/** * @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" #define SG_TRACE #define SG_SUFFIX _diag_avx2_256_16 #include "sg_helper.h" #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_srli_si256_rpl(a,imm) _mm256_or_si256(_mm256_slli_si256(_mm256_permute2x128_si256(a, a, _MM_SHUFFLE(3,0,0,1)), 16-imm), _mm256_srli_si256(a, imm)) static inline void arr_store_si256( int8_t *array, __m256i vWH, int32_t i, int32_t s1Len, int32_t j, int32_t s2Len) { if (0 <= i+0 && i+0 < s1Len && 0 <= j-0 && j-0 < s2Len) { array[1LL*(i+0)*s2Len + (j-0)] = (int8_t)_mm256_extract_epi16_rpl(vWH, 15); } if (0 <= i+1 && i+1 < s1Len && 0 <= j-1 && j-1 < s2Len) { array[1LL*(i+1)*s2Len + (j-1)] = (int8_t)_mm256_extract_epi16_rpl(vWH, 14); } if (0 <= i+2 && i+2 < s1Len && 0 <= j-2 && j-2 < s2Len) { array[1LL*(i+2)*s2Len + (j-2)] = (int8_t)_mm256_extract_epi16_rpl(vWH, 13); } if (0 <= i+3 && i+3 < s1Len && 0 <= j-3 && j-3 < s2Len) { array[1LL*(i+3)*s2Len + (j-3)] = (int8_t)_mm256_extract_epi16_rpl(vWH, 12); } if (0 <= i+4 && i+4 < s1Len && 0 <= j-4 && j-4 < s2Len) { array[1LL*(i+4)*s2Len + (j-4)] = (int8_t)_mm256_extract_epi16_rpl(vWH, 11); } if (0 <= i+5 && i+5 < s1Len && 0 <= j-5 && j-5 < s2Len) { array[1LL*(i+5)*s2Len + (j-5)] = (int8_t)_mm256_extract_epi16_rpl(vWH, 10); } if (0 <= i+6 && i+6 < s1Len && 0 <= j-6 && j-6 < s2Len) { array[1LL*(i+6)*s2Len + (j-6)] = (int8_t)_mm256_extract_epi16_rpl(vWH, 9); } if (0 <= i+7 && i+7 < s1Len && 0 <= j-7 && j-7 < s2Len) { array[1LL*(i+7)*s2Len + (j-7)] = (int8_t)_mm256_extract_epi16_rpl(vWH, 8); } if (0 <= i+8 && i+8 < s1Len && 0 <= j-8 && j-8 < s2Len) { array[1LL*(i+8)*s2Len + (j-8)] = (int8_t)_mm256_extract_epi16_rpl(vWH, 7); } if (0 <= i+9 && i+9 < s1Len && 0 <= j-9 && j-9 < s2Len) { array[1LL*(i+9)*s2Len + (j-9)] = (int8_t)_mm256_extract_epi16_rpl(vWH, 6); } if (0 <= i+10 && i+10 < s1Len && 0 <= j-10 && j-10 < s2Len) { array[1LL*(i+10)*s2Len + (j-10)] = (int8_t)_mm256_extract_epi16_rpl(vWH, 5); } if (0 <= i+11 && i+11 < s1Len && 0 <= j-11 && j-11 < s2Len) { array[1LL*(i+11)*s2Len + (j-11)] = (int8_t)_mm256_extract_epi16_rpl(vWH, 4); } if (0 <= i+12 && i+12 < s1Len && 0 <= j-12 && j-12 < s2Len) { array[1LL*(i+12)*s2Len + (j-12)] = (int8_t)_mm256_extract_epi16_rpl(vWH, 3); } if (0 <= i+13 && i+13 < s1Len && 0 <= j-13 && j-13 < s2Len) { array[1LL*(i+13)*s2Len + (j-13)] = (int8_t)_mm256_extract_epi16_rpl(vWH, 2); } if (0 <= i+14 && i+14 < s1Len && 0 <= j-14 && j-14 < s2Len) { array[1LL*(i+14)*s2Len + (j-14)] = (int8_t)_mm256_extract_epi16_rpl(vWH, 1); } if (0 <= i+15 && i+15 < s1Len && 0 <= j-15 && j-15 < s2Len) { array[1LL*(i+15)*s2Len + (j-15)] = (int8_t)_mm256_extract_epi16_rpl(vWH, 0); } } #define FNAME parasail_sg_flags_trace_diag_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, int s1_beg, int