <maxid>, calculate sequence weights using the
* BLOSUM algorithm (Henikoff and Henikoff, PNAS
* 89:10915-10919, 1992). These weights are stored
* internally in the <msa> object, replacing any weights
* that may have already been there. Weights are $\geq 0$
* and they sum to <msa->nseq>.
*
* The algorithm does a single linkage clustering by
* fractional id, defines clusters such that no two clusters
* have a pairwise link $\geq$ <maxid>), and assigns
* weights of $\frac{1}{M_i}$ to each of the $M_i$
* sequences in each cluster $i$. The <maxid> threshold
* is a fractional pairwise identity, in the range
* $0..1$.
*
* The <msa> may be in either digitized or text mode.
* Digital mode is preferred, so that the pairwise identity
* calculations deal with degenerate residue symbols
* properly.
*
* Returns: <eslOK> on success, and the weights inside <msa> have been
* modified.
*
* Throws: <eslEMEM> on allocation error. <eslEINVAL> if a pairwise
* identity calculation fails because of corrupted sequence
* data. In either case, the <msa> is unmodified.
*
* Xref: [Henikoff92]; squid::weight.c::BlosumWeights().
*/
int
esl_msaweight_BLOSUM(ESL_MSA *msa, double maxid)
{
int *c = NULL; /* cluster assignments for each sequence */
int *nmem = NULL; /* number of seqs in each cluster */
int nc; /* number of clusters */
int i; /* loop counter */
int status;
/* Contract checks
*/
ESL_DASSERT1( (maxid >= 0. && maxid <= 1.) );
ESL_DASSERT1( (msa->nseq >= 1) );
ESL_DASSERT1( (msa->alen >= 1) );
if (msa->nseq == 1) { msa->wgt[0] = 1.0; return eslOK; }
if ((status = esl_msacluster_SingleLinkage(msa, maxid, &c, NULL, &nc)) != eslOK) goto ERROR;
ESL_ALLOC(nmem, sizeof(int) * nc);
esl_vec_ISet(nmem, nc, 0);
for (i = 0; i < msa->nseq; i++) nmem[c[i]]++;
for (i = 0; i < msa->nseq; i++) msa->wgt[i] = 1. / (double) nmem[c[i]];
/* Make weights normalize up to nseq, and return.
*/
esl_vec_DNorm(msa->wgt, msa->nseq);
esl_vec_DScale(msa->wgt, msa->nseq, (double) msa->nseq);
msa->flags |= eslMSA_HASWGTS;
free(nmem);
free(c);
return eslOK;
ERROR:
if (c != NULL) free(c);
if (nmem != NULL) free(nmem);
return status;
}
static int
do_by_windows(ESL_GENCODE *gcode, ESL_GENCODE_WORKSTATE *wrk, ESL_SQFILE *sqfp)
{
ESL_SQ *sq = esl_sq_CreateDigital(gcode->nt_abc);
int windowsize = 4092; // can be any value, but a multiple of 3 makes most sense. windowsize can be +/-; + means reading top strand; - means bottom strand.
int contextsize = 2; // contextsize (adjacent window overlap) must be 2, or translation won't work properly.
int wstatus;
ESL_DASSERT1(( windowsize % 3 == 0 ));
while (( wstatus = esl_sqio_ReadWindow(sqfp, contextsize, windowsize, sq)) != eslEOF)
{
if (wstatus == eslEOD)
{
if ( (windowsize > 0 && wrk->do_watson) || (windowsize < 0 && wrk->do_crick))
esl_gencode_ProcessEnd(wrk, sq);
if (windowsize > 0 && ! wrk->do_crick) { esl_sq_Reuse(sq); continue; } // Don't switch to revcomp if we don't need do. Allows -W --watson to work on nonrewindable streams
if (windowsize < 0) esl_sq_Reuse(sq); // Do not Reuse the sq on the switch from watson to crick; ReadWindow needs sq->L
windowsize = -windowsize; // switch to other strand.
continue;
}
else if (wstatus == eslEFORMAT) esl_fatal("Parsing failed in sequence file %s:\n%s", sqfp->filename, esl_sqfile_GetErrorBuf(sqfp));
else if (wstatus == eslEINVAL) esl_fatal("Invalid residue(s) found in sequence file %s\n%s", sqfp->filename, esl_sqfile_GetErrorBuf(sqfp));
else if (wstatus != eslOK) esl_fatal("Unexpected error %d reading sequence file %s", wstatus, sqfp->filename);
/* If we're the first window in this input DNA sequence
* (or the first window in its revcomp), then initialize.
