/* Function: p7_MSVFilter()
* Synopsis: Calculates MSV score, vewy vewy fast, in limited precision.
* Incept: SRE, Wed Dec 26 15:12:25 2007 [Janelia]
*
* Purpose: Calculates an approximation of the MSV score for sequence
* <dsq> of length <L> residues, using optimized profile <om>,
* and a preallocated one-row DP matrix <ox>. Return the
* estimated MSV score (in nats) in <ret_sc>.
*
* Score may overflow (and will, on high-scoring
* sequences), but will not underflow.
*
* The model may be in any mode, because only its match
* emission scores will be used. The MSV filter inherently
* assumes a multihit local mode, and uses its own special
* state transition scores, not the scores in the profile.
*
* Args: dsq - digital target sequence, 1..L
* L - length of dsq in residues
* om - optimized profile
* ox - DP matrix
* ret_sc - RETURN: MSV score (in nats)
*
* 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.
* <eslERANGE> if the score overflows the limited range; in
* this case, this is a high-scoring hit.
*
* Throws: <eslEINVAL> if <ox> allocation is too small.
*/
int
p7_MSVFilter(const ESL_DSQ *dsq, int L, const P7_OPROFILE *om, P7_OMX *ox, float *ret_sc)
{
register __m128i mpv; /* previous row values */
register __m128i xEv; /* E state: keeps max for Mk->E as we go */
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 */
uint8_t xJ; /* special states' scores */
int i; /* counter over sequence positions 1..L */
int q; /* counter over vectors 0..nq-1 */
int Q = p7O_NQB(om->M); /* segment length: # of vectors */
__m128i *dp = ox->dpb[0]; /* we're going to use dp[0][0..q..Q-1], not {MDI}MX(q) macros*/
__m128i *rsc; /* will point at om->rbv[x] for residue x[i] */
__m128i xJv; /* vector for states score */
__m128i tjbmv; /* vector for cost of moving from either J or N through B to an M state */
__m128i tecv; /* vector for E->C cost */
__m128i basev; /* offset for scores */
__m128i ceilingv; /* saturateed simd value used to test for overflow */
__m128i tempv; /* work vector */
int cmp;
int status = eslOK;
/* Check that the DP matrix is ok for us. */
if (Q > ox->allocQ16) ESL_EXCEPTION(eslEINVAL, "DP matrix allocated too small");
ox->M = om->M;
/* Try highly optimized ssv filter first */
status = p7_SSVFilter(dsq, L, om, ret_sc);
if (status != eslENORESULT) return status;
/* 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 */
for (q = 0; q < Q; q++) dp[q] = _mm_setzero_si128();
xJ = 0;
/* saturate simd register for overflow test */
ceilingv = _mm_cmpeq_epi8(biasv, biasv);
basev = _mm_set1_epi8((int8_t) om->base_b);
tjbmv = _mm_set1_epi8((int8_t) om->tjb_b + (int8_t) om->tbm_b);
tecv = _mm_set1_epi8((int8_t) om->tec_b);
xJv = _mm_subs_epu8(biasv, biasv);
xBv = _mm_subs_epu8(basev, tjbmv);
#if p7_DEBUGGING
if (ox->debugging)
{
uint8_t xB;
xB = _mm_extract_epi16(xBv, 0);
xJ = _mm_extract_epi16(xJv, 0);
p7_omx_DumpMFRow(ox, 0, 0, 0, xJ, xB, xJ);
}
#endif
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 for the overflow condition */
tempv = _mm_adds_epu8(xEv, biasv);
tempv = _mm_cmpeq_epi8(tempv, ceilingv);
cmp = _mm_movemask_epi8(tempv);
/* Now the "special" states, which start from Mk->E (->C, ->J->B)
* Use shuffles instead of shifts so when the last max has completed,
* the last four elements of the simd register will contain the
* max value. Then the last shuffle will broadcast the max value
* to all simd elements.
