/* Function: p7_oprofile_MPIPackSize()
* Synopsis: Calculates size needed to pack an OPROFILE.
* Incept: MSF, Wed Oct 21, 2009 [Janelia]
*
* Purpose: Calculate an upper bound on the number of bytes
* that <p7_oprofile_MPIPack()> will need to pack an
* OPROFILE <om> in a packed MPI message for MPI
* communicator <comm>; return that number of bytes
* in <*ret_n>.
*
* Returns: <eslOK> on success, and <*ret_n> contains the answer.
*
* Throws: <eslESYS> if an MPI call fails, and <*ret_n> is 0.
*/
int
p7_oprofile_MPIPackSize(P7_OPROFILE *om, MPI_Comm comm, int *ret_n)
{
int status;
int n = 0;
int K = om->abc->Kp;
int len = 0;
int cnt;
int sz;
int Q4 = p7O_NQF(om->M);
int Q8 = p7O_NQW(om->M);
int Q16 = p7O_NQB(om->M);
int vsz = sizeof(__m128i);
/* MSV Filter information */
if (MPI_Pack_size(5, MPI_CHAR, comm, &sz) != 0) ESL_XEXCEPTION(eslESYS, "pack size failed"); n += sz;
if (MPI_Pack_size(1, MPI_FLOAT, comm, &sz) != 0) ESL_XEXCEPTION(eslESYS, "pack size failed"); n += sz;
if (MPI_Pack_size(vsz*Q16, MPI_CHAR, comm, &sz) != 0) ESL_XEXCEPTION(eslESYS, "pack size failed"); n += (K*sz);
/* Viterbi Filter information */
if (MPI_Pack_size(1, MPI_SHORT, comm, &sz) != 0) ESL_XEXCEPTION(eslESYS, "pack size failed"); n += ((p7O_NXSTATES*p7O_NXTRANS+2)*sz);
if (MPI_Pack_size(2, MPI_FLOAT, comm, &sz) != 0) ESL_XEXCEPTION(eslESYS, "pack size failed"); n += sz;
if (MPI_Pack_size(K*vsz*Q8, MPI_CHAR, comm, &sz) != 0) ESL_XEXCEPTION(eslESYS, "pack size failed"); n += sz;
if (MPI_Pack_size(8*vsz*Q8, MPI_CHAR, comm, &sz) != 0) ESL_XEXCEPTION(eslESYS, "pack size failed"); n += sz;
/* Forward/Backward information */
if (MPI_Pack_size(1, MPI_FLOAT, comm, &sz) != 0) ESL_XEXCEPTION(eslESYS, "pack size failed"); n += (p7O_NXSTATES*p7O_NXTRANS*sz);
if (MPI_Pack_size(K*vsz*Q4, MPI_CHAR, comm, &sz) != 0) ESL_XEXCEPTION(eslESYS, "pack size failed"); n += sz;
if (MPI_Pack_size(8*vsz*Q4, MPI_CHAR, comm, &sz) != 0) ESL_XEXCEPTION(eslESYS, "pack size failed"); n += sz;
/* disk offsets */
if (MPI_Pack_size(1, MPI_LONG_LONG_INT, comm, &sz) != 0) ESL_XEXCEPTION(eslESYS, "pack size failed"); n += ((p7_NOFFSETS+2)*sz);
/* annotation info */
if (om->name != NULL) len += strlen(om->name) + 1;
if (om->acc != NULL) len += strlen(om->acc) + 1;
if (om->desc != NULL) len += strlen(om->desc) + 1;
if (om->rf != NULL) len += strlen(om->rf) + 1;
if (om->mm != NULL) len += strlen(om->mm) + 1;
if (om->cs != NULL) len += strlen(om->cs) + 1;
if (om->consensus != NULL) len += strlen(om->consensus) + 1;
if (MPI_Pack_size(7, MPI_INT, comm, &sz) != 0) ESL_XEXCEPTION(eslESYS, "pack size failed"); n += sz;
if (MPI_Pack_size(len, MPI_CHAR, comm, &sz) != 0) ESL_XEXCEPTION(eslESYS, "pack size failed"); n += sz;
cnt = p7_NEVPARAM + p7_NCUTOFFS + p7_MAXABET;
if (MPI_Pack_size(cnt, MPI_FLOAT, comm, &sz) != 0) ESL_XEXCEPTION(eslESYS, "pack size failed"); n += sz;
/* current model size */
if (MPI_Pack_size(4, MPI_INT, comm, &sz) != 0) ESL_XEXCEPTION(eslESYS, "pack size failed"); n += sz;
if (MPI_Pack_size(1, MPI_FLOAT, comm, &sz) != 0) ESL_XEXCEPTION(eslESYS, "pack size failed"); n += sz;
*ret_n = n;
return eslOK;
ERROR:
*ret_n = 0;
return status;
}
/* Function: p7_omx_Create()
* Synopsis: Create an optimized dynamic programming matrix.
* Incept: SRE, Tue Nov 27 08:48:20 2007 [Janelia]
*
* Purpose: Allocates a reusable, resizeable <P7_OMX> for models up to
* size <allocM> and target sequences up to length
* <allocL/allocXL>, for use by any of the various optimized
* DP routines.
*
* To allocate the very memory-efficient one-row matrix
* used by *Filter() and *Score() functions that only
* calculate scores, <allocM=M>, <allocL=0>, and
* <allocXL=0>.