* sq->C is the actual context overlap; 0=1st window; 2 (i.e. C)= subsequent.
*/
if (sq->C == 0)
{
if (sq->n < 3) continue; // DNA sequence too short; skip it, don't even bother to revcomp, go to next sequence.
if ( (windowsize > 0 && wrk->do_watson) || (windowsize < 0 && wrk->do_crick))
esl_gencode_ProcessStart(gcode, wrk, sq);
}
if ( (windowsize > 0 && wrk->do_watson) || (windowsize < 0 && wrk->do_crick))
esl_gencode_ProcessPiece(gcode, wrk, sq);
}
esl_sq_Destroy(sq);
return eslOK;
}
/* Function: p7_filtermx_Create()
* Synopsis: Create a one-row DP matrix for MSV, VF.
*
* Purpose: Allocate a reusable, resizeable one-row <P7_FILTERMX>
* suitable for MSV and Viterbi filter calculations on
* query profiles of up to <allocM> consensus positions.
*
* <allocM> must be $\leq$ 100,000. This is an H3 design
* limit.
*
* Args: allocM - initial allocation size, in profile positions. (>=1, <=100000)
*
* Returns: ptr to new <P7_FILTERMX>
*
* Throws: <NULL> on allocation failure.
*/
P7_FILTERMX *
p7_filtermx_Create_neon64(int allocM)
{
#ifdef HAVE_NEON64
P7_FILTERMX *fx = NULL;
int status;
/* Contract checks / argument validation */
ESL_DASSERT1( (allocM >= 1 && allocM <= 100000) );
ESL_ALLOC(fx, sizeof(P7_FILTERMX));
fx->simd = NEON64;
fx->M = 0;
fx->dp = NULL;
fx->dp_mem = NULL;
fx->allocM = 0;
fx->type = p7F_NONE;
#ifdef p7_DEBUGGING
fx->do_dumping= FALSE;
fx->dfp = NULL;
#endif
// ISA we're using
/* 16B per vector * (MDI)states * ~M/4 vectors + alignment slop */
ESL_ALLOC(fx->dp_mem, (sizeof(esl_neon_128i_t) * p7F_NSCELLS * P7_NVW(allocM)) + (p7_VALIGN-1));
fx->allocM = allocM;
/* Manual memory alignment incantation: */
fx->dp = (esl_neon_128i_t *) ( (unsigned long int) ( (char *) fx->dp_mem + (p7_VALIGN-1) ) & p7_VALIMASK);
return fx;
ERROR:
p7_filtermx_Destroy(fx);
return NULL;
#endif //HAVE_NEON64
#ifndef HAVE_NEON64
return NULL;
#endif
}
/* Function: p7_filtermx_Create()
* Synopsis: Create a one-row DP matrix for MSV, VF.
*
* Purpose: Allocate a reusable, resizeable one-row <P7_FILTERMX>
* suitable for MSV and Viterbi filter calculations on
* query profiles of up to <allocM> consensus positions.
*
* <allocM> must be $\leq$ 100,000. This is an H3 design
* limit.
*
* Args: allocM - initial allocation size, in profile positions. (>=1, <=100000)
*
* Returns: ptr to new <P7_FILTERMX>
*
* Throws: <NULL> on allocation failure.
