/* Function: p7_ViterbiFilter()
* Synopsis: Calculates Viterbi score, vewy vewy fast, in limited precision.
* Incept: SRE, Tue Nov 27 09:15:24 2007 [Janelia]
*
* Purpose: Calculates an approximation of the Viterbi score for sequence
* <dsq> of length <L> residues, using optimized profile <om>,
* and a preallocated one-row DP matrix <ox>. Return the
* estimated Viterbi score (in nats) in <ret_sc>.
*
* Score may overflow (and will, on high-scoring
* sequences), but will not underflow.
*
* The model must be in a local alignment mode; other modes
* cannot provide the necessary guarantee of no underflow.
*
* This is a striped SIMD Viterbi implementation using Intel
* VMX integer intrinsics \citep{Farrar07}, in reduced
* precision (signed words, 16 bits).
*
* Args: dsq - digital target sequence, 1..L
* L - length of dsq in residues
* om - optimized profile
* ox - DP matrix
* ret_sc - RETURN: Viterbi score (in nats)
*
* Returns: <eslOK> on success;
* <eslERANGE> if the score overflows; in this case
* <*ret_sc> is <eslINFINITY>, and the sequence can
* be treated as a high-scoring hit.
*
* Throws: <eslEINVAL> if <ox> allocation is too small, or if
* profile isn't in a local alignment mode. (Must be in local
* alignment mode because that's what helps us guarantee
* limited dynamic range.)
*
* Xref: [Farrar07] for ideas behind striped SIMD DP.
* J2/46-47 for layout of HMMER's striped SIMD DP.
* J2/50 for single row DP.
* J2/60 for reduced precision (epu8)
* J2/65 for initial benchmarking
* J2/66 for precision maximization
* J4/138-140 for reimplementation in 16-bit precision
*/
int
p7_ViterbiFilter(const ESL_DSQ *dsq, int L, const P7_OPROFILE *om, P7_OMX *ox, float *ret_sc)
{
vector signed short mpv, dpv, ipv; /* previous row values */
vector signed short sv; /* temp storage of 1 curr row value in progress */
vector signed short dcv; /* delayed storage of D(i,q+1) */
vector signed short xEv; /* E state: keeps max for Mk->E as we go */
vector signed short xBv; /* B state: splatted vector of B[i-1] for B->Mk calculations */
vector signed short Dmaxv; /* keeps track of maximum D cell on row */
int16_t xE, xB, xC, xJ, xN; /* special states' scores */
int16_t Dmax; /* maximum D cell score on row */
int i; /* counter over sequence positions 1..L */
int q; /* counter over vectors 0..nq-1 */
int Q; /* segment length: # of vectors */
vector signed short *dp; /* using {MDI}MX(q) macro requires initialization of <dp> */
vector signed short *rsc; /* will point at om->ru[x] for residue x[i] */
vector signed short *tsc; /* will point into (and step thru) om->tu */
vector signed short negInfv;
Q = p7O_NQW(om->M);
dp = ox->dpw[0];
/* Check that the DP matrix is ok for us. */
if (Q > ox->allocQ8) ESL_EXCEPTION(eslEINVAL, "DP matrix allocated too small");
if (om->mode != p7_LOCAL && om->mode != p7_UNILOCAL) ESL_EXCEPTION(eslEINVAL, "Fast filter only works for local alignment");
ox->M = om->M;
negInfv = esl_vmx_set_s16((signed short)-32768);
/* Initialization. In unsigned arithmetic, -infinity is -32768
*/
for (q = 0; q < Q; q++)
MMXo(q) = IMXo(q) = DMXo(q) = negInfv;
xN = om->base_w;
xB = xN + om->xw[p7O_N][p7O_MOVE];
xJ = -32768;
xC = -32768;
xE = -32768;
#if p7_DEBUGGING
if (ox->debugging) p7_omx_DumpVFRow(ox, 0, xE, 0, xJ, xB, xC); /* first 0 is <rowi>: do header. second 0 is xN: always 0 here. */
#endif
for (i = 1; i <= L; i++)
{
rsc = om->rwv[dsq[i]];
tsc = om->twv;
dcv = negInfv; /* "-infinity" */
xEv = negInfv;
Dmaxv = negInfv;
xBv = esl_vmx_set_s16(xB);
/* Right shifts by 1 value (2 bytes). 4,8,12,x becomes x,4,8,12.