*
* To allocate the reasonably memory-efficient linear
* arrays used by *Parser() functions that only keep
* special (X) state scores, <allocM=M>, <allocL=0>,
* and <allocXL=L>.
*
* To allocate a complete matrix suitable for functions
* that need the whole DP matrix for traceback, sampling,
* posterior decoding, or reestimation, <allocM=M> and
* <allocL=allocXL=L>.
*
* Returns: a pointer to the new <P7_OMX>.
*
* Throws: <NULL> on allocation failure.
*/
P7_OMX *
p7_omx_Create(int allocM, int allocL, int allocXL)
{
P7_OMX *ox = NULL;
int i;
int status;
ESL_ALLOC(ox, sizeof(P7_OMX));
ox->dp_mem = NULL;
ox->dpb = NULL;
ox->dpw = NULL;
ox->dpf = NULL;
ox->xmx = NULL;
ox->x_mem = NULL;
/* DP matrix will be allocated for allocL+1 rows 0,1..L; allocQ4*p7X_NSCELLS columns */
ox->allocR = allocL+1;
ox->validR = ox->allocR;
ox->allocQ4 = p7O_NQF(allocM);
ox->allocQ8 = p7O_NQW(allocM);
ox->allocQ16 = p7O_NQB(allocM);
ox->ncells = ox->allocR * ox->allocQ4 * 4; /* # of DP cells allocated, where 1 cell contains MDI */
/* floats always dominate; +15 for alignment */
ESL_ALLOC(ox->dp_mem, sizeof(vector float) * ox->allocR * ox->allocQ4 * p7X_NSCELLS + 15);
ESL_ALLOC(ox->dpb, sizeof(vector unsigned char *) * ox->allocR);
ESL_ALLOC(ox->dpw, sizeof(vector signed short *) * ox->allocR);
ESL_ALLOC(ox->dpf, sizeof(vector float *) * ox->allocR);
/* DP memory shared by <dpb>, <dpw>, <dpf> */
ox->dpb[0] = (vector unsigned char *) ((unsigned long int) ((char *) ox->dp_mem + 15) & (~0xf));
ox->dpw[0] = (vector signed short *) ox->dpb[0];
ox->dpf[0] = (vector float *) ox->dpb[0];
for (i = 1; i <= allocL; i++) {
ox->dpf[i] = ox->dpf[0] + i * ox->allocQ4 * p7X_NSCELLS;
ox->dpw[i] = ox->dpw[0] + i * ox->allocQ8 * p7X_NSCELLS;
ox->dpb[i] = ox->dpb[0] + i * ox->allocQ16;
}
ox->allocXR = allocXL+1;
ESL_ALLOC(ox->x_mem, sizeof(float) * ox->allocXR * p7X_NXCELLS + 15);
ox->xmx = (float *) ((unsigned long int) ((char *) ox->x_mem + 15) & (~0xf));
ox->M = 0;
ox->L = 0;
ox->totscale = 0.0;
ox->has_own_scales = TRUE; /* most matrices are Forward, control their own scale factors */
#ifdef p7_DEBUGGING
ox->debugging = FALSE;
ox->dfp = NULL;
#endif
return ox;
ERROR:
p7_omx_Destroy(ox);
return NULL;
}
/* 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_omx_DumpVFRow()
* Synopsis: Dump current row of ViterbiFilter (int16) part of <ox> matrix.
* Incept: SRE, Wed Jul 30 16:43:21 2008 [Janelia]
*
* Purpose: Dump current row of ViterbiFilter (int16) part of DP
* matrix <ox> for diagnostics, and include the values of
* specials <xE>, etc. The index <rowi> for the current row
* is used as a row label.
*
* If <rowi> is 0, print a header first too.
*
* The output format is coordinated with <p7_gmx_Dump()> to
* facilitate comparison to a known answer.
*
* Returns: <eslOK> on success.
*
* Throws: <eslEMEM> on allocation failure.
*/
int
p7_omx_DumpVFRow(P7_OMX *ox, int rowi, int16_t xE, int16_t xN, int16_t xJ, int16_t xB, int16_t xC)
{
vector signed short *dp = ox->dpw[0]; /* must set <dp> before using {MDI}MX macros */
int M = ox->M;
int Q = p7O_NQW(M);
int16_t *v = NULL; /* array of unstriped, uninterleaved scores */
int q,z,k;
union { vector signed short v; int16_t i[8]; } tmp;
int status;
ESL_ALLOC(v, sizeof(int16_t) * ((Q*8)+1));
v[0] = 0;
/* Header (if we're on the 0th row)
*/
if (rowi == 0)
{
fprintf(ox->dfp, " ");
for (k = 0; k <= M; k++) fprintf(ox->dfp, "%6d ", k);
fprintf(ox->dfp, "%6s %6s %6s %6s %6s\n", "E", "N", "J", "B", "C");
fprintf(ox->dfp, " ");
for (k = 0; k <= M+5; k++) fprintf(ox->dfp, "%6s ", "------");
fprintf(ox->dfp, "\n");
}
/* Unpack and unstripe, then print M's. */
for (q = 0; q < Q; q++) {
tmp.v = MMXo(q);
for (z = 0; z < 8; z++) v[q+Q*z+1] = tmp.i[z];
}
fprintf(ox->dfp, "%4d M ", rowi);
for (k = 0; k <= M; k++) fprintf(ox->dfp, "%6d ", v[k]);
/* The specials */
fprintf(ox->dfp, "%6d %6d %6d %6d %6d\n", xE, xN, xJ, xB, xC);
/* Unpack and unstripe, then print I's. */
for (q = 0; q < Q; q++) {
tmp.v = IMXo(q);
for (z = 0; z < 8; z++) v[q+Q*z+1] = tmp.i[z];
}
fprintf(ox->dfp, "%4d I ", rowi);
for (k = 0; k <= M; k++) fprintf(ox->dfp, "%6d ", v[k]);
fprintf(ox->dfp, "\n");
/* Unpack, unstripe, then print D's. */
for (q = 0; q < Q; q++) {
tmp.v = DMXo(q);
for (z = 0; z < 8; z++) v[q+Q*z+1] = tmp.i[z];
}
fprintf(ox->dfp, "%4d D ", rowi);
for (k = 0; k <= M; k++) fprintf(ox->dfp, "%6d ", v[k]);
fprintf(ox->dfp, "\n\n");
free(v);
return eslOK;
ERROR:
free(v);
return status;
}
/* Function: p7_omx_GrowTo()
* Synopsis: Assure that a DP matrix is big enough.
