/** * @file * * @author jeffrey.daily@gmail.com * * Copyright (c) 2015 Battelle Memorial Institute. */ #include "config.h" #include #include #include #include "parasail.h" #include "parasail/memory.h" #include "parasail/internal_avx.h" #define SWAP(A,B) { __m256i* tmp = A; A = B; B = tmp; } #define _mm256_cmplt_epi16_rpl(a,b) _mm256_cmpgt_epi16(b,a) #if HAVE_AVX2_MM256_INSERT_EPI16 #define _mm256_insert_epi16_rpl _mm256_insert_epi16 #else static inline __m256i _mm256_insert_epi16_rpl(__m256i a, int16_t i, int imm) { __m256i_16_t A; A.m = a; A.v[imm] = i; return A.m; } #endif #if HAVE_AVX2_MM256_EXTRACT_EPI16 #define _mm256_extract_epi16_rpl _mm256_extract_epi16 #else static inline int16_t _mm256_extract_epi16_rpl(__m256i a, int imm) { __m256i_16_t A; A.m = a; return A.v[imm]; } #endif #define _mm256_slli_si256_rpl(a,imm) _mm256_alignr_epi8(a, _mm256_permute2x128_si256(a, a, _MM_SHUFFLE(0,0,3,0)), 16-imm) static inline void arr_store( __m256i *array, __m256i vH, int32_t t, int32_t seglen, int32_t d) { _mm256_store_si256(array + (1LL*d*seglen+t), vH); } static inline __m256i arr_load( __m256i *array, int32_t t, int32_t seglen, int32_t d) { return _mm256_load_si256(array + (1LL*d*seglen+t)); } #define FNAME parasail_nw_trace_striped_avx2_256_16 #define PNAME parasail_nw_trace_striped_profile_avx2_256_16 parasail_result_t* FNAME( const char * const restrict s1, const int s1Len, const char * const restrict s2, const int s2Len, const int open, const int gap, const parasail_matrix_t *matrix) { /* declare local variables */ parasail_profile_t *profile = NULL; parasail_result_t *result = NULL; /* validate inputs */ PARASAIL_CHECK_NULL(s2); PARASAIL_CHECK_GT0(s2Len); PARASAIL_CHECK_GE0(open); PARASAIL_CHECK_GE0(gap); PARASAIL_CHECK_NULL(matrix); if (matrix->type == PARASAIL_MATRIX_TYPE_SQUARE) { PARASAIL_CHECK_NULL(s1); PARASAIL_CHECK_GT0(s1Len); } /* initialize local variables */ profile = parasail_profile_create_avx_256_16(s1, s1Len, matrix); if (!profile) return NULL; result = PNAME(profile, s2, s2Len, open, gap); parasail_profile_free(profile); return result; } parasail_result_t* PNAME( const parasail_profile_t * const restrict profile, const char * const restrict s2, const int s2Len, const int open, const int gap) { /* declare local variables */ int32_t i = 0; int32_t j = 0; int32_t k = 0; int32_t s1Len = 0; int32_t end_query = 0; int32_t end_ref = 0; const parasail_matrix_t *matrix = NULL; int32_t segWidth = 0; int32_t segLen = 0; int32_t offset = 0; int32_t position = 0; __m256i* restrict vProfile = NULL; __m256i* restrict pvHStore = NULL; __m256i* restrict pvHLoad = NULL; __m256i* restrict pvE = NULL; __m256i* restrict pvEaStore = NULL; __m256i* restrict pvEaLoad = NULL; __m256i* restrict pvHT = NULL; int16_t* restrict boundary = NULL; __m256i vGapO; __m256i vGapE; int16_t NEG_LIMIT = 0; int16_t POS_LIMIT = 0; int16_t score = 0; __m256i vNegLimit; __m256i vPosLimit; __m256i