/* SSE optimized implementation for Infernal */ #ifndef CM_IMPL_SSE_INCLUDED #define CM_IMPL_SSE_INCLUDED #include #include #include "esl_config.h" #include "p7_config.h" #include "config.h" #include "easel.h" #include "esl_hmm.h" #include "hmmer.h" #include "infernal.h" #define BYTEMAX 255 #define WORDMAX 0x7fff /***************************************************************** * 1. CM_OPTIMIZED: an optimized score profile *****************************************************************/ typedef struct cm_optimized_s { /* basic model information */ int M; char *sttype; int *cfirst; int *cnum; int *plast; int *pnum; const ESL_ALPHABET *abc; /* epi16-specific information */ float scale_w; int16_t **tsc; int16_t **oesc; int16_t *beginsc; int16_t *endsc; int16_t el_selfsc; } CM_OPTIMIZED; /***************************************************************** * 2. CCM_BG: a background model (for bias composition correction) *****************************************************************/ typedef struct ccm_bg_s { float p1; /* null model's self-loop probability */ float *f; /* residue frequencies [0..K-1] */ ESL_HMM *fhmm; /* 2-state HMM filter null model in prefilters */ } CCM_BG; /***************************************************************** * 3. CM_CONSENSUS: a match-only profile *****************************************************************/ typedef struct cm_consensus_s { int M; char *sttype; int *next; /* Overloaded: we'll usually fill this in, even though next[i] = i+1 for most cases. For B_st, next will be the _right_ child (and the left will always be i+1 */ ESL_ALPHABET *abc; /* floating point values */ float e_fraglen; /* expected length (bases) of model fragments */ float p_rfrag; /* probability of a recursive fragment (i.e., proportion of frags that are not E-terminal */ float sc_frag; /* log prob for each fragment - LOG(1/#frags) */ /* Reduced-precision uchar scores */ uint8_t tsb_S_Sa, tsb_S_SM, tsb_S_e, tsb_M_S; float scale_b; uint8_t base_b; uint8_t bias_b; uint8_t **oesc; /* Matched-composition null model */ float *mcompo; /* Residue composition for matches */ float *fcompo; /* Total residue composition, including unaligned positions */ CCM_BG *bg; } CM_CONSENSUS; /***************************************************************** * 4. GammaHitMx_epu8: hit collection for 8-bit scanning models *****************************************************************/ /* Structure GammaHitMx_epu8: gamma semi-HMM used for optimal hit resolution * of a CM_CONSENSUS scan. */ typedef struct gammahitmx_epu8_s { int L; /* length of sequence */ int *mx; /* [0..L] SHMM DP matrix for optimum nonoverlap resolution */ int *gback; /* [0..L] traceback pointers for SHMM */ float *savesc; /* [0..L] saves score of hit added to best parse at j */ float cutoff; /* minimum score to report */ int i0; /* position of first residue in sequence (gamma->mx[0] corresponds to this residue) */ int iamgreedy; /* TRUE to use RSEARCH's greedy overlap resolution alg, FALSE to use optimal alg */ } GammaHitMx_epu8; /***************************************************************** * 5. Structure for returning hit coordinates for 16-bit models *****************************************************************/ typedef struct hitcoord_epi16_s { int16_t score; int16_t v; int j; int16_t d; } HitCoord_epi16; /***************************************************************** * 5. Declarations of the external API. *****************************************************************/ /* cm_optimized.c */ uint8_t biased_byteify(CM_CONSENSUS *ccm, float sc); uint8_t unbiased_byteify(CM_CONSENSUS *ccm, float sc); /* wordify() * Converts log probability score to a rounded signed 16-bit integer cost. * Both emissions and transitions for ViterbiFilter get this treatment. * No