/* Function: p7_Null2_ByTrace()
* Synopsis: Assign null2 scores to an envelope by the sampling method.
* Incept: SRE, Mon Aug 18 10:22:49 2008 [Janelia]
*
* Purpose: Identical to <p7_GNull2_ByTrace()> except that
* <om>, <wrk> are SSE optimized versions of the profile
* and the residue posterior probability matrix. See
* <p7_GNull2_ByTrace()> documentation.
*/
int
p7_Null2_ByTrace(const P7_OPROFILE *om, const P7_TRACE *tr, int zstart, int zend, P7_OMX *wrk, float *null2)
{
union { __m128 v; float p[4]; } u;
int Q = p7O_NQF(om->M);
int Ld = 0;
float *xmx = wrk->xmx; /* enables use of XMXo macro */
float norm;
float xfactor;
__m128 sv;
__m128 *rp;
int q, r, s;
int x;
int z;
/* We'll use the i=0 row in wrk for working space: dp[0][] and xmx[][0]. */
for (q = 0; q < Q; q++)
{
wrk->dpf[0][q*3 + p7X_M] = _mm_setzero_ps();
wrk->dpf[0][q*3 + p7X_I] = _mm_setzero_ps();
}
XMXo(0,p7X_N) = 0.0;
XMXo(0,p7X_C) = 0.0;
XMXo(0,p7X_J) = 0.0;
/* Calculate emitting state usage in this particular trace segment */
for (z = zstart; z <= zend; z++)
{
if (tr->i[z] == 0) continue; /* quick test for whether this trace elem emitted or not */
Ld++;
if (tr->k[z] > 0) /* must be an M or I */
{ /* surely there's an easier way? but our workspace is striped, interleaved quads... */
s = ( (tr->st[z] == p7T_M) ? p7X_M : p7X_I);
q = p7X_NSCELLS * ( (tr->k[z] - 1) % Q) + p7X_M;
r = (tr->k[z] - 1) / Q;
u.v = wrk->dpf[0][q];
u.p[r] += 1.0; /* all this to increment a count by one! */
wrk->dpf[0][q] = u.v;
}
else /* emitted an x_i with no k; must be an N,C,J */
{
switch (tr->st[z]) {
case p7T_N: XMXo(0,p7X_N) += 1.0; break;
case p7T_C: XMXo(0,p7X_C) += 1.0; break;
case p7T_J: XMXo(0,p7X_J) += 1.0; break;
}
}
}
norm = 1.0 / (float) Ld;
sv = _mm_set1_ps(norm);
for (q = 0; q < Q; q++)
{
wrk->dpf[0][q*3 + p7X_M] = _mm_mul_ps(wrk->dpf[0][q*3 + p7X_M], sv);
wrk->dpf[0][q*3 + p7X_I] = _mm_mul_ps(wrk->dpf[0][q*3 + p7X_I], sv);
}
XMXo(0,p7X_N) *= norm;
XMXo(0,p7X_C) *= norm;
XMXo(0,p7X_J) *= norm;
/* Calculate null2's emission odds, by taking posterior weighted sum
* over all emission vectors used in paths explaining the domain.
*/
xfactor = XMXo(0,p7X_N) + XMXo(0,p7X_C) + XMXo(0,p7X_J);
for (x = 0; x < om->abc->K; x++)
{
sv = _mm_setzero_ps();
rp = om->rfv[x];
for (q = 0; q < Q; q++)
{
sv = _mm_add_ps(sv, _mm_mul_ps(wrk->dpf[0][q*3 + p7X_M], *rp)); rp++;
sv = _mm_add_ps(sv, wrk->dpf[0][q*3 + p7X_I]); /* insert emission odds implicitly 1.0 */
// sv = _mm_add_ps(sv, _mm_mul_ps(wrk->dpf[0][q*3 + p7X_I], *rp)); rp++;
}
esl_sse_hsum_ps(sv, &(null2[x]));
null2[x] += xfactor;
}
/* now null2[x] = \frac{f_d(x)}{f_0(x)} for all x in alphabet,
* 0..K-1, where f_d(x) are the ad hoc "null2" residue frequencies
* for this envelope.
