/* 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
}
/* Function: p7_oprofile_Write()
* Synopsis: Write an optimized profile in two files.
*
* 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.
*
* Throws: <eslEWRITE> on any write failure, such as filling
* the disk.
*/
int
p7_oprofile_Write(FILE *ffp, FILE *pfp, P7_OPROFILE *om)
{
int Q4 = P7_NVF(om->M);
int Q8 = P7_NVW(om->M);
int Q16 = P7_NVB(om->M);
int Q16x = P7_NVB(om->M) + p7O_EXTRA_SB;
int n = strlen(om->name);
int x;
/* <ffp> is the part of the oprofile that MSVFilter() needs */
if (fwrite((char *) &(v3f_fmagic), sizeof(uint32_t), 1, ffp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) &(om->M), sizeof(int), 1, ffp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) &(om->abc->type), sizeof(int), 1, ffp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) &n, sizeof(int), 1, ffp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) om->name, sizeof(char), n+1, ffp) != n+1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) &(om->max_length),sizeof(int), 1, ffp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) &(om->tbm_b), sizeof(uint8_t), 1, ffp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) &(om->tec_b), sizeof(uint8_t), 1, ffp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) &(om->tjb_b), sizeof(uint8_t), 1, ffp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) &(om->scale_b), sizeof(float), 1, ffp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) &(om->base_b), sizeof(uint8_t), 1, ffp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) &(om->bias_b), sizeof(uint8_t), 1, ffp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
for (x = 0; x < om->abc->Kp; x++)
if (fwrite( (char *) om->sbv[x], sizeof(__m128i), Q16x, ffp) != Q16x) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
for (x = 0; x < om->abc->Kp; x++)
if (fwrite( (char *) om->rbv[x], sizeof(__m128i), Q16, ffp) != Q16) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) om->evparam, sizeof(float), p7_NEVPARAM, ffp) != p7_NEVPARAM) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) om->offs, sizeof(off_t), p7_NOFFSETS, ffp) != p7_NOFFSETS) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) om->compo, sizeof(float), p7_MAXABET, ffp) != p7_MAXABET) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) &(v3f_fmagic), sizeof(uint32_t), 1, ffp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed"); /* sentinel */
/* <pfp> gets the rest of the oprofile */
if (fwrite((char *) &(v3f_pmagic), sizeof(uint32_t), 1, pfp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) &(om->M), sizeof(int), 1, pfp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) &(om->abc->type), sizeof(int), 1, pfp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) &n, sizeof(int), 1, pfp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) om->name, sizeof(char), n+1, pfp) != n+1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (om->acc == NULL) {
n = 0;
if (fwrite((char *) &n, sizeof(int), 1, pfp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
} else {
n = strlen(om->acc);
if (fwrite((char *) &n, sizeof(int), 1, pfp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) om->acc, sizeof(char), n+1, pfp) != n+1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
}
if (om->desc == NULL) {
n = 0;
if (fwrite((char *) &n, sizeof(int), 1, pfp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
} else {
n = strlen(om->desc);
if (fwrite((char *) &n, sizeof(int), 1, pfp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) om->desc, sizeof(char), n+1, pfp) != n+1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
}