*/
tempv = _mm_shuffle_epi32(xEv, _MM_SHUFFLE(2, 3, 0, 1));
xEv = _mm_max_epu8(xEv, tempv);
tempv = _mm_shuffle_epi32(xEv, _MM_SHUFFLE(0, 1, 2, 3));
xEv = _mm_max_epu8(xEv, tempv);
tempv = _mm_shufflelo_epi16(xEv, _MM_SHUFFLE(2, 3, 0, 1));
xEv = _mm_max_epu8(xEv, tempv);
tempv = _mm_srli_si128(xEv, 1);
xEv = _mm_max_epu8(xEv, tempv);
xEv = _mm_shuffle_epi32(xEv, _MM_SHUFFLE(0, 0, 0, 0));
/* immediately detect overflow */
if (cmp != 0x0000)
{
*ret_sc = eslINFINITY;
return eslERANGE;
}
xEv = _mm_subs_epu8(xEv, tecv);
xJv = _mm_max_epu8(xJv,xEv);
xBv = _mm_max_epu8(basev, xJv);
xBv = _mm_subs_epu8(xBv, tjbmv);
#if p7_DEBUGGING
if (ox->debugging)
{
uint8_t xB, xE;
xB = _mm_extract_epi16(xBv, 0);
xE = _mm_extract_epi16(xEv, 0);
xJ = _mm_extract_epi16(xJv, 0);
p7_omx_DumpMFRow(ox, i, xE, 0, xJ, xB, xJ);
}
#endif
} /* end loop over sequence residues 1..L */
xJ = (uint8_t) _mm_extract_epi16(xJv, 0);
/* finally C->T, and add our missing precision on the NN,CC,JJ back */
*ret_sc = ((float) (xJ - 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 */
return eslOK;
}
/* Function: p7_MSVFilter()
* Synopsis: Calculates MSV score, vewy vewy fast, in limited precision.
* Incept: SRE, Wed Dec 26 15:12:25 2007 [Janelia]
*
* Purpose: Calculates an approximation of the MSV score for sequence
* <dsq> of length <L> residues, using optimized profile <om>,
* and a preallocated one-row DP matrix <ox>. Return the
* estimated MSV score (in nats) in <ret_sc>.
*
* Score may overflow (and will, on high-scoring
* sequences), but will not underflow.
*
* The model may be in any mode, because only its match
* emission scores will be used. The MSV filter inherently
* assumes a multihit local mode, and uses its own special
* state transition scores, not the scores in the profile.
*
* Args: dsq - digital target sequence, 1..L
* L - length of dsq in residues
* om - optimized profile
* ox - DP matrix
* ret_sc - RETURN: MSV score (in nats)
*
* 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.
* <eslERANGE> if the score overflows the limited range; in
* this case, this is a high-scoring hit.
*
* Throws: <eslEINVAL> if <ox> allocation is too small.
*/
int
p7_MSVFilter(const ESL_DSQ *dsq, int L, const P7_OPROFILE *om, P7_OMX *ox, float *ret_sc)
{
vector unsigned char mpv; /* previous row values */
vector unsigned char xEv; /* E state: keeps max for Mk->E as we go */
vector unsigned char xBv; /* B state: splatted vector of B[i-1] for B->Mk calculations */
vector unsigned char sv; /* temp storage of 1 curr row value in progress */
vector unsigned char biasv; /* emission bias in a vector */
uint8_t xJ; /* special states' scores */
int i; /* counter over sequence positions 1..L */
int q; /* counter over vectors 0..nq-1 */
int Q = p7O_NQB(om->M); /* segment length: # of vectors */
vector unsigned char *dp; /* we're going to use dp[0][0..q..Q-1], not {MDI}MX(q) macros*/
vector unsigned char *rsc; /* will point at om->rbv[x] for residue x[i] */
vector unsigned char zerov; /* vector of zeros */
vector unsigned char xJv; /* vector for states score */
vector unsigned char tjbmv; /* vector for B->Mk cost */
vector unsigned char tecv; /* vector for E->C cost */
vector unsigned char basev; /* offset for scores */
vector unsigned char ceilingv; /* saturateed simd value used to test for overflow */
vector unsigned char tempv;
/* Check that the DP matrix is ok for us. */
if (Q > ox->allocQ16) ESL_EXCEPTION(eslEINVAL, "DP matrix allocated too small");
ox->M = om->M;
/* Initialization. In offset unsigned arithmetic, -infinity is 0, and 0 is om->base.