*/
/* make valid scores for all degeneracies, by averaging the odds ratios. */
esl_abc_FAvgScVec(om->abc, null2);
null2[om->abc->K] = 1.0; /* gap character */
null2[om->abc->Kp-2] = 1.0; /* nonresidue "*" */
null2[om->abc->Kp-1] = 1.0; /* missing data "~" */
return eslOK;
}
/* Function: p7_Null2_ByExpectation()
* Synopsis: Calculate null2 model from posterior probabilities.
* Incept: SRE, Mon Aug 18 08:32:55 2008 [Janelia]
*
* Purpose: Identical to <p7_GNull2_ByExpectation()> except that
* <om>, <pp> are SSE optimized versions of the profile
* and the residue posterior probability matrix. See
* <p7_GNull2_ByExpectation()> documentation.
*
* Args: om - profile, in any mode, target length model set to <L>
* pp - posterior prob matrix, for <om> against domain envelope <dsq+i-1> (offset)
* null2 - RETURN: null2 log odds scores per residue; <0..Kp-1>; caller allocated space
*/
int
p7_Null2_ByExpectation(const P7_OPROFILE *om, const P7_OMX *pp, float *null2)
{
int M = om->M;
int Ld = pp->L;
int Q = p7O_NQF(M);
float *xmx = pp->xmx; /* enables use of XMXo(i,s) macro */
float norm;
__m128 *rp;
__m128 sv;
float xfactor;
int i,q,x;
/* Calculate expected # of times that each emitting state was used
* in generating the Ld residues in this domain.
* The 0 row in <wrk> is used to hold these numbers.
*/
memcpy(pp->dpf[0], pp->dpf[1], sizeof(__m128) * 3 * Q);
XMXo(0,p7X_N) = XMXo(1,p7X_N);
XMXo(0,p7X_C) = XMXo(1,p7X_C); /* 0.0 */
XMXo(0,p7X_J) = XMXo(1,p7X_J); /* 0.0 */
for (i = 2; i <= Ld; i++)
{
for (q = 0; q < Q; q++)
{
pp->dpf[0][q*3 + p7X_M] = _mm_add_ps(pp->dpf[i][q*3 + p7X_M], pp->dpf[0][q*3 + p7X_M]);
pp->dpf[0][q*3 + p7X_I] = _mm_add_ps(pp->dpf[i][q*3 + p7X_I], pp->dpf[0][q*3 + p7X_I]);
}
XMXo(0,p7X_N) += XMXo(i,p7X_N);
XMXo(0,p7X_C) += XMXo(i,p7X_C);
XMXo(0,p7X_J) += XMXo(i,p7X_J);
}
/* Convert those expected #'s to frequencies, to use as posterior weights. */
norm = 1.0 / (float) Ld;
sv = _mm_set1_ps(norm);
for (q = 0; q < Q; q++)
{
pp->dpf[0][q*3 + p7X_M] = _mm_mul_ps(pp->dpf[0][q*3 + p7X_M], sv);
pp->dpf[0][q*3 + p7X_I] = _mm_mul_ps(pp->dpf[0][q*3 + p7X_I], sv);
}
XMXo(0,p7X_N) *= norm;
XMXo(0,p7X_C) *= norm;
XMXo(0,p7X_J) *= norm;
/* Calculate null2's emission odds, by taking posterior weighted sum
* over all emission vectors used in paths explaining the domain.
*/
xfactor = XMXo(0, p7X_N) + XMXo(0, p7X_C) + XMXo(0, p7X_J);
for (x = 0; x < om->abc->K; x++)
{
sv = _mm_setzero_ps();
rp = om->rfv[x];
for (q = 0; q < Q; q++)
{
sv = _mm_add_ps(sv, _mm_mul_ps(pp->dpf[0][q*3 + p7X_M], *rp)); rp++;
sv = _mm_add_ps(sv, pp->dpf[0][q*3 + p7X_I]); /* insert odds implicitly 1.0 */
// sv = _mm_add_ps(sv, _mm_mul_ps(pp->dpf[0][q*3 + p7X_I], *rp)); rp++;
}
esl_sse_hsum_ps(sv, &(null2[x]));
null2[x] += xfactor;
}
/* now null2[x] = \frac{f_d(x)}{f_0(x)} for all x in alphabet,
* 0..K-1, where f_d(x) are the ad hoc "null2" residue frequencies
* for this envelope.