if (fwrite((char *) om->rf, sizeof(char), om->M+2, pfp) != om->M+2) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) om->mm, sizeof(char), om->M+2, pfp) != om->M+2) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) om->cs, sizeof(char), om->M+2, pfp) != om->M+2) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) om->consensus, sizeof(char), om->M+2, pfp) != om->M+2) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
/* ViterbiFilter part */
if (fwrite((char *) om->twv, sizeof(__m128i), 8*Q8, pfp) != 8*Q8) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
for (x = 0; x < om->abc->Kp; x++)
if (fwrite( (char *) om->rwv[x], sizeof(__m128i), Q8, pfp) != Q8) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
for (x = 0; x < p7O_NXSTATES; x++)
if (fwrite( (char *) om->xw[x], sizeof(int16_t), p7O_NXTRANS, pfp) != p7O_NXTRANS) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) &(om->scale_w), sizeof(float), 1, pfp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) &(om->base_w), sizeof(int16_t), 1, pfp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) &(om->ddbound_w), sizeof(int16_t), 1, pfp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) &(om->ncj_roundoff), sizeof(float), 1, pfp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
/* Forward/Backward part */
if (fwrite((char *) om->tfv, sizeof(__m128), 8*Q4, pfp) != 8*Q4) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
for (x = 0; x < om->abc->Kp; x++)
if (fwrite( (char *) om->rfv[x], sizeof(__m128), Q4, pfp) != Q4) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
for (x = 0; x < p7O_NXSTATES; x++)
if (fwrite( (char *) om->xf[x], sizeof(float), p7O_NXTRANS, pfp) != p7O_NXTRANS) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) om->cutoff, sizeof(float), p7_NCUTOFFS, pfp) != p7_NCUTOFFS) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) &(om->nj), sizeof(float), 1, pfp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) &(om->mode), sizeof(int), 1, pfp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) &(om->L) , sizeof(int), 1, pfp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed");
if (fwrite((char *) &(v3f_pmagic), sizeof(uint32_t), 1, pfp) != 1) ESL_EXCEPTION_SYS(eslEWRITE, "oprofile write failed"); /* sentinel */
return eslOK;
}
/* Function: p7_oprofile_ReadMSV()
* Synopsis: Read MSV filter part of an optimized profile.
*
* Purpose: Read the MSV filter part of a profile from the
* <.h3f> file associated with an open HMM file <hfp>.
* Allocate a new model, populate it with this minimal
* MSV filter information, and return a pointer to it
* in <*ret_om>.
*
* Our alphabet may get set by the first HMM we read. If
* <*byp_abc> is <NULL> at start, create a new alphabet and
* return a pointer to it in <*byp_abc>. If <*byp_abc> is
* non-<NULL>, it is assumed to be a pointer to an existing
* alphabet; we verify that the HMM's alphabet matches it
* and <*ret_abc> isn't changed. This is the same
* convention used by <p7_hmmfile_Read()>.
*
* The <.h3f> file was opened automatically, if it existed,
* when the HMM file was opened with <p7_hmmfile_OpenE()>.
*
* When no more HMMs remain in the file, return <eslEOF>.
*
* Args: hfp - open HMM file, with associated .h3p file
* byp_abc - BYPASS: <*byp_abc == ESL_ALPHABET *> if known;
* <*byp_abc == NULL> if desired;
* <NULL> if unwanted.
* ret_om - RETURN: newly allocated <om> with MSV filter
* data filled in.
*
* Returns: <eslOK> on success. <*ret_om> is allocated here;
* caller free's with <p7_oprofile_Destroy()>.
* <*byp_abc> is allocated here if it was requested;
* caller free's with <esl_alphabet_Destroy()>.
*
* Returns <eslEFORMAT> if <hfp> has no <.h3f> file open,
* or on any parsing error.
*
* Returns <eslEINCOMPAT> if the HMM we read is incompatible
* with the existing alphabet <*byp_abc> led us to expect.
*
* On any returned error, <hfp->errbuf> contains an
* informative error message.
*
* Throws: <eslEMEM> on allocation error.