* Incept: SRE, Thu Dec 20 09:27:07 2007 [Janelia]
*
* Purpose: Assures that an optimized DP matrix <ox> is allocated for
* a model up to <allocM> in length; if not, reallocate to
* make it so.
*
* Because the optimized matrix is one-row, only the model
* length matters; the target sequence length isn't
* relevant.
*
* Returns: <eslOK> on success, and <gx> may be reallocated upon
* return; any data that may have been in <gx> must be
* assumed to be invalidated.
*
* Throws: <eslEMEM> on allocation failure, and any data that may
* have been in <gx> must be assumed to be invalidated.
*/
int
p7_omx_GrowTo(P7_OMX *ox, int allocM, int allocL, int allocXL)
{
void *p;
int nqf = p7O_NQF(allocM); /* segment length; total # of striped vectors for uchar */
int nqw = p7O_NQW(allocM); /* segment length; total # of striped vectors for float */
int nqb = p7O_NQB(allocM); /* segment length; total # of striped vectors for float */
size_t ncells = (allocL+1) * nqf * 4;
int reset_row_pointers = FALSE;
int i;
int status;
/* If all possible dimensions are already satisfied, the matrix is fine */
if (ox->allocQ4*4 >= allocM && ox->validR > allocL && ox->allocXR >= allocXL+1) return eslOK;
/* If the main matrix is too small in cells, reallocate it;
* and we'll need to realign/reset the row pointers later.
*/
if (ncells > ox->ncells)
{
ESL_RALLOC(ox->dp_mem, p, sizeof(vector float) * (allocL+1) * nqf * p7X_NSCELLS + 15);
ox->ncells = ncells;
reset_row_pointers = TRUE;
}
/* If the X beams are too small, reallocate them. */
if (allocXL+1 >= ox->allocXR)
{
ESL_RALLOC(ox->x_mem, p, sizeof(float) * (allocXL+1) * p7X_NXCELLS + 15);
ox->allocXR = allocXL+1;
ox->xmx = (float *) ((unsigned long int) ((char *) ox->x_mem + 15) & (~0xf));
}
/* If there aren't enough rows, reallocate the row pointers; we'll
* realign and reset them later.
*/
if (allocL >= ox->allocR)
{
ESL_RALLOC(ox->dpb, p, sizeof(vector unsigned char *) * (allocL+1));
ESL_RALLOC(ox->dpw, p, sizeof(vector signed short * ) * (allocL+1));
ESL_RALLOC(ox->dpf, p, sizeof(vector float *) * (allocL+1));
ox->allocR = allocL+1;
reset_row_pointers = TRUE;
}
/* must we widen the rows? */
if (allocM > ox->allocQ4*4)
reset_row_pointers = TRUE;
/* must we set some more valid row pointers? */
if (allocL >= ox->validR)
reset_row_pointers = TRUE;
/* now reset the row pointers, if needed */
if (reset_row_pointers)
{
ox->dpb[0] = (vector unsigned char *) ((unsigned long int) ((char *) ox->dp_mem + 15) & (~0xf));
ox->dpw[0] = (vector signed short *) ox->dpb[0];
ox->dpf[0] = (vector float *) ox->dpb[0];
ox->validR = ESL_MIN( ox->ncells / (nqf * 4), ox->allocR);
for (i = 1; i < ox->validR; i++)
{
ox->dpb[i] = ox->dpb[0] + i * nqb;
ox->dpw[i] = ox->dpw[0] + i * nqw * p7X_NSCELLS;
ox->dpf[i] = ox->dpf[0] + i * nqf * p7X_NSCELLS;
}
ox->allocQ4 = nqf;
ox->allocQ8 = nqw;
ox->allocQ16 = nqb;
}
ox->M = 0;
ox->L = 0;
return eslOK;
ERROR:
return status;
}
/* Function: p7_oprofile_Write()
* Synopsis: Write an optimized profile in two files.
* Incept: SRE, Wed Jan 21 10:35:28 2009 [Janelia]
*
* Purpose: Write the MSV filter part of <om> to open binary stream
* <ffp>, and the rest of the model to <pfp>. These two
* streams will typically be <.h3f> and <.h3p> files
* being created by hmmpress.