*/
/* make valid scores for all degeneracies, by averaging the odds ratios. */
esl_abc_FAvgScVec(om->abc, null2);
null2[om->abc->K] = 1.0; /* gap character */
null2[om->abc->Kp-2] = 1.0; /* nonresidue "*" */
null2[om->abc->Kp-1] = 1.0; /* missing data "~" */
return eslOK;
}
/* Function: p7_OptimalAccuracy()
* Synopsis: DP fill of an optimal accuracy alignment calculation.
* Incept: SRE, Mon Aug 18 11:04:48 2008 [Janelia]
*
* Purpose: Calculates the fill step of the optimal accuracy decoding
* algorithm \citep{Kall05}.
*
* Caller provides the posterior decoding matrix <pp>,
* which was calculated by Forward/Backward on a target sequence
* of length <pp->L> using the query model <om>.
*
* Caller also provides a DP matrix <ox>, allocated for a full
* <om->M> by <L> comparison. The routine fills this in
* with OA scores.
*
* Args: gm - query profile
* pp - posterior decoding matrix created by <p7_GPosteriorDecoding()>
* gx - RESULT: caller provided DP matrix for <gm->M> by <L>
* ret_e - RETURN: expected number of correctly decoded positions
*
* Returns: <eslOK> on success, and <*ret_e> contains the final OA
* score, which is the expected number of correctly decoded
* positions in the target sequence (up to <L>).
*
* Throws: (no abnormal error conditions)
*/
int
p7_OptimalAccuracy(const P7_OPROFILE *om, const P7_OMX *pp, P7_OMX *ox, float *ret_e)
{
vector float mpv, dpv, ipv; /* previous row values */
vector float sv; /* temp storage of 1 curr row value in progress */
vector float xEv; /* E state: keeps max for Mk->E as we go */
vector float xBv; /* B state: splatted vector of B[i-1] for B->Mk calculations */
vector float dcv;
float *xmx = ox->xmx;
vector float *dpc = ox->dpf[0]; /* current row, for use in {MDI}MO(dpp,q) access macro */
vector float *dpp; /* previous row, for use in {MDI}MO(dpp,q) access macro */
vector float *ppp; /* quads in the <pp> posterior probability matrix */
vector float *tp; /* quads in the <om->tfv> transition scores */
vector float zerov;
vector float infv;
int M = om->M;
int Q = p7O_NQF(M);
int q;
int j;
int i;
float t1, t2;
zerov = (vector float) vec_splat_u32(0);
infv = esl_vmx_set_float(-eslINFINITY);
ox->M = om->M;
ox->L = pp->L;
for (q = 0; q < Q; q++) MMO(dpc, q) = IMO(dpc,q) = DMO(dpc,q) = infv;
XMXo(0, p7X_E) = -eslINFINITY;
XMXo(0, p7X_N) = 0.;
XMXo(0, p7X_J) = -eslINFINITY;
XMXo(0, p7X_B) = 0.;
XMXo(0, p7X_C) = -eslINFINITY;
for (i = 1; i <= pp->L; i++)
{
dpp = dpc; /* previous DP row in OA matrix */
dpc = ox->dpf[i]; /* current DP row in OA matrix */