*/
int
p7_oprofile_ReadMSV(P7_HMMFILE *hfp, ESL_ALPHABET **byp_abc, P7_OPROFILE **ret_om)
{
P7_OPROFILE *om = NULL;
ESL_ALPHABET *abc = NULL;
uint32_t magic;
off_t roff;
int M, Q16, Q16x;
int x,n;
int alphatype;
int status;
hfp->errbuf[0] = '\0';
if (hfp->ffp == NULL) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "no MSV profile file; hmmpress probably wasn't run");
if (feof(hfp->ffp)) { status = eslEOF; goto ERROR; } /* normal EOF: no more profiles */
/* keep track of the starting offset of the MSV model */
roff = ftello(hfp->ffp);
if (! fread( (char *) &magic, sizeof(uint32_t), 1, hfp->ffp)) { status = eslEOF; goto ERROR; }
if (magic == v3a_fmagic) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "binary auxfiles are in an outdated HMMER format (3/a); please hmmpress your HMM file again");
if (magic == v3b_fmagic) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "binary auxfiles are in an outdated HMMER format (3/b); please hmmpress your HMM file again");
if (magic == v3c_fmagic) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "binary auxfiles are in an outdated HMMER format (3/c); please hmmpress your HMM file again");
if (magic == v3d_fmagic) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "binary auxfiles are in an outdated HMMER format (3/d); please hmmpress your HMM file again");
if (magic == v3e_fmagic) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "binary auxfiles are in an outdated HMMER format (3/e); please hmmpress your HMM file again");
if (magic != v3f_fmagic) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "bad magic; not an HMM database?");
if (! fread( (char *) &M, sizeof(int), 1, hfp->ffp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read model size M");
if (! fread( (char *) &alphatype, sizeof(int), 1, hfp->ffp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read alphabet type");
Q16 = P7_NVB(M);
Q16x = P7_NVB(M) + p7O_EXTRA_SB;
/* Set or verify alphabet. */
if (byp_abc == NULL || *byp_abc == NULL) { /* alphabet unknown: whether wanted or unwanted, make a new one */
if ((abc = esl_alphabet_Create(alphatype)) == NULL) ESL_XFAIL(eslEMEM, hfp->errbuf, "allocation failed: alphabet");
} else { /* alphabet already known: verify it against what we see in the HMM */
abc = *byp_abc;
if (abc->type != alphatype)
ESL_XFAIL(eslEINCOMPAT, hfp->errbuf, "Alphabet type mismatch: was %s, but current profile says %s",
esl_abc_DecodeType(abc->type), esl_abc_DecodeType(alphatype));
}
/* Now we know the sizes of things, so we can allocate. */
if ((om = p7_oprofile_Create(M, abc)) == NULL) ESL_XFAIL(eslEMEM, hfp->errbuf, "allocation failed: oprofile");
om->M = M;
om->roff = roff;
if (! fread((char *) &n, sizeof(int), 1, hfp->ffp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read name length");
ESL_ALLOC(om->name, sizeof(char) * (n+1));
if (! fread((char *) om->name, sizeof(char), n+1, hfp->ffp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read name");
if (! fread((char *) &(om->max_length),sizeof(int), 1, hfp->ffp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read max_length");
if (! fread((char *) &(om->tbm_b), sizeof(uint8_t), 1, hfp->ffp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read tbm");
if (! fread((char *) &(om->tec_b), sizeof(uint8_t), 1, hfp->ffp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read tec");
if (! fread((char *) &(om->tjb_b), sizeof(uint8_t), 1, hfp->ffp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read tjb");
if (! fread((char *) &(om->scale_b), sizeof(float), 1, hfp->ffp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read scale");
if (! fread((char *) &(om->base_b), sizeof(uint8_t), 1, hfp->ffp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read base");
if (! fread((char *) &(om->bias_b), sizeof(uint8_t), 1, hfp->ffp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read bias");
for (x = 0; x < abc->Kp; x++)
if (! fread((char *) om->sbv[x], sizeof(__m128i), Q16x, hfp->ffp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read ssv scores at %d [residue %c]", x, abc->sym[x]);
for (x = 0; x < abc->Kp; x++)
if (! fread((char *) om->rbv[x], sizeof(__m128i), Q16, hfp->ffp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read msv scores at %d [residue %c]", x, abc->sym[x]);
if (! fread((char *) om->evparam, sizeof(float), p7_NEVPARAM, hfp->ffp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read stat params");
if (! fread((char *) om->offs, sizeof(off_t), p7_NOFFSETS, hfp->ffp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read hmmpfam offsets");
if (! fread((char *) om->compo, sizeof(float), p7_MAXABET, hfp->ffp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read model composition");
/* record ends with magic sentinel, for detecting binary file corruption */
if (! fread( (char *) &magic, sizeof(uint32_t), 1, hfp->ffp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "no sentinel magic: .h3f file corrupted?");
if (magic != v3f_fmagic) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "bad sentinel magic; .h3f file corrupted?");
/* keep track of the ending offset of the MSV model */
om->eoff = ftello(hfp->ffp) - 1;;
/* MSV models are always multilocal. ReadRest() might override this later; that's ok. */
om->mode = p7_LOCAL;
om->nj = 1.0f;
if (byp_abc != NULL) *byp_abc = abc; /* pass alphabet (whether new or not) back to caller, if caller wanted it */
*ret_om = om;
return eslOK;
ERROR:
if (abc != NULL && (byp_abc == NULL || *byp_abc == NULL)) esl_alphabet_Destroy(abc); /* destroy alphabet if we created it here */
if (om != NULL) p7_oprofile_Destroy(om);
*ret_om = NULL;
return status;
}
/* Function: p7_oprofile_ReadInfoMSV()
* Synopsis: Read MSV filter info, but not the scores.