*/
P7_FILTERMX *
p7_filtermx_Create_avx(int allocM)
{
P7_FILTERMX *fx = NULL;
int status;
/* Contract checks / argument validation */
ESL_DASSERT1( (allocM >= 1 && allocM <= 100000) );
#ifdef HAVE_AVX2
ESL_ALLOC(fx, sizeof(P7_FILTERMX));
fx->simd = AVX;
fx->dp_AVX = NULL;
fx->dp_mem_AVX = NULL;
fx->allocM_AVX = 0;
fx->type = p7F_NONE;
#ifdef p7_DEBUGGING
fx->do_dumping= FALSE;
fx->dfp = NULL;
#endif
/* 32B per vector * (MDI)states * ~M/4 vectors + alignment slop */
ESL_ALLOC(fx->dp_mem_AVX, (sizeof(__m256i) * p7F_NSCELLS * P7_NVW_AVX(allocM)) + (p7_VALIGN_AVX-1));
fx->allocM_AVX = allocM;
/* Manual memory alignment incantation: */
fx->dp_AVX = (__m256i *) ( (unsigned long int) ( (char *) fx->dp_mem_AVX + (p7_VALIGN_AVX-1) ) & p7_VALIMASK_AVX);
return fx;
ERROR:
p7_filtermx_Destroy(fx);
return NULL;
#endif //HAVE_AVX2
#ifndef HAVE_AVX2
return NULL;
#endif
}
/* wei_binned_func():
* Returns the negative log likelihood of a binned data sample,
* in the API of the conjugate gradient descent optimizer in esl_minimizer.
*/
static double
wei_binned_func(double *p, int nparam, void *dptr)
{
struct wei_binned_data *data = (struct wei_binned_data *) dptr;
ESL_HISTOGRAM *h = data->h;
double lambda, tau;
double logL;
double ai,bi;
int i;
double tmp;
/* Unpack what the optimizer gave us.
*/
lambda = exp(p[0]); /* see below for c.o.v. notes */
tau = exp(p[1]);
logL = 0.;
for (i = h->cmin; i <= h->imax; i++)
{
if (h->obs[i] == 0) continue;
ai = esl_histogram_Bin2LBound(h,i);
bi = esl_histogram_Bin2UBound(h,i);
if (ai < data->mu) ai = data->mu;
tmp = esl_wei_cdf(bi, data->mu, lambda, tau) -
esl_wei_cdf(ai, data->mu, lambda, tau);
/* for cdf~1.0, numerical roundoff error can create tmp<0 by a
* teensy amount; tolerate that, but catch anything worse */
ESL_DASSERT1( (tmp + 1e-7 > 0.));
if (tmp <= 0.) return eslINFINITY;
logL += h->obs[i] * log(tmp);
}
return -logL; /* goal: minimize NLL */
}
/* Function: esl_mixgev_FitComplete()
*
* Purpose: Given <n> observed data values <x[0..n-1]>, and
* an initial guess at a mixture GEV fit to those data
* <mg>, use conjugate gradient descent to perform
* a locally optimal maximum likelihood mixture
* GEV parameter fit to the data.
*
* To obtain a reasonable initial guess for <mg>,
* see <esl_mixgev_FitGuess()>.
*
* Args: x - observed data, <x[0..n-1]>.
* n - number of samples in <x>
* mg - mixture GEV to estimate, w/ params set to
* an initial guess.
*
* Returns: <eslOK> on success, and <mg> contains local
* ML estimate for mixture GEV parameters.
*
* Throws: <eslEMEM> on allocation error, and <mg> is unchanged
* from its initial state.
*/
int
esl_mixgev_FitComplete(double *x, int n, ESL_MIXGEV *mg)
{
struct mixgev_data data;
int status;
double *p = NULL;
double *u = NULL;
double *wrk = NULL;
double tol;
int np;
double fx;
int k;
int i;
tol = 1e-6;
/* Determine number of free parameters and allocate
*/
np = mg->K-1; /* K-1 mix coefficients free */
for (k = 0; k < mg->K; k++)
np += (mg->isgumbel[k])? 2 : 3;
ESL_ALLOC(p, sizeof(double) * np);
ESL_ALLOC(u, sizeof(double) * np);
ESL_ALLOC(wrk, sizeof(double) * np * 4);
/* Copy shared info into the "data" structure
*/
data.x = x;
data.n = n;
data.wrk = wrk;
data.mg = mg;
/* From mg, create the parameter vector.