*/
dp = ox->dpb[0];
for (q = 0; q < Q; q++) dp[q] = vec_splat_u8(0);
xJ = 0;
biasv = esl_vmx_set_u8(om->bias_b);
zerov = vec_splat_u8(0);
/* saturate simd register for overflow test */
tempv = vec_splat_u8(1);
ceilingv = (vector unsigned char)vec_cmpeq(biasv, biasv);
ceilingv = vec_subs(ceilingv, biasv);
ceilingv = vec_subs(ceilingv, tempv);
basev = esl_vmx_set_u8((int8_t) om->base_b);
tecv = esl_vmx_set_u8((int8_t) om->tec_b);
tjbmv = esl_vmx_set_u8((int8_t) om->tjb_b + (int8_t) om->tbm_b);
xJv = vec_subs(biasv, biasv);
xBv = vec_subs(basev, tjbmv);
#if p7_DEBUGGING
if (ox->debugging)
{
unsigned char xB;
vec_ste(xBv, 0, &xB);
vec_ste(xJv, 0, &xJ);
p7_omx_DumpMFRow(ox, 0, 0, 0, xJ, xB, xJ);
}
#endif
for (i = 1; i <= L; i++)
{
rsc = om->rbv[dsq[i]];
xEv = vec_splat_u8(0);
// xBv = vec_sub(xBv, tbmv);
/* 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 = vec_sld(zerov, dp[Q-1], 15);
for (q = 0; q < Q; q++)
{
/* Calculate new MMXo(i,q); don't store it yet, hold it in sv. */
sv = vec_max(mpv, xBv);
sv = vec_adds(sv, biasv);
sv = vec_subs(sv, *rsc); rsc++;
xEv = vec_max(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 */
}
/* Now the "special" states, which start from Mk->E (->C, ->J->B)
* Use rotates instead of shifts so when the last max has completed,
* all elements of the simd register will contain the max value.
*/
tempv = vec_sld(xEv, xEv, 1);
xEv = vec_max(xEv, tempv);
tempv = vec_sld(xEv, xEv, 2);
xEv = vec_max(xEv, tempv);
tempv = vec_sld(xEv, xEv, 4);
xEv = vec_max(xEv, tempv);
tempv = vec_sld(xEv, xEv, 8);
xEv = vec_max(xEv, tempv);
/* immediately detect overflow */
if (vec_any_gt(xEv, ceilingv))
{
*ret_sc = eslINFINITY;
return eslERANGE;
}
xEv = vec_subs(xEv, tecv);
xJv = vec_max(xJv,xEv);
xBv = vec_max(basev, xJv);
xBv = vec_subs(xBv, tjbmv);
#if p7_DEBUGGING
if (ox->debugging)
{
unsigned char xB, xE;
vec_ste(xBv, 0, &xB);
vec_ste(xEv, 0, &xE);
vec_ste(xJv, 0, &xJ);
p7_omx_DumpMFRow(ox, i, xE, 0, xJ, xB, xJ);
}
#endif
} /* end loop over sequence residues 1..L */
/* finally C->T, and add our missing precision on the NN,CC,JJ back */
vec_ste(xJv, 0, &xJ);
*ret_sc = ((float) (xJ - 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 */
return eslOK;
}