s1_end, int s2_beg, int s2_end) { /* declare local variables */ int32_t N = 0; int32_t PAD = 0; int32_t PAD2 = 0; int32_t s1Len_PAD = 0; int32_t s2Len_PAD = 0; int16_t * restrict s1 = NULL; int16_t * restrict s2B = NULL; int16_t * restrict _H_pr = NULL; int16_t * restrict _F_pr = NULL; int16_t * restrict s2 = NULL; int16_t * restrict H_pr = NULL; int16_t * restrict F_pr = NULL; parasail_result_t *result = NULL; int32_t i = 0; int32_t j = 0; int32_t s1Len = 0; int32_t end_query = 0; int32_t end_ref = 0; int16_t NEG_LIMIT = 0; int16_t POS_LIMIT = 0; int16_t score = 0; __m256i vNegLimit; __m256i vPosLimit; __m256i vSaturationCheckMin; __m256i vSaturationCheckMax; __m256i vNegInf; __m256i vOpen; __m256i vGap; __m256i vOne; __m256i vN; __m256i vGapN; __m256i vNegOne; __m256i vI; __m256i vJreset; __m256i vMaxHRow; __m256i vMaxHCol; __m256i vLastVal; __m256i vEndI; __m256i vEndJ; __m256i vILimit; __m256i vILimit1; __m256i vJLimit; __m256i vJLimit1; __m256i vIBoundary; __m256i vTDiag; __m256i vTIns; __m256i vTDel; __m256i vTDiagE; __m256i vTInsE; __m256i vTDiagF; __m256i vTDelF; /* 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 stack variables */ N = 16; /* 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 = s1Len-1; end_ref = s2Len-1; 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; vNegInf = _mm256_set1_epi16(NEG_LIMIT); vOpen = _mm256_set1_epi16(open); vGap = _mm256_set1_epi16(gap); vOne = _mm256_set1_epi16(1); vN = _mm256_set1_epi16(N); vGapN = s1_beg ? _mm256_set1_epi16(0) : _mm256_set1_epi16(gap*N); vNegOne = _mm256_set1_epi16(-1); vI = _mm256_set_epi16(0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15); vJreset = _mm256_set_epi16(0,-1,-2,-3,-4,-5,-6,-7,-8,-9,-10,-11,-12,-13,-14,-15); vMaxHRow = vNegInf; vMaxHCol = vNegInf; vLastVal = vNegInf; vEndI = vNegInf; vEndJ = vNegInf; vILimit = _mm256_set1_epi16(s1Len); vILimit1 = _mm256_subs_epi16(vILimit, vOne); vJLimit = _mm256_set1_epi16(s2Len); vJLimit1 = _mm256_subs_epi16(vJLimit, vOne); vIBoundary = s1_beg ? _mm256_set1_epi16(0) : _mm256_set_epi16( -open-0*gap, -open-1*gap, -open-2*gap, -open-3*gap, -open-4*gap, -open-5*gap, -open-6*gap, -open-7*gap, -open-8*gap, -open-9*gap, -open-10*gap, -open-11*gap, -open-12*gap, -open-13*gap, -open-14*gap, -open-15*gap); vTDiag = _mm256_set1_epi16(PARASAIL_DIAG); vTIns = _mm256_set1_epi16(PARASAIL_INS); vTDel = _mm256_set1_epi16(PARASAIL_DEL); 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); /* initialize result */ result = parasail_result_new_trace(s1Len, s2Len, 32, sizeof(int8_t)); if (!result) return NULL; /* set known flags */ result->flag |= PARASAIL_FLAG_SG | PARASAIL_FLAG_DIAG | PARASAIL_FLAG_TRACE | PARASAIL_FLAG_BITS_16 | PARASAIL_FLAG_LANES_16; result->flag |= s1_beg ? PARASAIL_FLAG_SG_S1_BEG : 0; result->flag |= s1_end ? PARASAIL_FLAG_SG_S1_END : 0; result->flag |= s2_beg ? PARASAIL_FLAG_SG_S2_BEG : 0; result->flag |= s2_end ? PARASAIL_FLAG_SG_S2_END : 0; /* initialize heap variables */ s2B= parasail_memalign_int16_t(32, s2Len+PAD2); _H_pr = parasail_memalign_int16_t(32, s2Len+PAD2); _F_pr = parasail_memalign_int16_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_int16_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))]; const int * const restrict matrow4 = &matrix->matrix[matrix->size * ((matrix->type == PARASAIL_MATRIX_TYPE_SQUARE) ? s1[i+4] : ((i+4 >= s1Len) ? s1Len-1 : i+4))]; const int * const restrict matrow5 = &matrix->matrix[matrix->size * ((matrix->type == PARASAIL_MATRIX_TYPE_SQUARE) ? s1[i+5] : ((i+5 >= s1Len) ? s1Len-1 : i+5))]; const int * const restrict matrow6 = &matrix->matrix[matrix->size * ((matrix->type == PARASAIL_MATRIX_TYPE_SQUARE) ? s1[i+6] : ((i+6 >= s1Len) ? s1Len-1 : i+6))]; const int * const restrict matrow7 = &matrix->matrix[matrix->size * ((matrix->type == PARASAIL_MATRIX_TYPE_SQUARE) ? s1[i+7] : ((i+7 >= s1Len) ? s1Len-1 : i+7))]; const int * const restrict matrow8 = &matrix->matrix[matrix->size * ((matrix->type == PARASAIL_MATRIX_TYPE_SQUARE) ? s1[i+8] : ((i+8 >= s1Len) ? s1Len-1 : i+8))]; const int * const restrict matrow9 = &matrix->matrix[matrix->size * ((matrix->type == PARASAIL_MATRIX_TYPE_SQUARE) ? s1[i+9] : ((i+9 >= s1Len) ? s1Len-1 : i+9))]; const int * const restrict matrow10 = &matrix->matrix[matrix->size * ((matrix->type == PARASAIL_MATRIX_TYPE_SQUARE) ? s1[i+10] : ((i+10 >= s1Len) ? s1Len-1 : i+10))]; const int * const restrict matrow11 = &matrix->matrix[matrix->size * ((matrix->type == PARASAIL_MATRIX_TYPE_SQUARE) ? s1[i+11] : ((i+11 >= s1Len) ? s1Len-1 : i+11))]; const int * const restrict matrow12 = &matrix->matrix[matrix->size * ((matrix->type == PARASAIL_MATRIX_TYPE_SQUARE) ? s1[i+12] : ((i+12 >= s1Len) ? s1Len-1 : i+12))]; const int * const restrict matrow13 = &matrix->matrix[matrix->size * ((matrix->type == PARASAIL_MATRIX_TYPE_SQUARE) ? s1[i+13] : ((i+13 >= s1Len) ? s1Len-1 : i+13))]; const int * const restrict matrow14 = &matrix->matrix[matrix->size * ((matrix->type == PARASAIL_MATRIX_TYPE_SQUARE) ? s1[i+14] : ((i+14 >= s1Len) ? s1Len-1 : i+14))]; const int * const restrict matrow15 = &matrix->matrix[matrix->size * ((matrix->type == PARASAIL_MATRIX_TYPE_SQUARE) ? s1[i+15] : ((i+15 >= s1Len) ? s1Len-1 : i+15))]; __m256i vIltLimit = _mm256_cmplt_epi16_rpl(vI, vILimit); __m256i vIeqLimit1 = _mm256_cmpeq_epi16(vI, vILimit1); vNH = _mm256_srli_si256_rpl(vNH, 2); vNH = _mm256_insert_epi16_rpl(vNH, H_pr[-1], 15); vWH = _mm256_srli_si256_rpl(vWH, 2); vWH = _mm256_insert_epi16_rpl(vWH, s1_beg ? 