*/
mixgev_pack_paramvector(p, np, mg);
/* Define the step size vector u.
*/
i = 0;
for (k = 1; k < mg->K; k++) u[i++] = 1.0;
for (k = 0; k < mg->K; k++)
{
u[i++] = 1.0;
u[i++] = 1.0;
if (! mg->isgumbel[k]) u[i++] = 0.02;
}
ESL_DASSERT1( (np == i) );
/* Feed it all to the mighty optimizer.
*/
status = esl_min_ConjugateGradientDescent(p, u, np, &mixgev_complete_func, NULL,
(void *) (&data), tol, wrk, &fx);
if (status != eslOK) goto ERROR;
/* Convert the final parameter vector back to a mixture GEV
*/
mixgev_unpack_paramvector(p, np, mg);
free(p);
free(u);
free(wrk);
return eslOK;
ERROR:
if (p != NULL) free(p);
if (u != NULL) free(u);
if (wrk != NULL) free(wrk);
return status;
}
/* Function: esl_keyhash_CreateCustom()
* Synopsis: Allocate a new keyhash with customized initial allocations.
*
* Purpose: Create a new hash table, initially allocating for
* a hash table of size <hashsize> entries, <kalloc>
* keys, and a total key string length of <salloc>.
* <hashsize> must be a power of 2, and all allocations
* must be $\geq 0$.
*
* The object will still expand as needed, so the reason to
* use a customized allocation is when you're trying to
* minimize memory footprint and you expect your keyhash to
* be smaller than the default (of up to 128 keys, of total
* length up to 2048).
*
* Throws: <NULL> on allocation failure.
*/
ESL_KEYHASH *
esl_keyhash_CreateCustom(uint32_t hashsize, int kalloc, int salloc)
{
ESL_DASSERT1((hashsize && ((hashsize & (hashsize-1)) == 0))); /* hashsize is a power of 2 (bitshifting trickery) */
return keyhash_create(hashsize, kalloc, salloc);
}
static int
profillic_esl_msafile_profile_Read(ESLX_MSAFILE *afp, ESL_MSA **ret_msa, ProfileType * profile_ptr )
{
/// \note Right now this isn't actually using the open file pointer; for convenience I just use the profile.fromFile( <filename> ) method.
/// \todo Use convenience fns in esl_buffer.h; see eg hmmer-3.1/easel/esl_msafile_stockholm.c for examples...
ESL_MSA *msa = NULL;
string profile_string;
char *buf;
long len;
int seqidx;
int status;
char errmsg2[eslERRBUFSIZE];
ESL_DASSERT1((afp->format == eslMSAFILE_PROFILLIC));
const char * const seqname = "Galosh Profile Consensus";
const char * const msaname = "Galosh Profile";
uint32_t profile_length;
galosh::Sequence<typename ProfileType::ProfileResidueType> consensus_sequence;
stringstream tmp_consensus_output_stream;
uint32_t pos_i;
if (profile_ptr == NULL) { ESL_EXCEPTION(eslEINCONCEIVABLE, "profile_ptr is NULL in profillic_esl_msafile_profile_Read(..)!"); }
//if (feof(afp->bf->fp)) { status = eslEOF; goto ERROR; }
afp->errmsg[0] = '\0';
// Read in the galosh profile (from profillic)
//fseek( afp->bf->fp, 0, SEEK_END ); // go to the end
//len = afp->bf->ftell( afp->bf->fp ); // get the position at the end (length)
//fseek( afp->bf->fp, 0, SEEK_SET ); // go to the beginning again.
//ESL_ALLOC_CPP( char, buf, sizeof( char ) * len ); //malloc buffer
//fread( buf, len, 1, afp->bf->fp ); //read into buffer
//profile_string = buf;
//profile_ptr->fromString( profile_string );
profile_ptr->fromFile( afp->bf->filename );
//if (buf) free(buf);
// \todo WHY WON'T THIS WORK? See HACKs in profillic-hmmbuild.cpp to work around it.