*
* Purpose: Read just enough of the MSV filter header from the
* <.h3f> file associated with an open HMM file <hfp>
* to skip ahead to the next MSV filter. Allocate a new
* model, populate it with just the file offsets of this
* model and return a pointer to it in <*ret_om>.
*
* The <.h3f> file was opened automatically, if it existed,
* when the HMM file was opened with <p7_hmmfile_OpenE()>.
*
* When no more HMMs remain in the file, return <eslEOF>.
*
* Args: hfp - open HMM file, with associated .h3p file
* byp_abc - BYPASS: <*byp_abc == ESL_ALPHABET *> if known;
* <*byp_abc == NULL> if desired;
* <NULL> if unwanted.
* ret_om - RETURN: newly allocated <om> with partial MSV
* filter data filled in.
*
* Returns: <eslOK> on success. <*ret_om> is allocated here;
* caller free's with <p7_oprofile_Destroy()>.
* <*byp_abc> is allocated here if it was requested;
* caller free's with <esl_alphabet_Destroy()>.
*
* Returns <eslEFORMAT> if <hfp> has no <.h3f> file open,
* or on any parsing error.
*
* Returns <eslEINCOMPAT> if the HMM we read is incompatible
* with the existing alphabet <*byp_abc> led us to expect.
*
* On any returned error, <hfp->errbuf> contains an
* informative error message.
*
* Throws: <eslEMEM> on allocation error.
*/
int
p7_oprofile_ReadInfoMSV(P7_HMMFILE *hfp, ESL_ALPHABET **byp_abc, P7_OPROFILE **ret_om)
{
P7_OPROFILE *om = NULL;
ESL_ALPHABET *abc = NULL;
uint32_t magic;
off_t roff;
int M, Q16, Q16x;
int n;
int alphatype;
int status;
hfp->errbuf[0] = '\0';
if (hfp->ffp == NULL) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "no MSV profile file; hmmpress probably wasn't run");
if (feof(hfp->ffp)) { status = eslEOF; goto ERROR; } /* normal EOF: no more profiles */
/* keep track of the starting offset of the MSV model */
roff = ftello(hfp->ffp);
if (! fread( (char *) &magic, sizeof(uint32_t), 1, hfp->ffp)) { status = eslEOF; goto ERROR; }
if (magic == v3a_fmagic) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "binary auxfiles are in an outdated HMMER format (3/a); please hmmpress your HMM file again");
if (magic == v3b_fmagic) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "binary auxfiles are in an outdated HMMER format (3/b); please hmmpress your HMM file again");
if (magic == v3c_fmagic) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "binary auxfiles are in an outdated HMMER format (3/c); please hmmpress your HMM file again");
if (magic == v3d_fmagic) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "binary auxfiles are in an outdated HMMER format (3/d); please hmmpress your HMM file again");
if (magic == v3e_fmagic) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "binary auxfiles are in an outdated HMMER format (3/e); please hmmpress your HMM file again");
if (magic != v3f_fmagic) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "bad magic; not an HMM database?");
if (! fread( (char *) &M, sizeof(int), 1, hfp->ffp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read model size M");
if (! fread( (char *) &alphatype, sizeof(int), 1, hfp->ffp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read alphabet type");
Q16 = P7_NVB(M);
Q16x = P7_NVB(M) + p7O_EXTRA_SB;