vSaturationCheckMin; __m256i vSaturationCheckMax; parasail_result_t *result = NULL; __m256i vTIns; __m256i vTDel; __m256i vTDiag; __m256i vTDiagE; __m256i vTInsE; __m256i vTDiagF; __m256i vTDelF; __m256i vTMask; __m256i vFTMask; /* validate inputs */ PARASAIL_CHECK_NULL(profile); PARASAIL_CHECK_NULL(profile->profile16.score); PARASAIL_CHECK_NULL(profile->matrix); PARASAIL_CHECK_GT0(profile->s1Len); PARASAIL_CHECK_NULL(s2); PARASAIL_CHECK_GT0(s2Len); PARASAIL_CHECK_GE0(open); PARASAIL_CHECK_GE0(gap); /* initialize stack variables */ i = 0; j = 0; k = 0; s1Len = profile->s1Len; end_query = s1Len-1; end_ref = s2Len-1; matrix = profile->matrix; segWidth = 16; /* number of values in vector unit */ segLen = (s1Len + segWidth - 1) / segWidth; offset = (s1Len - 1) % segLen; position = (segWidth - 1) - (s1Len - 1) / segLen; vProfile = (__m256i*)profile->profile16.score; vGapO = _mm256_set1_epi16(open); vGapE = _mm256_set1_epi16(gap); NEG_LIMIT = (-open < matrix->min ? INT16_MIN + open : INT16_MIN - matrix->min) + 1; POS_LIMIT = INT16_MAX - matrix->max - 1; score = NEG_LIMIT; vNegLimit = _mm256_set1_epi16(NEG_LIMIT); vPosLimit = _mm256_set1_epi16(POS_LIMIT); vSaturationCheckMin = vPosLimit; vSaturationCheckMax = vNegLimit; vTIns = _mm256_set1_epi16(PARASAIL_INS); vTDel = _mm256_set1_epi16(PARASAIL_DEL); vTDiag = _mm256_set1_epi16(PARASAIL_DIAG); vTDiagE = _mm256_set1_epi16(PARASAIL_DIAG_E); vTInsE = _mm256_set1_epi16(PARASAIL_INS_E); vTDiagF = _mm256_set1_epi16(PARASAIL_DIAG_F); vTDelF = _mm256_set1_epi16(PARASAIL_DEL_F); vTMask = _mm256_set1_epi16(PARASAIL_ZERO_MASK); vFTMask = _mm256_set1_epi16(PARASAIL_F_MASK); /* initialize result */ result = parasail_result_new_trace(segLen, s2Len, 32, sizeof(__m256i)); if (!result) return NULL; /* set known flags */ result->flag |= PARASAIL_FLAG_NW | PARASAIL_FLAG_STRIPED | PARASAIL_FLAG_TRACE | PARASAIL_FLAG_BITS_16 | PARASAIL_FLAG_LANES_16; /* initialize heap variables */ pvHStore = parasail_memalign___m256i(32, segLen); pvHLoad = parasail_memalign___m256i(32, segLen); pvE = parasail_memalign___m256i(32, segLen); pvEaStore = parasail_memalign___m256i(32, segLen); pvEaLoad = parasail_memalign___m256i(32, segLen); pvHT = parasail_memalign___m256i(32, segLen); boundary = parasail_memalign_int16_t(32, s2Len+1); /* validate heap variables */ if (!pvHStore) return NULL; if (!pvHLoad) return NULL; if (!pvE) return NULL; if (!pvEaStore) return NULL; if (!pvEaLoad) return NULL; if (!pvHT) return NULL; if (!boundary) return NULL; /* initialize H and E */ { int32_t index = 0; for (i=0; itrace->trace_table, vTDiagE, i, segLen, 0); } /* outer loop over database sequence */ for (j=0; jmapper[(unsigned char)s2[j]] * segLen; /* Swap the 2 H buffers. */ SWAP(pvHLoad, pvHStore) SWAP(pvEaLoad, pvEaStore) /* inner loop to process the query sequence */ for (i=0; itrace->trace_table, i, segLen, j); __m256i case1 = _mm256_cmpeq_epi16(vH, vH_dag); __m256i case2 = _mm256_cmpeq_epi16(vH, vF); __m256i vT = _mm256_blendv_epi8( _mm256_blendv_epi8(vTIns, vTDel, case2), vTDiag, case1); _mm256_store_si256(pvHT + i, vT); vT = _mm256_or_si256(vT, vTAll); arr_store(result->trace->trace_table, vT, i, segLen, j); } vEF_opn = _mm256_subs_epi16(vH, vGapO); /* Update vE value. */ vE_ext = _mm256_subs_epi16(vE, vGapE); vE = _mm256_max_epi16(vEF_opn, vE_ext); _mm256_store_si256(pvE + i, vE); { __m256i vEa = _mm256_load_si256(pvEaLoad + i); __m256i vEa_ext = _mm256_subs_epi16(vEa, vGapE); vEa = _mm256_max_epi16(vEF_opn, vEa_ext); _mm256_store_si256(pvEaStore + i, vEa); if (j+1trace->trace_table, vT, i, segLen, j+1); } } /* Update vF value. */ vF_ext = _mm256_subs_epi16(vF, vGapE); vF = _mm256_max_epi16(vEF_opn, vF_ext); if (i+1trace->trace_table, i+1, segLen, j); __m256i cond = _mm256_cmpgt_epi16(vEF_opn, vF_ext); __m256i vT = _mm256_blendv_epi8(vTDelF, vTDiagF, cond); vT = _mm256_or_si256(vT, vTAll); arr_store(result->trace->trace_table, vT, i+1, segLen, j); } /* Load the next vH. */ vH = _mm256_load_si256(pvHLoad + i); } /* Lazy_F loop: has been revised to disallow adjecent insertion and * then deletion, so don't update E(i, i), learn from SWPS3 */ vFa_ext = vF_ext; vFa = vF; for (k=0; ktrace->trace_table, i, segLen, j); vT = _mm256_load_si256(pvHT + i); vT = _mm256_blendv_epi8(vT, vTDel, cond); _mm256_store_si256(pvHT + i, vT); vTAll = _mm256_and_si256(vTAll, vTMask); vTAll = _mm256_or_si256(vTAll, vT); arr_store(result->trace->trace_table, vTAll, i, segLen, j); } /* Update vF value. */ { __m256i vTAll = arr_load(result->trace->trace_table, i, segLen, j); __m256i cond = _mm256_cmpgt_epi16(vEF_opn, vFa_ext); __m256i vT = _mm256_blendv_epi8(vTDelF, vTDiagF, cond); vTAll = _mm256_and_si256(vTAll, vFTMask); vTAll = _mm256_or_si256(vTAll, vT); arr_store(result->trace->trace_table, vTAll, i, segLen, j); } vEF_opn = _mm256_subs_epi16(vH, vGapO); vF_ext = _mm256_subs_epi16(vF, vGapE); { __m256i vEa = _mm256_load_si256(pvEaLoad + i); __m256i vEa_ext = _mm256_subs_epi16(vEa, vGapE); vEa = _mm256_max_epi16(vEF_opn, vEa_ext); _mm256_store_si256(pvEaStore + i, vEa); if (j+1trace->trace_table, vT, i, segLen, j+1); } } if (! _mm256_movemask_epi8( _mm256_or_si256( _mm256_cmpgt_epi16(vF_ext, vEF_opn), _mm256_cmpeq_epi16(vF_ext, vEF_opn)))) goto end; /*vF = _mm256_max_epi16(vEF_opn, vF_ext);*/ vF = vF_ext; vFa_ext = _mm256_subs_epi16(vFa, vGapE); vFa = _mm256_max_epi16(vEF_opn, vFa_ext); vHp = _mm256_load_si256(pvHLoad + i); } } end: { } } /* extract last value from the last column */ { __m256i vH = _mm256_load_si256(pvHStore + offset); for (k=0; kflag |= PARASAIL_FLAG_SATURATED; score = 0; end_query = 0; end_ref = 0; } result->score = score; result->end_query = end_query; result->end_ref = end_ref; parasail_free(boundary); parasail_free(pvHT); parasail_free(pvEaLoad); parasail_free(pvEaStore); parasail_free(pvE); parasail_free(pvHLoad); parasail_free(pvHStore); return result; }