/** * @file * * @author jeffrey.daily@gmail.com * * Copyright (c) 2015 Battelle Memorial Institute. */ #include "config.h" #include #include #include "parasail.h" #include "parasail/memory.h" #include "parasail/internal_avx.h" #if HAVE_AVX2_MM256_INSERT_EPI64 #define _mm256_insert_epi64_rpl _mm256_insert_epi64 #else static inline __m256i _mm256_insert_epi64_rpl(__m256i a, int64_t i, int imm) { __m256i_64_t A; A.m = a; A.v[imm] = i; return A.m; } #endif #if HAVE_AVX2_MM256_SET1_EPI64X #define _mm256_set1_epi64x_rpl _mm256_set1_epi64x #else static inline __m256i _mm256_set1_epi64x_rpl(int64_t i) { __m256i_64_t A; A.v[0] = i; A.v[1] = i; A.v[2] = i; A.v[3] = i; return A.m; } #endif static inline __m256i _mm256_max_epi64_rpl(__m256i a, __m256i b) { __m256i_64_t A; __m256i_64_t B; A.m = a; B.m = b; A.v[0] = (A.v[0]>B.v[0]) ? A.v[0] : B.v[0]; A.v[1] = (A.v[1]>B.v[1]) ? A.v[1] : B.v[1]; A.v[2] = (A.v[2]>B.v[2]) ? A.v[2] : B.v[2]; A.v[3] = (A.v[3]>B.v[3]) ? A.v[3] : B.v[3]; return A.m; } #if HAVE_AVX2_MM256_SET_EPI64X #define _mm256_set_epi64x_rpl _mm256_set_epi64x #else static inline __m256i _mm256_set_epi64x_rpl(int64_t e3, int64_t e2, int64_t e1, int64_t e0) { __m256i_64_t A; A.v[0] = e0; A.v[1] = e1; A.v[2] = e2; A.v[3] = e3; return A.m; } #endif #if HAVE_AVX2_MM256_EXTRACT_EPI64 #define _mm256_extract_epi64_rpl _mm256_extract_epi64 #else static inline int64_t _mm256_extract_epi64_rpl(__m256i a, int imm) { __m256i_64_t A; A.m = a; return A.v[imm]; } #endif static inline __m256i _mm256_min_epi64_rpl(__m256i a, __m256i b) { __m256i_64_t A; __m256i_64_t B; A.m = a; B.m = b; A.v[0] = (A.v[0]type == PARASAIL_MATRIX_TYPE_SQUARE) { PARASAIL_CHECK_NULL(_s1); PARASAIL_CHECK_GT0(_s1Len); } /* initialize stack variables */ N = 4; /* number of values in vector */ PAD = N-1; PAD2 = PAD*2; s1Len = matrix->type == PARASAIL_MATRIX_TYPE_SQUARE ? _s1Len : matrix->length; s1Len_PAD = s1Len+PAD; s2Len_PAD = s2Len+PAD; i = 0; j = 0; end_query = 0; end_ref = 0; NEG_LIMIT = (-open < matrix->min ? INT64_MIN + open : INT64_MIN - matrix->min) + 1; POS_LIMIT = INT64_MAX - matrix->max - 1; score = NEG_LIMIT; vNegLimit = _mm256_set1_epi64x_rpl(NEG_LIMIT); vPosLimit = _mm256_set1_epi64x_rpl(POS_LIMIT); vSaturationCheckMin = vPosLimit; vSaturationCheckMax = vNegLimit; vNegInf = _mm256_set1_epi64x_rpl(NEG_LIMIT); vNegInf0 = _mm256_srli_si256_rpl(vNegInf, 8); /* shift in a 0 */ vOpen = _mm256_set1_epi64x_rpl(open); vGap = _mm256_set1_epi64x_rpl(gap); vZero = _mm256_set1_epi64x_rpl(0); vOne = _mm256_set1_epi64x_rpl(1); vN = _mm256_set1_epi64x_rpl(N); vNegOne = _mm256_set1_epi64x_rpl(-1); vI = _mm256_set_epi64x_rpl(0,1,2,3); vJreset = _mm256_set_epi64x_rpl(0,-1,-2,-3); vMaxH = vNegInf; vEndI = vNegInf; vEndJ = vNegInf; vILimit = _mm256_set1_epi64x_rpl(s1Len); vJLimit = _mm256_set1_epi64x_rpl(s2Len); vTDiag = _mm256_set1_epi64x_rpl(PARASAIL_DIAG); vTIns = _mm256_set1_epi64x_rpl(PARASAIL_INS); vTDel = _mm256_set1_epi64x_rpl(PARASAIL_DEL); vTZero = _mm256_set1_epi64x_rpl(PARASAIL_ZERO); vTDiagE = _mm256_set1_epi64x_rpl(PARASAIL_DIAG_E); vTInsE = _mm256_set1_epi64x_rpl(PARASAIL_INS_E); vTDiagF = _mm256_set1_epi64x_rpl(PARASAIL_DIAG_F); vTDelF = _mm256_set1_epi64x_rpl(PARASAIL_DEL_F); /* initialize result */ result = parasail_result_new_trace(s1Len, s2Len, 32, sizeof(int8_t)); if (!result) return NULL; /* set known flags */ result->flag |= PARASAIL_FLAG_SW | PARASAIL_FLAG_DIAG | PARASAIL_FLAG_TRACE | PARASAIL_FLAG_BITS_64 | PARASAIL_FLAG_LANES_4; /* initialize heap variables */ s2B= parasail_memalign_int64_t(32, s2Len+PAD2); _H_pr = parasail_memalign_int64_t(32, s2Len+PAD2); _F_pr = parasail_memalign_int64_t(32, s2Len+PAD2); s2 = s2B+PAD; /* will allow later for negative indices */ H_pr = _H_pr+PAD; F_pr = _F_pr+PAD; /* validate heap variables */ if (!s2B) return NULL; if (!