*
* Args: ffp - open binary stream for saving MSV filter part
* pfp - open binary stream for saving rest of profile
* om - optimized profile to save
*
* Returns: <eslOK> on success.
*
* Returns <eslFAIL> on any write failure; for example,
* if disk is full.
*
* Throws: (no abnormal error conditions)
*/
int
p7_oprofile_Write(FILE *ffp, FILE *pfp, P7_OPROFILE *om)
{
int Q4 = p7O_NQF(om->M);
int Q8 = p7O_NQW(om->M);
int Q16 = p7O_NQB(om->M);
int n = strlen(om->name);
int x;
/* <ffp> is the part of the oprofile that MSVFilter() needs */
if (fwrite((char *) &(v3b_fmagic), sizeof(uint32_t), 1, ffp) != 1) return eslFAIL;
if (fwrite((char *) &(om->M), sizeof(int), 1, ffp) != 1) return eslFAIL;
if (fwrite((char *) &(om->abc->type), sizeof(int), 1, ffp) != 1) return eslFAIL;
if (fwrite((char *) &n, sizeof(int), 1, ffp) != 1) return eslFAIL;
if (fwrite((char *) om->name, sizeof(char), n+1, ffp) != n+1) return eslFAIL;
if (fwrite((char *) &(om->tbm_b), sizeof(uint8_t), 1, ffp) != 1) return eslFAIL;
if (fwrite((char *) &(om->tec_b), sizeof(uint8_t), 1, ffp) != 1) return eslFAIL;
if (fwrite((char *) &(om->tjb_b), sizeof(uint8_t), 1, ffp) != 1) return eslFAIL;
if (fwrite((char *) &(om->scale_b), sizeof(float), 1, ffp) != 1) return eslFAIL;
if (fwrite((char *) &(om->base_b), sizeof(uint8_t), 1, ffp) != 1) return eslFAIL;
if (fwrite((char *) &(om->bias_b), sizeof(uint8_t), 1, ffp) != 1) return eslFAIL;
for (x = 0; x < om->abc->Kp; x++)
if (fwrite( (char *) om->rbv[x], sizeof(__m128i), Q16, ffp) != Q16) return eslFAIL;
if (fwrite((char *) om->evparam, sizeof(float), p7_NEVPARAM, ffp) != p7_NEVPARAM) return eslFAIL;
if (fwrite((char *) om->offs, sizeof(off_t), p7_NOFFSETS, ffp) != p7_NOFFSETS) return eslFAIL;
if (fwrite((char *) om->compo, sizeof(float), p7_MAXABET, ffp) != p7_MAXABET) return eslFAIL;
/* <pfp> gets the rest of the oprofile */
if (fwrite((char *) &(v3b_pmagic), sizeof(uint32_t), 1, pfp) != 1) return eslFAIL;
if (fwrite((char *) &(om->M), sizeof(int), 1, pfp) != 1) return eslFAIL;
if (fwrite((char *) &(om->abc->type), sizeof(int), 1, pfp) != 1) return eslFAIL;
if (fwrite((char *) &n, sizeof(int), 1, pfp) != 1) return eslFAIL;
if (fwrite((char *) om->name, sizeof(char), n+1, pfp) != n+1) return eslFAIL;
if (om->acc == NULL) {
n = 0;
if (fwrite((char *) &n, sizeof(int), 1, pfp) != 1) return eslFAIL;
} else {
n = strlen(om->acc);
if (fwrite((char *) &n, sizeof(int), 1, pfp) != 1) return eslFAIL;
if (fwrite((char *) om->acc, sizeof(char), n+1, pfp) != n+1) return eslFAIL;
}
if (om->desc == NULL) {
n = 0;
if (fwrite((char *) &n, sizeof(int), 1, pfp) != 1) return eslFAIL;
} else {
n = strlen(om->desc);
if (fwrite((char *) &n, sizeof(int), 1, pfp) != 1) return eslFAIL;
if (fwrite((char *) om->desc, sizeof(char), n+1, pfp) != n+1) return eslFAIL;
}
if (fwrite((char *) om->rf, sizeof(char), om->M+2, pfp) != om->M+2) return eslFAIL;
if (fwrite((char *) om->cs, sizeof(char), om->M+2, pfp) != om->M+2) return eslFAIL;
if (fwrite((char *) om->consensus, sizeof(char), om->M+2, pfp) != om->M+2) return eslFAIL;
/* ViterbiFilter part */
if (fwrite((char *) om->twv, sizeof(__m128i), 8*Q8, pfp) != 8*Q8) return eslFAIL;
for (x = 0; x < om->abc->Kp; x++)
if (fwrite( (char *) om->rwv[x], sizeof(__m128i), Q8, pfp) != Q8) return eslFAIL;
for (x = 0; x < p7O_NXSTATES; x++)
if (fwrite( (char *) om->xw[x], sizeof(int16_t), p7O_NXTRANS, pfp) != p7O_NXTRANS) return eslFAIL;
if (fwrite((char *) &(om->scale_w), sizeof(float), 1, pfp) != 1) return eslFAIL;
if (fwrite((char *) &(om->base_w), sizeof(int16_t), 1, pfp) != 1) return eslFAIL;
if (fwrite((char *) &(om->ddbound_w), sizeof(int16_t), 1, pfp) != 1) return eslFAIL;
if (fwrite((char *) &(om->ncj_roundoff), sizeof(float), 1, pfp) != 1) return eslFAIL;
/* Forward/Backward part */
if (fwrite((char *) om->tfv, sizeof(__m128), 8*Q4, pfp) != 8*Q4) return eslFAIL;
for (x = 0; x < om->abc->Kp; x++)
if (fwrite( (char *) om->rfv[x], sizeof(__m128), Q4, pfp) != Q4) return eslFAIL;
for (x = 0; x < p7O_NXSTATES; x++)
if (fwrite( (char *) om->xf[x], sizeof(float), p7O_NXTRANS, pfp) != p7O_NXTRANS) return eslFAIL;
if (fwrite((char *) om->cutoff, sizeof(float), p7_NCUTOFFS, pfp) != p7_NCUTOFFS) return eslFAIL;
if (fwrite((char *) &(om->nj), sizeof(float), 1, pfp) != 1) return eslFAIL;
if (fwrite((char *) &(om->mode), sizeof(int), 1, pfp) != 1) return eslFAIL;
if (fwrite((char *) &(om->L) , sizeof(int), 1, pfp) != 1) return eslFAIL;
return eslOK;
}
/* Function: p7_oprofile_ReadRest()
* Synopsis: Read the rest of an optimized profile.