* Because ia32 is littlendian, this means a left bit shift.
* Zeros shift on automatically; replace it with -32768.
*/
mpv = MMXo(Q-1); mpv = vec_sld(negInfv, mpv, 14);
dpv = DMXo(Q-1); dpv = vec_sld(negInfv, dpv, 14);
ipv = IMXo(Q-1); ipv = vec_sld(negInfv, ipv, 14);
for (q = 0; q < Q; q++)
{
/* Calculate new MMXo(i,q); don't store it yet, hold it in sv. */
sv = vec_adds(xBv, *tsc); tsc++;
sv = vec_max (sv, vec_adds(mpv, *tsc)); tsc++;
sv = vec_max (sv, vec_adds(ipv, *tsc)); tsc++;
sv = vec_max (sv, vec_adds(dpv, *tsc)); tsc++;
sv = vec_adds(sv, *rsc); rsc++;
xEv = vec_max(xEv, sv);
/* Load {MDI}(i-1,q) into mpv, dpv, ipv;
* {MDI}MX(q) is then the current, not the prev row
*/
mpv = MMXo(q);
dpv = DMXo(q);
ipv = IMXo(q);
/* Do the delayed stores of {MD}(i,q) now that memory is usable */
MMXo(q) = sv;
DMXo(q) = dcv;
/* Calculate the next D(i,q+1) partially: M->D only;
* delay storage, holding it in dcv
*/
dcv = vec_adds(sv, *tsc); tsc++;
Dmaxv = vec_max(dcv, Dmaxv);
/* Calculate and store I(i,q) */
sv = vec_adds(mpv, *tsc); tsc++;
IMXo(q)= vec_max(sv, vec_adds(ipv, *tsc)); tsc++;
}
/* Now the "special" states, which start from Mk->E (->C, ->J->B) */
xE = esl_vmx_hmax_s16(xEv);
if (xE >= 32767) { *ret_sc = eslINFINITY; return eslERANGE; } /* immediately detect overflow */
xN = xN + om->xw[p7O_N][p7O_LOOP];
xC = ESL_MAX(xC + om->xw[p7O_C][p7O_LOOP], xE + om->xw[p7O_E][p7O_MOVE]);
xJ = ESL_MAX(xJ + om->xw[p7O_J][p7O_LOOP], xE + om->xw[p7O_E][p7O_LOOP]);
xB = ESL_MAX(xJ + om->xw[p7O_J][p7O_MOVE], xN + om->xw[p7O_N][p7O_MOVE]);
/* and now xB will carry over into next i, and xC carries over after i=L */
/* Finally the "lazy F" loop (sensu [Farrar07]). We can often
* prove that we don't need to evaluate any D->D paths at all.
*
* The observation is that if we can show that on the next row,
* B->M(i+1,k) paths always dominate M->D->...->D->M(i+1,k) paths
* for all k, then we don't need any D->D calculations.
*
* The test condition is:
* max_k D(i,k) + max_k ( TDD(k-2) + TDM(k-1) - TBM(k) ) < xB(i)
* So:
* max_k (TDD(k-2) + TDM(k-1) - TBM(k)) is precalc'ed in om->dd_bound;
* max_k D(i,k) is why we tracked Dmaxv;
* xB(i) was just calculated above.
*/
Dmax = esl_vmx_hmax_s16(Dmaxv);
if (Dmax + om->ddbound_w > xB)
{
/* Now we're obligated to do at least one complete DD path to be sure. */
/* dcv has carried through from end of q loop above */
dcv = vec_sld(negInfv, dcv, 14);
tsc = om->twv + 7*Q; /* set tsc to start of the DD's */
for (q = 0; q < Q; q++)
{
DMXo(q) = vec_max(dcv, DMXo(q));
dcv = vec_adds(DMXo(q), *tsc); tsc++;
}
/* We may have to do up to three more passes; the check
* is for whether crossing a segment boundary can improve
* our score.