bias term needed, because we use signed words. * e.g. a score of +3.2, with scale 500.0, becomes +1600. */ static inline int16_t wordify(float scale_w, float sc) { sc = roundf(scale_w * sc); if (sc >= 32767.0) return 32767; else if (sc <= -32768.0) return -32768; else return (int16_t) sc; } CM_OPTIMIZED* cm_optimized_Convert(const CM_t *cm); void cm_optimized_Free(CM_OPTIMIZED *ocm); CM_CONSENSUS* cm_consensus_Convert(CM_t *cm); void cm_consensus_Free(CM_CONSENSUS *ccm); float ccm_explen(CM_CONSENSUS *ccm, float t1, float t2, float t3); int ccm_SetCompo(CM_CONSENSUS *ccm, float f_S_Sa, float f_S_SM, float f_S_e); CCM_BG* ccm_bg_CreateUniform(CM_CONSENSUS *ccm); void ccm_bg_Destroy(CCM_BG *bg); int ccm_bg_NullOne(const CCM_BG *bg, int L, float *ret_sc); int ccm_bg_SetFilter(CM_CONSENSUS *ccm, float mL, float sL); int ccm_bg_FilterScore(CCM_BG *bg, ESL_DSQ *dsq, int L, float *ret_sc); /* sse_cmcons_hitmx.c */ GammaHitMx_epu8 *CreateGammaHitMx_epu8(int L, int i0, int be_greedy, int offset_zero, float cutoff, int do_backward); void FreeGammaHitMx_epu8(GammaHitMx_epu8 *gamma); int UpdateGammaHitMx_epu8(CM_CONSENSUS *ccm, char *errbuf, GammaHitMx_epu8 *gamma, int j, __m128i *alpha_row, CM_TOPHITS *hitlist, int W, int sW); void TBackGammaHitMx_epu8 (GammaHitMx_epu8 *gamma, CM_TOPHITS *hitlist, int i0, int j0); /* sse_cmcons_mscyk.c */ int SSE_MSCYK(CM_CONSENSUS *ccm, char *errbuf, int W, ESL_DSQ *dsq, int i0, int j0, uint8_t cutoff, CM_TOPHITS *hitlist, int do_null3, float **ret_vsc, float *ret_sc); CM_TOPHITS * ResolveMSCYK(CM_TOPHITS *initial, int i0, int j0, int W, float cutoff); /* sse_cm_dpsearch.c */ int SSE_CYKScan(CM_t *cm, char *errbuf, CM_SCAN_MX *smx, ESL_DSQ *dsq, int i0, int j0, float cutoff, CM_TOPHITS *hitlist, int do_null3, float **ret_vsc, float *ret_sc); /* sse_cm_dpsmall.c */ float SSE_CYKInside(CM_t *cm, ESL_DSQ *dsq, int L, int r, int i0, int j0, Parsetree_t **ret_tr); float SSE_CYKInsideScore(CM_t *cm, ESL_DSQ *dsq, int L, int r, int i0, int j0); float SSE_CYKDemands(CM_t *cm, int L, int be_quiet); float SSE_CYKDivideAndConquer(CM_t *cm, ESL_DSQ *dsq, int L, int r, int i0, int j0, Parsetree_t **ret_tr); int SSE_CYKFilter_epi16(CM_OPTIMIZED *ocm, ESL_DSQ *dsq, int L, int vroot, int vend, int i0, int j0, int allow_begin, int *ret_b, int *ret_bsc, HitCoord_epi16 *ret_coord); /* sse_util.c */ __m128 alt_rightshift_ps(__m128 a, __m128 b); void vecprint_ps(__m128 a); /* Function: sse_setlw_neginfv() * Date: DLK, Tue May 12 2009 * * Purpose: Returns a vector containing * <{ -32768 a[1] a[2] a[3] a[4] a[5] a[6] a[7] }> * * This is designed as a limited-use * replacement for the call: * _mm_insert_epi16(a, -32768, 0) * which suffers from a compiler * bug in gcc 3.4.x */ static inline __m128i sse_setlw_neginfv(__m128i a) { __m128i mask = _mm_setr_epi16(0x0000,0xffff,0xffff,0xffff,0xffff,0xffff,0xffff,0xffff); __m128i setv = _mm_setr_epi16(-32768,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000); return _mm_or_si128(_mm_and_si128(a,mask),setv); } /* Function: sse_select_si128() * Date: DLK, Tue June 9 2009 * * Purposes: Bitwise vector select for __m128i */ static inline __m128i sse_select_si128(__m128i a, __m128i b, __m128i mask) { b = _mm_and_si128(b, mask); a = _mm_andnot_si128(mask, a); return _mm_or_si128(a, b); } #endif /* CM_IMPL_SSE_INCLUDED */ /***************************************************************** * Infernal - inference of RNA secondary structure alignments * Version 1.1.1; July 2014 * Copyright (C) 2014 Howard Hughes Medical Institute. * Other copyrights also apply. See the COPYRIGHT file for a full list. * * Infernal is distributed under the terms of the GNU General Public License * (GPLv3). See the LICENSE file for details. *****************************************************************/