* Incept: SRE, Wed Jan 21 11:04:56 2009 [Janelia]
*
* Purpose: Read the rest of an optimized profile <om> from
* the <.h3p> file associated with an open HMM
* file <hfp>.
*
* This is the second part of a two-part calling sequence.
* The <om> here must be the result of a previous
* successful <p7_oprofile_ReadMSV()> call on the same
* open <hfp>.
*
* Args: hfp - open HMM file, from which we've previously
* called <p7_oprofile_ReadMSV()>.
* om - optimized profile that was successfully
* returned by <p7_oprofile_ReadMSV()>.
*
* Returns: <eslOK> on success, and <om> is now a complete
* optimized profile.
*
* Returns <eslEFORMAT> if <hfp> has no <.h3p> file open,
* or on any parsing error, and set <hfp->errbuf> to
* an informative error message.
*
* Throws: <eslESYS> if an <fseek()> fails to reposition the
* binary <.h3p> file.
*
* <eslEMEM> on allocation error.
*/
int
p7_oprofile_ReadRest(P7_HMMFILE *hfp, P7_OPROFILE *om)
{
uint32_t magic;
int M, Q4, Q8;
int x,n;
char *name = NULL;
int alphatype;
int status;
#ifdef HMMER_THREADS
/* lock the mutex to prevent other threads from reading from the optimized
* profile at the same time.
*/
if (hfp->syncRead)
{
if (pthread_mutex_lock (&hfp->readMutex) != 0) ESL_EXCEPTION(eslESYS, "mutex lock failed");
}
#endif
if (hfp->errbuf != NULL) hfp->errbuf[0] = '\0';
if (hfp->pfp == NULL) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "no MSV profile file; hmmpress probably wasn't run");
/* Position the <hfp->pfp> using offset stored in <om> */
if (fseeko(hfp->pfp, om->offs[p7_POFFSET], SEEK_SET) != 0) ESL_EXCEPTION(eslESYS, "fseeko() failed");
if (! fread( (char *) &magic, sizeof(uint32_t), 1, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read magic");
if (magic == v3a_pmagic) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "this is an outdated HMM database (3/a format); please hmmpress your HMM file again");
if (magic != v3b_pmagic) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "bad magic; not an HMM database file?");
if (! fread( (char *) &M, sizeof(int), 1, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read model size M");
if (! fread( (char *) &alphatype, sizeof(int), 1, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read alphabet type");
if (! fread( (char *) &n, sizeof(int), 1, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read name length");
if (M != om->M) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "p/f model length mismatch");
if (alphatype != om->abc->type) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "p/f alphabet type mismatch");
ESL_ALLOC(name, sizeof(char) * (n+1));
if (! fread( (char *) name, sizeof(char), n+1, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read name");
if (strcmp(name, om->name) != 0) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "p/f name mismatch");
if (! fread((char *) &n, sizeof(int), 1, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read accession length");
if (n > 0) {
ESL_ALLOC(om->acc, sizeof(char) * (n+1));
if (! fread( (char *) om->acc, sizeof(char), n+1, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read accession");
}
if (! fread((char *) &n, sizeof(int), 1, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read description length");
if (n > 0) {
ESL_ALLOC(om->desc, sizeof(char) * (n+1));
if (! fread( (char *) om->desc, sizeof(char), n+1, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read description");
}
if (! fread((char *) om->rf, sizeof(char), M+2, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read rf annotation");
if (! fread((char *) om->cs, sizeof(char), M+2, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read cs annotation");
if (! fread((char *) om->consensus, sizeof(char), M+2, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read consensus annotation");
Q4 = p7O_NQF(om->M);
Q8 = p7O_NQW(om->M);
if (! fread((char *) om->twv, sizeof(__m128i), 8*Q8, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read <tu>, vitfilter transitions");