ppp = pp->dpf[i]; /* current row in the posterior probabilities per position */
tp = om->tfv; /* transition probabilities */
dcv = infv;
xEv = infv;
xBv = esl_vmx_set_float(XMXo(i-1, p7X_B));
mpv = vec_sld(infv, MMO(dpp,Q-1), 12); /* Right shifts by 4 bytes. 4,8,12,x becomes x,4,8,12. */
dpv = vec_sld(infv, DMO(dpp,Q-1), 12);
ipv = vec_sld(infv, IMO(dpp,Q-1), 12);
for (q = 0; q < Q; q++)
{
sv = vec_and(vec_cmpgt(*tp, zerov), xBv); tp++;
sv = vec_max(sv, vec_and(vec_cmpgt(*tp, zerov), mpv)); tp++;
sv = vec_max(sv, vec_and(vec_cmpgt(*tp, zerov), ipv)); tp++;
sv = vec_max(sv, vec_and(vec_cmpgt(*tp, zerov), dpv)); tp++;
sv = vec_add(sv, *ppp); ppp += 2;
xEv = vec_max(xEv, sv);
mpv = MMO(dpp,q);
dpv = DMO(dpp,q);
ipv = IMO(dpp,q);
MMO(dpc,q) = sv;
DMO(dpc,q) = dcv;
dcv = vec_and(vec_cmpgt(*tp, zerov), sv); tp++;
sv = vec_and(vec_cmpgt(*tp, zerov), mpv); tp++;
sv = vec_max(sv, vec_and(vec_cmpgt(*tp, zerov), ipv)); tp++;
IMO(dpc,q) = vec_add(sv, *ppp); ppp++;
}
/* dcv has carried through from end of q loop above; store it
* in first pass, we add M->D and D->D path into DMX
*/
dcv = vec_sld(infv, dcv, 12);
tp = om->tfv + 7*Q; /* set tp to start of the DD's */
for (q = 0; q < Q; q++)
{
DMO(dpc, q) = vec_max(dcv, DMO(dpc, q));
dcv = vec_and(vec_cmpgt(*tp, zerov), DMO(dpc,q)); tp++;
}
/* fully serialized D->D; can optimize later */
for (j = 1; j < 4; j++)
{
dcv = vec_sld(infv, dcv, 12);
tp = om->tfv + 7*Q;
for (q = 0; q < Q; q++)
{
DMO(dpc, q) = vec_max(dcv, DMO(dpc, q));
dcv = vec_and(vec_cmpgt(*tp, zerov), dcv); tp++;
}
}
/* D->E paths */
for (q = 0; q < Q; q++) xEv = vec_max(xEv, DMO(dpc,q));
/* Specials */
XMXo(i,p7X_E) = esl_vmx_hmax_float(xEv);
t1 = ( (om->xf[p7O_J][p7O_LOOP] == 0.0) ? 0.0 : ox->xmx[(i-1)*p7X_NXCELLS+p7X_J] + pp->xmx[i*p7X_NXCELLS+p7X_J]);
t2 = ( (om->xf[p7O_E][p7O_LOOP] == 0.0) ? 0.0 : ox->xmx[ i *p7X_NXCELLS+p7X_E]);
ox->xmx[i*p7X_NXCELLS+p7X_J] = ESL_MAX(t1, t2);
t1 = ( (om->xf[p7O_C][p7O_LOOP] == 0.0) ? 0.0 : ox->xmx[(i-1)*p7X_NXCELLS+p7X_C] + pp->xmx[i*p7X_NXCELLS+p7X_C]);
t2 = ( (om->xf[p7O_E][p7O_MOVE] == 0.0) ? 0.0 : ox->xmx[ i *p7X_NXCELLS+p7X_E]);
ox->xmx[i*p7X_NXCELLS+p7X_C] = ESL_MAX(t1, t2);
ox->xmx[i*p7X_NXCELLS+p7X_N] = ((om->xf[p7O_N][p7O_LOOP] == 0.0) ? 0.0 : ox->xmx[(i-1)*p7X_NXCELLS+p7X_N] + pp->xmx[i*p7X_NXCELLS+p7X_N]);
t1 = ( (om->xf[p7O_N][p7O_MOVE] == 0.0) ? 0.0 : ox->xmx[i*p7X_NXCELLS+p7X_N]);
t2 = ( (om->xf[p7O_J][p7O_MOVE] == 0.0) ? 0.0 : ox->xmx[i*p7X_NXCELLS+p7X_J]);
ox->xmx[i*p7X_NXCELLS+p7X_B] = ESL_MAX(t1, t2);
}
*ret_e = ox->xmx[pp->L*p7X_NXCELLS+p7X_C];
return eslOK;
}