*/
do {
dcv = vec_sld(negInfv, dcv, 14);
tsc = om->twv + 7*Q; /* set tsc to start of the DD's */
for (q = 0; q < Q; q++)
{
if (! vec_any_gt(dcv, DMXo(q))) break;
DMXo(q) = vec_max(dcv, DMXo(q));
dcv = vec_adds(DMXo(q), *tsc); tsc++;
}
} while (q == Q);
}
else /* not calculating DD? then just store the last M->D vector calc'ed.*/
DMXo(0) = vec_sld(negInfv, dcv, 14);
#if p7_DEBUGGING
if (ox->debugging) p7_omx_DumpVFRow(ox, i, xE, 0, xJ, xB, xC);
#endif
} /* end loop over sequence residues 1..L */
/* finally C->T */
if (xC > -32768)
{
*ret_sc = (float) xC + (float) om->xw[p7O_C][p7O_MOVE] - (float) om->base_w;
/* *ret_sc += L * om->ncj_roundoff; see J4/150 for rationale: superceded by -3.0nat approximation*/
*ret_sc /= om->scale_w;
*ret_sc -= 3.0; /* the NN/CC/JJ=0,-3nat approximation: see J5/36. That's ~ L \log \frac{L}{L+3}, for our NN,CC,JJ contrib */
}
else *ret_sc = -eslINFINITY;
return eslOK;
}
/* Function: p7_ViterbiFilter_longtarget()
* Synopsis: Finds windows within potentially long sequence blocks with Viterbi
* scores above threshold (vewy vewy fast, in limited precision)
*
* Purpose: Calculates an approximation of the Viterbi score for regions
* of sequence <dsq>, using optimized profile <om>, and a pre-
* allocated one-row DP matrix <ox>, and captures the positions
* at which such regions exceed the score required to be
* significant in the eyes of the calling function (usually
* p=0.001).
*
* The resulting landmarks are converted to subsequence
* windows by the calling function
*
* The model must be in a local alignment mode; other modes
* cannot provide the necessary guarantee of no underflow.
*
* This is a striped SIMD Viterbi implementation using Intel
* VMX integer intrinsics \citep{Farrar07}, in reduced
* precision (signed words, 16 bits).
*
* Args: dsq - digital target sequence, 1..L
* L - length of dsq in residues
* om - optimized profile
* ox - DP matrix
* filtersc - null or bias correction, required for translating a P-value threshold into a score threshold
* P - p-value below which a region is captured as being above threshold
* windowlist - RETURN: array of hit windows (start and end of diagonal) for the above-threshold areas
*
* Returns: <eslOK> on success;
*
* Throws: <eslEINVAL> if <ox> allocation is too small, or if
* profile isn't in a local alignment mode. (Must be in local
* alignment mode because that's what helps us guarantee
* limited dynamic range.)
*
* Xref: See p7_ViterbiFilter()
*/
int
p7_ViterbiFilter_longtarget(const ESL_DSQ *dsq, int L, const P7_OPROFILE *om, P7_OMX *ox,
float filtersc, double P, P7_HMM_WINDOWLIST *windowlist)
{
vector signed short mpv, dpv, ipv; /* previous row values */
vector signed short sv; /* temp storage of 1 curr row value in progress */
vector signed short dcv; /* delayed storage of D(i,q+1) */
vector signed short xEv; /* E state: keeps max for Mk->E as we go */
vector signed short xBv; /* B state: splatted vector of B[i-1] for B->Mk calculations */
vector signed short Dmaxv; /* keeps track of maximum D cell on row */
int16_t xE, xB, xC, xJ, xN; /* special states' scores */
int16_t Dmax; /* maximum D cell score on row */
int i; /* counter over sequence positions 1..L */
int q; /* counter over vectors 0..nq-1 */
int Q = p7O_NQW(om->M); /* segment length: # of vectors */
vector signed short *dp = ox->dpw[0]; /* using {MDI}MX(q) macro requires initialization of <dp> */
vector signed short *rsc; /* will point at om->ru[x] for residue x[i] */