for (x = 0; x < om->abc->Kp; x++)
if (! fread( (char *) om->rwv[x], sizeof(__m128i), Q8, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read <ru>[%d], vitfilter emissions for sym %c", x, om->abc->sym[x]);
for (x = 0; x < p7O_NXSTATES; x++)
if (! fread( (char *) om->xw[x], sizeof(int16_t), p7O_NXTRANS, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read <xu>[%d], vitfilter special transitions", x);
if (! fread((char *) &(om->scale_w), sizeof(float), 1, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read scale_w");
if (! fread((char *) &(om->base_w), sizeof(int16_t), 1, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read base_w");
if (! fread((char *) &(om->ddbound_w), sizeof(int16_t), 1, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read ddbound_w");
if (! fread((char *) &(om->ncj_roundoff), sizeof(float), 1, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read ddbound_w");
if (! fread((char *) om->tfv, sizeof(__m128), 8*Q4, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read <tf> transitions");
for (x = 0; x < om->abc->Kp; x++)
if (! fread( (char *) om->rfv[x], sizeof(__m128), Q4, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read <rf>[%d] emissions for sym %c", x, om->abc->sym[x]);
for (x = 0; x < p7O_NXSTATES; x++)
if (! fread( (char *) om->xf[x], sizeof(float), p7O_NXTRANS, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read <xf>[%d] special transitions", x);
if (! fread((char *) om->cutoff, sizeof(float), p7_NCUTOFFS, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read Pfam score cutoffs");
if (! fread((char *) &(om->nj), sizeof(float), 1, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read nj");
if (! fread((char *) &(om->mode), sizeof(int), 1, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read mode");
if (! fread((char *) &(om->L) , sizeof(int), 1, hfp->pfp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read L");
#ifdef HMMER_THREADS
if (hfp->syncRead)
{
if (pthread_mutex_unlock (&hfp->readMutex) != 0) ESL_EXCEPTION(eslESYS, "mutex unlock failed");
}
#endif
free(name);
return eslOK;
ERROR:
#ifdef HMMER_THREADS
if (hfp->syncRead)
{
if (pthread_mutex_unlock (&hfp->readMutex) != 0) ESL_EXCEPTION(eslESYS, "mutex unlock failed");
}
#endif
if (name != NULL) free(name);
return status;
}
/* Function: p7_oprofile_MPIUnpack()
* Synopsis: Unpacks an OPROFILE from an MPI buffer.
* Incept: MSF, Wed Oct 21, 2009 [Janelia]
*
* Purpose: Unpack a newly allocated OPROFILE from MPI packed buffer
* <buf>, starting from position <*pos>, where the total length
* of the buffer in bytes is <n>.
*
* Caller may or may not already know what alphabet the OPROFILE
* is expected to be in. A reference to the current
* alphabet is passed in <abc>. If the alphabet is unknown,
* pass <*abc = NULL>, and when the OPROFILE is received, an
* appropriate new alphabet object is allocated and passed
* back to the caller via <*abc>. If the alphabet is
* already known, <*abc> is that alphabet, and the new
* OPROFILE's alphabet type is verified to agree with it. This
* mechanism allows an application to let the first OPROFILE
* determine the alphabet type for the application, while
* still keeping the alphabet under the application's scope
* of control.
*
* Returns: <eslOK> on success. <*pos> is updated to the position of
* the next element in <buf> to unpack (if any). <*ret_om>
* contains a newly allocated OPROFILE, which the caller is
* responsible for free'ing. If <*abc> was passed as
* <NULL>, it now points to an <ESL_ALPHABET> object that
* was allocated here; caller is responsible for free'ing
* this.
*
* Returns <eslEINCOMPAT> if the OPROFILE is in a different
* alphabet than <*abc> said to expect. In this case,
* <*abc> is unchanged, <*buf> and <*nalloc> may have been
* changed, and <*ret_om> is <NULL>.
*
* Throws: <eslESYS> on an MPI call failure. <eslEMEM> on allocation failure.
* In either case, <*ret_om> is <NULL>, and the state of <buf>
* and <*pos> is undefined and should be considered to be corrupted.