/* Set or verify alphabet. */
if (byp_abc == NULL || *byp_abc == NULL) { /* alphabet unknown: whether wanted or unwanted, make a new one */
if ((abc = esl_alphabet_Create(alphatype)) == NULL) ESL_XFAIL(eslEMEM, hfp->errbuf, "allocation failed: alphabet");
} else { /* alphabet already known: verify it against what we see in the HMM */
abc = *byp_abc;
if (abc->type != alphatype)
ESL_XFAIL(eslEINCOMPAT, hfp->errbuf, "Alphabet type mismatch: was %s, but current profile says %s",
esl_abc_DecodeType(abc->type), esl_abc_DecodeType(alphatype));
}
/* Now we know the sizes of things, so we can allocate. */
if ((om = p7_oprofile_Create(M, abc)) == NULL) ESL_XFAIL(eslEMEM, hfp->errbuf, "allocation failed: oprofile");
om->M = M;
om->roff = roff;
/* calculate the remaining length of the msv model */
om->name = NULL;
if (!fread((char *) &n, sizeof(int), 1, hfp->ffp)) ESL_XFAIL(eslEFORMAT, hfp->errbuf, "failed to read name length");
roff += (sizeof(int) * 5); /* magic, model size, alphabet type, max length, name length */
roff += (sizeof(char) * (n + 1)); /* name string and terminator '\0' */
roff += (sizeof(float) + sizeof(uint8_t) * 5); /* transition costs, bias, scale and base */
roff += (sizeof(__m128i) * abc->Kp * Q16x); /* ssv scores */
roff += (sizeof(__m128i) * abc->Kp * Q16); /* msv scores */
roff += (sizeof(float) * p7_NEVPARAM); /* stat params */
roff += (sizeof(off_t) * p7_NOFFSETS); /* hmmscan offsets */
roff += (sizeof(float) * p7_MAXABET); /* model composition */
roff += sizeof(uint32_t); /* sentinel magic */
/* keep track of the ending offset of the MSV model */
p7_oprofile_Position(hfp, roff);
om->eoff = ftello(hfp->ffp) - 1;
/* MSV models are always multilocal. ReadRest() might override this later; that's ok. */
om->mode = p7_LOCAL;
om->nj = 1.0f;
if (byp_abc != NULL) *byp_abc = abc; /* pass alphabet (whether new or not) back to caller, if caller wanted it */
*ret_om = om;
return eslOK;
ERROR:
if (abc != NULL && (byp_abc == NULL || *byp_abc == NULL)) esl_alphabet_Destroy(abc); /* destroy alphabet if we created it here */
if (om != NULL) p7_oprofile_Destroy(om);
*ret_om = NULL;
return status;
}
/* Function: p7_SSVFilter_longtarget()
* Synopsis: Finds windows with SSV scores above some threshold (vewy vewy fast, in limited precision)
*
* Purpose: Calculates an approximation of the SSV (single ungapped diagonal)
* score for regions of sequence <dsq> of length <L> residues, using
* optimized profile <om>, and a preallocated 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,
* which depends on the <bg> and <p> (usually p=0.02 for nhmmer).
* Note that this variant performs only SSV computations, never
* passing through the J state - the score required to pass SSV at
* the default threshold (or less restrictive) is sufficient to
* pass MSV in essentially all DNA models we've tested.
*
* Above-threshold diagonals are captured into a preallocated list
* <windowlist>. Rather than simply capturing positions at which a
* score threshold is reached, this function establishes windows
* around those high-scoring positions, using scores in <msvdata>.
* These windows can be merged by the calling function.