vector signed short *tsc; /* will point into (and step thru) om->tu */
vector signed short negInfv;
int16_t sc_thresh;
float invP;
int z;
union { vector signed short v; int16_t i[8]; } tmp;
windowlist->count = 0;
/*
* In p7_ViterbiFilter, converting from a scaled int Viterbi score
* S (aka xE the score getting to state E) to a probability
* goes like this:
* vsc = S + om->xw[p7O_E][p7O_MOVE] + om->xw[p7O_C][p7O_MOVE] - om->base_w
* ret_sc /= om->scale_w;
* vsc -= 3.0;
* P = esl_gumbel_surv((vfsc - filtersc) / eslCONST_LOG2 , om->evparam[p7_VMU], om->evparam[p7_VLAMBDA]);
* and we're computing the threshold vsc, so invert it:
* (vsc - filtersc) / eslCONST_LOG2 = esl_gumbel_invsurv( P, om->evparam[p7_VMU], om->evparam[p7_VLAMBDA])
* vsc = filtersc + eslCONST_LOG2 * esl_gumbel_invsurv( P, om->evparam[p7_VMU], om->evparam[p7_VLAMBDA])
* vsc += 3.0
* vsc *= om->scale_w
* S = vsc - (float)om->xw[p7O_E][p7O_MOVE] - (float)om->xw[p7O_C][p7O_MOVE] + (float)om->base_w
*/
invP = esl_gumbel_invsurv(P, om->evparam[p7_VMU], om->evparam[p7_VLAMBDA]);
sc_thresh = (int) ceil ( ( (filtersc + (eslCONST_LOG2 * invP) + 3.0) * om->scale_w )
- (float)om->xw[p7O_E][p7O_MOVE] - (float)om->xw[p7O_C][p7O_MOVE] + (float)om->base_w );
/* Check that the DP matrix is ok for us. */
if (Q > ox->allocQ8) ESL_EXCEPTION(eslEINVAL, "DP matrix allocated too small");
if (om->mode != p7_LOCAL && om->mode != p7_UNILOCAL) ESL_EXCEPTION(eslEINVAL, "Fast filter only works for local alignment");
ox->M = om->M;
negInfv = esl_vmx_set_s16((signed short)-32768);
/* Initialization. In unsigned arithmetic, -infinity is -32768
*/
for (q = 0; q < Q; q++)
MMXo(q) = IMXo(q) = DMXo(q) = negInfv;
xN = om->base_w;
xB = xN + om->xw[p7O_N][p7O_MOVE];
xJ = -32768;
xC = -32768;
xE = -32768;
#if p7_DEBUGGING
if (ox->debugging) p7_omx_DumpVFRow(ox, 0, xE, 0, xJ, xB, xC); /* first 0 is <rowi>: do header. second 0 is xN: always 0 here. */
#endif
for (i = 1; i <= L; i++)
{
rsc = om->rwv[dsq[i]];
tsc = om->twv;
dcv = negInfv; /* "-infinity" */
xEv = negInfv;
Dmaxv = negInfv;
xBv = esl_vmx_set_s16(xB);
/* Right shifts by 1 value (2 bytes). 4,8,12,x becomes x,4,8,12.
* Because ia32 is littlendian, this means a left bit shift.
* Zeros shift on automatically; replace it with -32768.
*/
mpv = MMXo(Q-1); mpv = vec_sld(negInfv, mpv, 14);
dpv = DMXo(Q-1); dpv = vec_sld(negInfv, dpv, 14);
ipv = IMXo(Q-1); ipv = vec_sld(negInfv, ipv, 14);
for (q = 0; q < Q; q++)
{
/* Calculate new MMXo(i,q); don't store it yet, hold it in sv. */
sv = vec_adds(xBv, *tsc); tsc++;
sv = vec_max (sv, vec_adds(mpv, *tsc)); tsc++;
sv = vec_max (sv, vec_adds(ipv, *tsc)); tsc++;
sv = vec_max (sv, vec_adds(dpv, *tsc)); tsc++;
sv = vec_adds(sv, *rsc); rsc++;
xEv = vec_max(xEv, sv);
/* Load {MDI}(i-1,q) into mpv, dpv, ipv;
* {MDI}MX(q) is then the current, not the prev row
*/
mpv = MMXo(q);
dpv = DMXo(q);
ipv = IMXo(q);
/* Do the delayed stores of {MD}(i,q) now that memory is usable */
MMXo(q) = sv;
DMXo(q) = dcv;
/* Calculate the next D(i,q+1) partially: M->D only;
* delay storage, holding it in dcv
*/
dcv = vec_adds(sv, *tsc); tsc++;
Dmaxv = vec_max(dcv, Dmaxv);
/* Calculate and store I(i,q) */
sv = vec_adds(mpv, *tsc); tsc++;
IMXo(q)= vec_max(sv, vec_adds(ipv, *tsc)); tsc++;
}
/* Now the "special" states, which start from Mk->E (->C, ->J->B) */
xE = esl_vmx_hmax_s16(xEv);
if (xE >= sc_thresh) {
//hit score threshold. Add a window to the list, then reset scores.