*/
int
p7_oprofile_MPIUnpack(char *buf, int n, int *pos, MPI_Comm comm, ESL_ALPHABET **abc, P7_OPROFILE **ret_om)
{
int status;
int M, K, atype;
int len;
int x;
int Q4, Q8, Q16;
int vsz = sizeof(vector float);
P7_OPROFILE *om = NULL;
if (MPI_Unpack(buf, n, pos, &M, 1, MPI_INT, comm) != 0) ESL_XEXCEPTION(eslESYS, "mpi unpack failed");
if (MPI_Unpack(buf, n, pos, &atype, 1, MPI_INT, comm) != 0) ESL_XEXCEPTION(eslESYS, "mpi unpack failed");
/* Set or verify the alphabet */
if (*abc == NULL) { /* still unknown: set it, pass control of it back to caller */
if ((*abc = esl_alphabet_Create(atype)) == NULL) { status = eslEMEM; goto ERROR; }
} else { /* already known: check it */
if ((*abc)->type != atype) { status = eslEINCOMPAT; goto ERROR; }
}
Q4 = p7O_NQF(M);
Q8 = p7O_NQW(M);
Q16 = p7O_NQB(M);
if ((om = p7_oprofile_Create(M, *abc)) == NULL) { status = eslEMEM; goto ERROR; }
om->M = M;
K = (*abc)->Kp;
/* model configuration */
if (MPI_Unpack(buf, n, pos, &om->L, 1, MPI_INT, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
if (MPI_Unpack(buf, n, pos, &om->mode, 1, MPI_INT, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
if (MPI_Unpack(buf, n, pos, &om->nj, 1, MPI_FLOAT, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
/* MSV Filter information */
if (MPI_Unpack(buf, n, pos, &om->tbm_b, 1, MPI_CHAR, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
if (MPI_Unpack(buf, n, pos, &om->tec_b, 1, MPI_CHAR, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
if (MPI_Unpack(buf, n, pos, &om->tjb_b, 1, MPI_CHAR, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
if (MPI_Unpack(buf, n, pos, &om->scale_b, 1, MPI_FLOAT, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
if (MPI_Unpack(buf, n, pos, &om->base_b, 1, MPI_CHAR, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
if (MPI_Unpack(buf, n, pos, &om->bias_b, 1, MPI_CHAR, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
for (x = 0; x < K; x++)
if (MPI_Unpack(buf, n, pos, om->rbv[x], vsz*Q16, MPI_CHAR, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
/* Viterbi Filter information */
if (MPI_Unpack(buf, n, pos, &om->scale_w, 1, MPI_FLOAT, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
if (MPI_Unpack(buf, n, pos, &om->base_w, 1, MPI_SHORT, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
if (MPI_Unpack(buf, n, pos, &om->ddbound_w, 1, MPI_SHORT, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
if (MPI_Unpack(buf, n, pos, &om->ncj_roundoff, 1, MPI_FLOAT, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
if (MPI_Unpack(buf, n, pos, om->twv, 8*vsz*Q8, MPI_CHAR, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
for (x = 0; x < p7O_NXSTATES; x++)
if (MPI_Unpack(buf, n, pos, om->xw[x], p7O_NXTRANS, MPI_SHORT, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
for (x = 0; x < K; x++)
if (MPI_Unpack(buf, n, pos, om->rwv[x], vsz*Q8, MPI_CHAR, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
/* Forward/Backward information */
if (MPI_Unpack(buf, n, pos, om->tfv, 8*vsz*Q4, MPI_CHAR, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
for (x = 0; x < p7O_NXSTATES; x++)
if (MPI_Unpack(buf, n, pos, om->xf[x], p7O_NXTRANS, MPI_FLOAT, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
for (x = 0; x < K; x++)
if (MPI_Unpack(buf, n, pos, om->rfv[x], vsz*Q4, MPI_CHAR, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
/* Forward/Backward information */
if (MPI_Unpack(buf, n, pos, om->offs, p7_NOFFSETS, MPI_LONG_LONG_INT, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
if (MPI_Unpack(buf, n, pos, &om->roff, 1, MPI_LONG_LONG_INT, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
if (MPI_Unpack(buf, n, pos, &om->eoff, 1, MPI_LONG_LONG_INT, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
/* Annotation information */
if (MPI_Unpack(buf, n, pos, &len, 1, MPI_INT, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
if (len > 0) {
ESL_ALLOC(om->name, len);
if (MPI_Unpack(buf, n, pos, om->name, len, MPI_CHAR, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
om->name[len-1] = '\0';
}
if (MPI_Unpack(buf, n, pos, &len, 1, MPI_INT, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
if (len > 0) {
ESL_ALLOC(om->acc, len);
if (MPI_Unpack(buf, n, pos, om->acc, len, MPI_CHAR, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
om->acc[len-1] = '\0';
}
if (MPI_Unpack(buf, n, pos, &len, 1, MPI_INT, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
if (len > 0) {
ESL_ALLOC(om->desc, len);
if (MPI_Unpack(buf, n, pos, om->desc, len, MPI_CHAR, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
om->desc[len-1] = '\0';
}
if (MPI_Unpack(buf, n, pos, &len, 1, MPI_INT, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
if (len > 0) {
ESL_ALLOC(om->rf, len);
if (MPI_Unpack(buf, n, pos, om->rf, len, MPI_CHAR, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
om->rf[len-1] = '\0';
}
if (MPI_Unpack(buf, n, pos, &len, 1, MPI_INT, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
if (len > 0) {
ESL_ALLOC(om->cs, len);
if (MPI_Unpack(buf, n, pos, om->cs, len, MPI_CHAR, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
om->cs[len-1] = '\0';
}
if (MPI_Unpack(buf, n, pos, &len, 1, MPI_INT, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
if (len > 0) {
ESL_ALLOC(om->consensus, len);
if (MPI_Unpack(buf, n, pos, om->consensus, len, MPI_CHAR, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
om->consensus[len-1] = '\0';
}
if (MPI_Unpack(buf, n, pos, om->evparam, p7_NEVPARAM, MPI_FLOAT, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
if (MPI_Unpack(buf, n, pos, om->cutoff, p7_NCUTOFFS, MPI_FLOAT, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
if (MPI_Unpack(buf, n, pos, om->compo, p7_MAXABET, MPI_FLOAT, comm) != 0) ESL_EXCEPTION(eslESYS, "mpi unpack failed");
*ret_om = om;
return eslOK;
ERROR:
if (om != NULL) p7_oprofile_Destroy(om);
return status;
}
/* Function: p7_oprofile_MPIPack()
* Synopsis: Packs an OPROFILE into MPI buffer.
* Incept: MSF, Wed Oct 21, 2009 [Janelia]
*
* Purpose: Packs OPROFILE <om> into an MPI packed message buffer <buf>
* of length <n> bytes, starting at byte position <*position>,
* for MPI communicator <comm>.