0 : (-open - i*gap), 15); H_pr[-1] = -open - (i+N)*gap; /* iterate over database sequence */ for (j=0; j PAD) { vSaturationCheckMin = _mm256_min_epi16(vSaturationCheckMin, vWH); vSaturationCheckMax = _mm256_max_epi16(vSaturationCheckMax, vWH); } /* trace table */ { __m256i case1 = _mm256_cmpeq_epi16(vWH, vNWH); __m256i case2 = _mm256_cmpeq_epi16(vWH, vF); __m256i vT = _mm256_blendv_epi8( _mm256_blendv_epi8(vTIns, vTDel, case2), vTDiag, case1); __m256i condE = _mm256_cmpgt_epi16(vE_opn, vE_ext); __m256i condF = _mm256_cmpgt_epi16(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-15] = (int16_t)_mm256_extract_epi16_rpl(vWH,0); F_pr[j-15] = (int16_t)_mm256_extract_epi16_rpl(vF,0); /* as minor diagonal vector passes across the i or j limit * boundary, extract the last value of the column or row */ { __m256i vJeqLimit1 = _mm256_cmpeq_epi16(vJ, vJLimit1); __m256i vJgtNegOne = _mm256_cmpgt_epi16(vJ, vNegOne); __m256i vJltLimit = _mm256_cmplt_epi16_rpl(vJ, vJLimit); __m256i cond_j = _mm256_and_si256(vIltLimit, vJeqLimit1); __m256i cond_i = _mm256_and_si256(vIeqLimit1, _mm256_and_si256(vJgtNegOne, vJltLimit)); __m256i cond_max_row = _mm256_cmpgt_epi16(vWH, vMaxHRow); __m256i cond_max_col = _mm256_cmpgt_epi16(vWH, vMaxHCol); __m256i cond_last_val = _mm256_and_si256(vIeqLimit1, vJeqLimit1); __m256i cond_all_row = _mm256_and_si256(cond_max_row, cond_i); __m256i cond_all_col = _mm256_and_si256(cond_max_col, cond_j); vMaxHRow = _mm256_blendv_epi8(vMaxHRow, vWH, cond_all_row); vMaxHCol = _mm256_blendv_epi8(vMaxHCol, vWH, cond_all_col); vLastVal = _mm256_blendv_epi8(vLastVal, vWH, cond_last_val); vEndI = _mm256_blendv_epi8(vEndI, vI, cond_all_col); vEndJ = _mm256_blendv_epi8(vEndJ, vJ, cond_all_row); } vJ = _mm256_adds_epi16(vJ, vOne); } vI = _mm256_adds_epi16(vI, vN); vIBoundary = _mm256_subs_epi16(vIBoundary, vGapN); } /* alignment ending position */ { int16_t max_row = NEG_LIMIT; int16_t max_col = NEG_LIMIT; int16_t last_val = NEG_LIMIT; int16_t *s = (int16_t*)&vMaxHRow; int16_t *t = (int16_t*)&vMaxHCol; int16_t *u = (int16_t*)&vLastVal; int16_t *i = (int16_t*)&vEndI; int16_t *j = (int16_t*)&vEndJ; int32_t k; for (k=0; k max_col || (*t == max_col && *i < end_query)) { max_col = *t; end_query = *i; } if (*s > max_row) { max_row = *s; end_ref = *j; } if (*u > last_val) { last_val = *u; } } if (s1_end && s2_end) { if (max_col > max_row || (max_col == max_row && end_ref == s2Len-1)) { score = max_col; end_ref = s2Len-1; } else { score = max_row; end_query = s1Len-1; } } else if (s1_end) { score = max_col; end_ref = s2Len-1; } else if (s2_end) { score = max_row; end_query = s1Len-1; } else { score = last_val; end_query = s1Len-1; end_ref = s2Len-1; } } if (_mm256_movemask_epi8(_mm256_or_si256( _mm256_cmplt_epi16_rpl(vSaturationCheckMin, vNegLimit), _mm256_cmpgt_epi16(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; } SG_IMPL_ALL