/* Unpack and unstripe, then find the position responsible for the hit */
for (q = 0; q < Q; q++) {
tmp.v = MMXo(q);
for (z = 0; z < 8; z++) { // unstripe
if ( tmp.i[z] == xE && (q+Q*z+1) <= om->M) {
// (q+Q*z+1) is the model position k at which the xE score is found
p7_hmmwindow_new(windowlist, 0, i, 0, (q+Q*z+1), 1, 0.0, p7_NOCOMPLEMENT );
}
}
MMXo(q) = IMXo(q) = DMXo(q) = negInfv; //reset score to start search for next vit window.
}
} else {
xN = xN + om->xw[p7O_N][p7O_LOOP];
xC = ESL_MAX(xC + om->xw[p7O_C][p7O_LOOP], xE + om->xw[p7O_E][p7O_MOVE]);
xJ = ESL_MAX(xJ + om->xw[p7O_J][p7O_LOOP], xE + om->xw[p7O_E][p7O_LOOP]);
xB = ESL_MAX(xJ + om->xw[p7O_J][p7O_MOVE], xN + om->xw[p7O_N][p7O_MOVE]);
/* and now xB will carry over into next i, and xC carries over after i=L */
/* Finally the "lazy F" loop (sensu [Farrar07]). We can often
* prove that we don't need to evaluate any D->D paths at all.
*
* The observation is that if we can show that on the next row,
* B->M(i+1,k) paths always dominate M->D->...->D->M(i+1,k) paths
* for all k, then we don't need any D->D calculations.
*
* The test condition is:
* max_k D(i,k) + max_k ( TDD(k-2) + TDM(k-1) - TBM(k) ) < xB(i)
* So:
* max_k (TDD(k-2) + TDM(k-1) - TBM(k)) is precalc'ed in om->dd_bound;
* max_k D(i,k) is why we tracked Dmaxv;
* xB(i) was just calculated above.
*/
Dmax = esl_vmx_hmax_s16(Dmaxv);
if (Dmax + om->ddbound_w > xB)
{
/* Now we're obligated to do at least one complete DD path to be sure. */
/* dcv has carried through from end of q loop above */
dcv = vec_sld(negInfv, dcv, 14);
tsc = om->twv + 7*Q; /* set tsc to start of the DD's */
for (q = 0; q < Q; q++)
{
DMXo(q) = vec_max(dcv, DMXo(q));
dcv = vec_adds(DMXo(q), *tsc); tsc++;
}
/* We may have to do up to three more passes; the check
* is for whether crossing a segment boundary can improve
* our score.
*/
do {
dcv = vec_sld(negInfv, dcv, 14);
tsc = om->twv + 7*Q; /* set tsc to start of the DD's */
for (q = 0; q < Q; q++)
{
if (! vec_any_gt(dcv, DMXo(q))) break;
DMXo(q) = vec_max(dcv, DMXo(q));
dcv = vec_adds(DMXo(q), *tsc); tsc++;
}
} while (q == Q);
}
else /* not calculating DD? then just store the last M->D vector calc'ed.*/
DMXo(0) = vec_sld(negInfv, dcv, 14);
#if p7_DEBUGGING
if (ox->debugging) p7_omx_DumpVFRow(ox, i, xE, 0, xJ, xB, xC);
#endif
}
} /* end loop over sequence residues 1..L */
return eslOK;
}