*
* The caller must know that <buf>'s allocation of <n>
* bytes is large enough to append the packed OPROFILE at
* position <*pos>. This typically requires a call to
* <p7_oprofile_MPIPackSize()> first, and reallocation if
* needed.
*
* Returns: <eslOK> on success; <buf> now contains the
* packed <om>, and <*position> is set to the byte
* immediately following the last byte of the OPROFILE
* in <buf>.
*
* Throws: <eslESYS> if an MPI call fails; or <eslEMEM> if the
* buffer's length <n> was overflowed in trying to pack
* <msa> into <buf>. In either case, the state of
* <buf> and <*position> is undefined, and both should
* be considered to be corrupted.
*/
int
p7_oprofile_MPIPack(P7_OPROFILE *om, char *buf, int n, int *pos, MPI_Comm comm)
{
int K = om->abc->Kp;
int atype = om->abc->type;
int len;
int x;
int Q4 = p7O_NQF(om->M);
int Q8 = p7O_NQW(om->M);
int Q16 = p7O_NQB(om->M);
int vsz = sizeof(vector float);
/* model configuration */
if (MPI_Pack(&om->M, 1, MPI_INT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (MPI_Pack(&atype, 1, MPI_INT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (MPI_Pack(&om->L, 1, MPI_INT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (MPI_Pack(&om->mode, 1, MPI_INT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (MPI_Pack(&om->nj, 1, MPI_FLOAT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
/* MSV Filter information */
if (MPI_Pack(&om->tbm_b, 1, MPI_CHAR, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (MPI_Pack(&om->tec_b, 1, MPI_CHAR, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (MPI_Pack(&om->tjb_b, 1, MPI_CHAR, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (MPI_Pack(&om->scale_b, 1, MPI_FLOAT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (MPI_Pack(&om->base_b, 1, MPI_CHAR, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (MPI_Pack(&om->bias_b, 1, MPI_CHAR, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
for (x = 0; x < K; x++)
if (MPI_Pack( om->rbv[x], vsz*Q16, MPI_CHAR, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
/* Viterbi Filter information */
if (MPI_Pack(&om->scale_w, 1, MPI_FLOAT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (MPI_Pack(&om->base_w, 1, MPI_SHORT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (MPI_Pack(&om->ddbound_w, 1, MPI_SHORT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (MPI_Pack(&om->ncj_roundoff, 1, MPI_FLOAT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (MPI_Pack( om->twv, 8*vsz*Q8, MPI_CHAR, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
for (x = 0; x < p7O_NXSTATES; x++)
if (MPI_Pack( om->xw[x], p7O_NXTRANS, MPI_SHORT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
for (x = 0; x < K; x++)
if (MPI_Pack( om->rwv[x], vsz*Q8, MPI_CHAR, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
/* Forward/Backward information */
if (MPI_Pack( om->tfv, 8*vsz*Q4, MPI_CHAR, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
for (x = 0; x < p7O_NXSTATES; x++)
if (MPI_Pack( om->xf[x], p7O_NXTRANS, MPI_FLOAT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
for (x = 0; x < K; x++)
if (MPI_Pack( om->rfv[x], vsz*Q4, MPI_CHAR, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
/* Forward/Backward information */
if (MPI_Pack( om->offs, p7_NOFFSETS, MPI_LONG_LONG_INT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (MPI_Pack(&om->roff, 1, MPI_LONG_LONG_INT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (MPI_Pack(&om->eoff, 1, MPI_LONG_LONG_INT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
/* Annotation information */
len = (om->name != NULL) ? strlen(om->name)+1 : 0;
if (MPI_Pack(&len, 1, MPI_INT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (len > 0)
if (MPI_Pack( om->name, len, MPI_CHAR, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
len = (om->acc != NULL) ? strlen(om->acc)+1 : 0;
if (MPI_Pack(&len, 1, MPI_INT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (len > 0)
if (MPI_Pack( om->acc, len, MPI_CHAR, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
len = (om->desc != NULL) ? strlen(om->desc)+1 : 0;
if (MPI_Pack(&len, 1, MPI_INT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (len > 0)
if (MPI_Pack( om->desc, len, MPI_CHAR, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
len = (om->rf != NULL) ? strlen(om->rf)+1 : 0;
if (MPI_Pack(&len, 1, MPI_INT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (len > 0)
if (MPI_Pack( om->rf, len, MPI_CHAR, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
len = (om->cs != NULL) ? strlen(om->cs)+1 : 0;
if (MPI_Pack(&len, 1, MPI_INT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (len > 0)
if (MPI_Pack( om->cs, len, MPI_CHAR, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
len = (om->consensus != NULL) ? strlen(om->consensus)+1 : 0;
if (MPI_Pack(&len, 1, MPI_INT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (len > 0)
if (MPI_Pack( om->consensus, len, MPI_CHAR, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (MPI_Pack( om->evparam, p7_NEVPARAM, MPI_FLOAT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (MPI_Pack( om->cutoff, p7_NCUTOFFS, MPI_FLOAT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (MPI_Pack( om->compo, p7_MAXABET, MPI_FLOAT, buf, n, pos, comm) != 0) ESL_EXCEPTION(eslESYS, "pack failed");
if (*pos > n) ESL_EXCEPTION(eslEMEM, "buffer overflow");
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;
}