//fseek( afp->bf->fp, 0, SEEK_END ); // go to the end (to signal there's no more profiles in the file, the next time we come to this function)
// Calculate the consensus sequence.
profile_length = profile_ptr->length();
consensus_sequence.reinitialize( profile_length );
for( pos_i = 0; pos_i < profile_length; pos_i++ ) {
consensus_sequence[ pos_i ] =
( *profile_ptr )[ pos_i ][ galosh::Emission::Match ].maximumValueType();
}
tmp_consensus_output_stream << consensus_sequence;
/* Allocate a growable MSA, and auxiliary parse data coupled to the MSA allocation */
#ifdef eslAUGMENT_ALPHABET
if (afp->abc && (msa = esl_msa_CreateDigital(afp->abc, 16, -1)) == NULL) { status = eslEMEM; goto ERROR; }
#endif
if (! afp->abc && (msa = esl_msa_Create( 16, -1)) == NULL) { status = eslEMEM; goto ERROR; }
// Set first-and-only seq to the consensus. This should set sqlen[0] to the profile's length and set ax to have length 1 and ax[0] to be the sequence itself. Also msa->sqname[0] to the "name" of that consensus sequence.
/* if nec, make room for the new seq */
if (msa->nseq >= msa->sqalloc && (status = esl_msa_Expand(msa)) != eslOK) return status;
seqidx = msa->nseq; // 0
msa->nseq++; // = 1
status = esl_strdup(seqname, -1, &(msa->sqname[seqidx]));
// NOTE: Could add description of this "sequence" here, using esl_msa_SetSeqDescription(msa, seqidx, desc).
#ifdef eslAUGMENT_ALPHABET
if (msa->flags & eslMSA_DIGITAL)
{
// NOTE (profillic): There was a bug in this; it had said .."esl_abc_dsqcat(msa->abc, " where it should have said .."esl_abc_dsqcat(msa->abc->inmap, "
if((status = esl_abc_dsqcat(msa->abc->inmap, &(msa->ax[seqidx]), &(msa->sqlen[seqidx]), tmp_consensus_output_stream.str().c_str(), profile_length)) != eslOK) {
/* invalid char(s), get informative error message */
if (esl_abc_ValidateSeq(msa->abc, tmp_consensus_output_stream.str().c_str(), profile_length, afp->errmsg) != eslOK)
ESL_XFAIL(eslEFORMAT, errmsg2, "%s (line %d): %s", msa->sqname[0], afp->linenumber, afp->errmsg);
}
}
#endif
if (! (msa->flags & eslMSA_DIGITAL))
{
status = esl_strcat(&(msa->aseq[seqidx]), 0, tmp_consensus_output_stream.str().c_str(), profile_length);
msa->sqlen[seqidx] = profile_length;
}
msa->alen = profile_length;
/// \todo OR read in a fasta file of sequences too.
/// \todo (Optional?) Set msa->name to the name of the profile (file?)
esl_strdup(msaname, -1, &(msa->name));
/// \todo make sure eslMSA_HASWGTS is FALSE .. OR set it to TRUE and set msa->wgt[idx] to 1.0.
/// \note Could have secondary structure (per sequence) too. msa->ss[0]. msa->sslen[0] should be the same as msa->sqlen[0].
/// \todo Investigate what msa->sa and msa->pp are for.
/* Give the newly parsed MSA a good
* going-over, and finalize the fields of the MSA data structure.
* verify_parse will fill in errmsg if it sees a problem.
*/
//if (verify_parse(msa, afp->errmsg) != eslOK) { status = eslEFORMAT; goto ERROR; }
if (( status = esl_msa_SetDefaultWeights(msa)) != eslOK) goto ERROR;
if (ret_msa != NULL) *ret_msa = msa; else esl_msa_Destroy(msa);
return eslOK;
ERROR:
if (msa != NULL) esl_msa_Destroy(msa);
if (ret_msa != NULL) *ret_msa = NULL;
return status;
}