*
*
* Args: dsq - digital target sequence, 1..L
* L - length of dsq in residues
* om - optimized profile
* ox - DP matrix
* msvdata - compact representation of substitution scores, for backtracking diagonals
* bg - the background model, 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 - preallocated container for all hits (resized if necessary)
*
*
* Note: We misuse the matrix <ox> here, using only a third of the
* first dp row, accessing it as <dp[0..Q-1]> rather than
* in triplets via <{MDI}MX(q)> macros, since we only need
* to store M state values. We know that if <ox> was big
* enough for normal DP calculations, it must be big enough
* to hold the MSVFilter calculation.
*
* Returns: <eslOK> on success.
*
* Throws: <eslEINVAL> if <ox> allocation is too small.
*/
int
p7_SSVFilter_longtarget_avx(const ESL_DSQ *dsq, int L, P7_OPROFILE *om, P7_FILTERMX *ox, const P7_SCOREDATA *msvdata,
P7_BG *bg, double P, P7_HMM_WINDOWLIST *windowlist)
{
#ifdef HAVE_AVX2
register __m128i mpv; /* previous row values */
register __m128i xEv; /* E state: keeps max for Mk->E for a single iteration */
register __m128i xBv; /* B state: splatted vector of B[i-1] for B->Mk calculations */
register __m128i sv; /* temp storage of 1 curr row value in progress */
register __m128i biasv; /* emission bias in a vector */
int i; /* counter over sequence positions 1..L */
int q; /* counter over vectors 0..nq-1 */
int Q = P7_NVB(om->M); /* segment length: # of vectors */
__m128i *dp = ox->dp; /* one DP row of memory */
__m128i *rsc; /* will point at om->rbv[x] for residue x[i] */
__m128i tjbmv; /* vector for J->B move cost + B->M move costs */
__m128i basev; /* offset for scores */
__m128i ceilingv; /* saturated simd value used to test for overflow */
__m128i tempv; /* work vector */
int cmp;
int k;
int n;
int end;
int rem_sc;
int start;
int target_end;
int target_start;
int max_end;
int max_sc;
int sc;
int pos_since_max;
float ret_sc;
union { __m128i v; uint8_t b[16]; } u;
int status;
/*
* Computing the score required to let P meet the F1 prob threshold
* In original code, converting from a scaled int MSV
* score S (the score getting to state E) to a probability goes like this:
* usc = S - om->tec_b - om->tjb_b - om->base_b;
* usc /= om->scale_b;
* usc -= 3.0;
* P = f ( (usc - nullsc) / eslCONST_LOG2 , mu, lambda)
* and we're computing the threshold usc, so reverse it:
* (usc - nullsc) / eslCONST_LOG2 = inv_f( P, mu, lambda)
* usc = nullsc + eslCONST_LOG2 * inv_f( P, mu, lambda)
* usc += 3
* usc *= om->scale_b
* S = usc + om->tec_b + om->tjb_b + om->base_b
*
* Here, I compute threshold with length model based on max_length. Doesn't
* matter much - in any case, both the bg and om models will change with roughly
* 1 bit for each doubling of the length model, so they offset.
*/
float nullsc;
__m128i sc_threshv;
uint8_t sc_thresh;
float invP = esl_gumbel_invsurv(P, om->evparam[p7_MMU], om->evparam[p7_MLAMBDA]);
p7_bg_SetLength(bg, om->max_length);
p7_oprofile_ReconfigMSVLength(om, om->max_length);
p7_bg_NullOne (bg, dsq, om->max_length, &nullsc);
sc_thresh = (int) ceil( ( ( nullsc + (invP * eslCONST_LOG2) + 3.0 ) * om->scale_b ) + om->base_b + om->tec_b + om->tjb_b );
sc_threshv = _mm_set1_epi8((int8_t) 255 - sc_thresh);
/* Resize the filter mx as needed */
if (( status = p7_filtermx_GrowTo(ox, om->M)) != eslOK) ESL_EXCEPTION(status, "Reallocation of SSV filter matrix failed");
/* Matrix type and size must be set early, not late: debugging dump functions need this information. */
ox->M = om->M;
ox->type = p7F_SSVFILTER;
/* Initialization. In offset unsigned arithmetic, -infinity is 0, and 0 is om->base.