_H_pr) return NULL; if (!_F_pr) return NULL; /* convert _s1 from char to int in range 0-23 */ if (matrix->type == PARASAIL_MATRIX_TYPE_SQUARE) { s1 = parasail_memalign_int64_t(32, s1Len+PAD); if (!s1) return NULL; for (i=0; imapper[(unsigned char)_s1[i]]; } /* pad back of s1 with dummy values */ for (i=s1Len; imapper[(unsigned char)_s2[j]]; } /* pad front of s2 with dummy values */ for (j=-PAD; j<0; ++j) { s2[j] = 0; /* point to first matrix row because we don't care */ } /* pad back of s2 with dummy values */ for (j=s2Len; jmatrix[matrix->size * ((matrix->type == PARASAIL_MATRIX_TYPE_SQUARE) ? s1[i+0] : ((i+0 >= s1Len) ? s1Len-1 : i+0))]; const int * const restrict matrow1 = &matrix->matrix[matrix->size * ((matrix->type == PARASAIL_MATRIX_TYPE_SQUARE) ? s1[i+1] : ((i+1 >= s1Len) ? s1Len-1 : i+1))]; const int * const restrict matrow2 = &matrix->matrix[matrix->size * ((matrix->type == PARASAIL_MATRIX_TYPE_SQUARE) ? s1[i+2] : ((i+2 >= s1Len) ? s1Len-1 : i+2))]; const int * const restrict matrow3 = &matrix->matrix[matrix->size * ((matrix->type == PARASAIL_MATRIX_TYPE_SQUARE) ? s1[i+3] : ((i+3 >= s1Len) ? s1Len-1 : i+3))]; __m256i vIltLimit = _mm256_cmplt_epi64_rpl(vI, vILimit); /* iterate over database sequence */ for (j=0; j PAD) { vSaturationCheckMin = _mm256_min_epi64_rpl(vSaturationCheckMin, vWH); vSaturationCheckMax = _mm256_max_epi64_rpl(vSaturationCheckMax, vWH); } /* trace table */ { __m256i cond_zero = _mm256_cmpeq_epi64(vWH, vZero); __m256i case1 = _mm256_cmpeq_epi64(vWH, vNWH); __m256i case2 = _mm256_cmpeq_epi64(vWH, vF); __m256i vT = _mm256_blendv_epi8( _mm256_blendv_epi8(vTIns, vTDel, case2), _mm256_blendv_epi8(vTDiag, vTZero, cond_zero), case1); __m256i condE = _mm256_cmpgt_epi64(vE_opn, vE_ext); __m256i condF = _mm256_cmpgt_epi64(vF_opn, vF_ext); __m256i vET = _mm256_blendv_epi8(vTInsE, vTDiagE, condE); __m256i vFT = _mm256_blendv_epi8(vTDelF, vTDiagF, condF); vT = _mm256_or_si256(vT, vET); vT = _mm256_or_si256(vT, vFT); arr_store_si256(result->trace->trace_table, vT, i, s1Len, j, s2Len); } H_pr[j-3] = (int64_t)_mm256_extract_epi64_rpl(vWH,0); F_pr[j-3] = (int64_t)_mm256_extract_epi64_rpl(vF,0); /* as minor diagonal vector passes across table, extract * max values within the i,j bounds */ { __m256i cond_valid_J = _mm256_and_si256( _mm256_cmpgt_epi64(vJ, vNegOne), _mm256_cmplt_epi64_rpl(vJ, vJLimit)); __m256i cond_valid_IJ = _mm256_and_si256(cond_valid_J, vIltLimit); __m256i cond_eq = _mm256_cmpeq_epi64(vWH, vMaxH); __m256i cond_max = _mm256_cmpgt_epi64(vWH, vMaxH); __m256i cond_all = _mm256_and_si256(cond_max, cond_valid_IJ); __m256i cond_Jlt = _mm256_cmplt_epi64_rpl(vJ, vEndJ); vMaxH = _mm256_blendv_epi8(vMaxH, vWH, cond_all); vEndI = _mm256_blendv_epi8(vEndI, vI, cond_all); vEndJ = _mm256_blendv_epi8(vEndJ, vJ, cond_all); cond_all = _mm256_and_si256(cond_Jlt, cond_eq); cond_all = _mm256_and_si256(cond_all, cond_valid_IJ); vEndI = _mm256_blendv_epi8(vEndI, vI, cond_all); vEndJ = _mm256_blendv_epi8(vEndJ, vJ, cond_all); } vJ = _mm256_add_epi64(vJ, vOne); } vI = _mm256_add_epi64(vI, vN); } /* alignment ending position */ { int64_t *t = (int64_t*)&vMaxH; int64_t *i = (int64_t*)&vEndI; int64_t *j = (int64_t*)&vEndJ; int32_t k; for (k=0; k score) { score = *t; end_query = *i; end_ref = *j; } else if (*t == score) { if (*j < end_ref) { end_query = *i; end_ref = *j; } else if (*j == end_ref && *i < end_query) { end_query = *i; end_ref = *j; } } } } if (_mm256_movemask_epi8(_mm256_or_si256( _mm256_cmplt_epi64_rpl(vSaturationCheckMin, vNegLimit), _mm256_cmpgt_epi64(vSaturationCheckMax, vPosLimit)))) { result->flag |= 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(_F_pr); parasail_free(_H_pr); parasail_free(s2B); if (matrix->type == PARASAIL_MATRIX_TYPE_SQUARE) { parasail_free(s1); } return result; }