*/
biasv = _mm_set1_epi8((int8_t) om->bias_b); /* yes, you can set1() an unsigned char vector this way */
ceilingv = _mm_cmpeq_epi8(biasv, biasv);
for (q = 0; q < Q; q++) dp[q] = _mm_setzero_si128();
basev = _mm_set1_epi8((int8_t) om->base_b);
tjbmv = _mm_set1_epi8((int8_t) om->tjb_b + (int8_t) om->tbm_b);
xBv = _mm_subs_epu8(basev, tjbmv);
for (i = 1; i <= L; i++) {
rsc = om->rbv[dsq[i]];
xEv = _mm_setzero_si128();
/* Right shifts by 1 byte. 4,8,12,x becomes x,4,8,12.
* Because ia32 is littlendian, this means a left bit shift.
* Zeros shift on automatically, which is our -infinity.
*/
mpv = _mm_slli_si128(dp[Q-1], 1);
for (q = 0; q < Q; q++) {
/* Calculate new MMXo(i,q); don't store it yet, hold it in sv. */
sv = _mm_max_epu8(mpv, xBv);
sv = _mm_adds_epu8(sv, biasv);
sv = _mm_subs_epu8(sv, *rsc); rsc++;
xEv = _mm_max_epu8(xEv, sv);
mpv = dp[q]; /* Load {MDI}(i-1,q) into mpv */
dp[q] = sv; /* Do delayed store of M(i,q) now that memory is usable */
}
/* test if the pthresh significance threshold has been reached;
* note: don't use _mm_cmpgt_epi8, because it's a signed comparison, which won't work on uint8s */
tempv = _mm_adds_epu8(xEv, sc_threshv);
tempv = _mm_cmpeq_epi8(tempv, ceilingv);
cmp = _mm_movemask_epi8(tempv);
if (cmp != 0) { //hit pthresh, so add position to list and reset values
//figure out which model state hit threshold
end = -1;
rem_sc = -1;
for (q = 0; q < Q; q++) { /// Unpack and unstripe, so we can find the state that exceeded pthresh
u.v = dp[q];
for (k = 0; k < 16; k++) { // unstripe
//(q+Q*k+1) is the model position k at which the xE score is found
if (u.b[k] >= sc_thresh && u.b[k] > rem_sc && (q+Q*k+1) <= om->M) {
end = (q+Q*k+1);
rem_sc = u.b[k];
}
}
dp[q] = _mm_set1_epi8(0); // while we're here ... this will cause values to get reset to xB in next dp iteration
}
//recover the diagonal that hit threshold
start = end;
target_end = target_start = i;
sc = rem_sc;
while (rem_sc > om->base_b - om->tjb_b - om->tbm_b) {
rem_sc -= om->bias_b - msvdata->msv_scores[start*om->abc->Kp + dsq[target_start]];
--start;
--target_start;
}
start++;
target_start++;
//extend diagonal further with single diagonal extension
k = end+1;
n = target_end+1;
max_end = target_end;
max_sc = sc;
pos_since_max = 0;
while (k<om->M && n<=L) {
sc += om->bias_b - msvdata->msv_scores[k*om->abc->Kp + dsq[n]];
if (sc >= max_sc) {
max_sc = sc;
max_end = n;
pos_since_max=0;
} else {
pos_since_max++;
if (pos_since_max == 5)
break;
}
k++;
n++;
}
end += (max_end - target_end);
k += (max_end - target_end);
target_end = max_end;
ret_sc = ((float) (max_sc - om->tjb_b) - (float) om->base_b);
ret_sc /= om->scale_b;
ret_sc -= 3.0; // that's ~ L \log \frac{L}{L+3}, for our NN,CC,JJ
p7_hmmwindow_new(windowlist, 0, target_start, k, end, end-start+1 , ret_sc, p7_NOCOMPLEMENT);
i = target_end; // skip forward
}
} /* end loop over sequence residues 1..L */
return eslOK;
#endif /* HAVE_AVX2 */
#ifndef HAVE_AVX2
return 0;
#endif
}
