libparasail-sys 0.2.1

/**
 * @file
 *
 * @author jeffrey.daily@gmail.com
 *
 * Copyright (c) 2015 Battelle Memorial Institute.
 */
#include "config.h"

#include <stdint.h>
#include <stdlib.h>

#include <immintrin.h>

#include "parasail.h"
#include "parasail/memory.h"
#include "parasail/internal_avx.h"

#define SG_SUFFIX _scan_avx2_256_16
#define SG_SUFFIX_PROF _scan_profile_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_slli_si256_rpl(a,imm) _mm256_alignr_epi8(a, _mm256_permute2x128_si256(a, a, _MM_SHUFFLE(0,0,3,0)), 16-imm)


#ifdef PARASAIL_TABLE
static inline void arr_store_si256(
        int *array,
        __m256i vH,
        int32_t t,
        int32_t seglen,
        int32_t d,
        int32_t dlen)
{
    array[1LL*( 0*seglen+t)*dlen + d] = (int16_t)_mm256_extract_epi16_rpl(vH,  0);
    array[1LL*( 1*seglen+t)*dlen + d] = (int16_t)_mm256_extract_epi16_rpl(vH,  1);
    array[1LL*( 2*seglen+t)*dlen + d] = (int16_t)_mm256_extract_epi16_rpl(vH,  2);
    array[1LL*( 3*seglen+t)*dlen + d] = (int16_t)_mm256_extract_epi16_rpl(vH,  3);
    array[1LL*( 4*seglen+t)*dlen + d] = (int16_t)_mm256_extract_epi16_rpl(vH,  4);
    array[1LL*( 5*seglen+t)*dlen + d] = (int16_t)_mm256_extract_epi16_rpl(vH,  5);
    array[1LL*( 6*seglen+t)*dlen + d] = (int16_t)_mm256_extract_epi16_rpl(vH,  6);
    array[1LL*( 7*seglen+t)*dlen + d] = (int16_t)_mm256_extract_epi16_rpl(vH,  7);
    array[1LL*( 8*seglen+t)*dlen + d] = (int16_t)_mm256_extract_epi16_rpl(vH,  8);
    array[1LL*( 9*seglen+t)*dlen + d] = (int16_t)_mm256_extract_epi16_rpl(vH,  9);
    array[1LL*(10*seglen+t)*dlen + d] = (int16_t)_mm256_extract_epi16_rpl(vH, 10);
    array[1LL*(11*seglen+t)*dlen + d] = (int16_t)_mm256_extract_epi16_rpl(vH, 11);
    array[1LL*(12*seglen+t)*dlen + d] = (int16_t)_mm256_extract_epi16_rpl(vH, 12);
    array[1LL*(13*seglen+t)*dlen + d] = (int16_t)_mm256_extract_epi16_rpl(vH, 13);
    array[1LL*(14*seglen+t)*dlen + d] = (int16_t)_mm256_extract_epi16_rpl(vH, 14);
    array[1LL*(15*seglen+t)*dlen + d] = (int16_t)_mm256_extract_epi16_rpl(vH, 15);
}
#endif

#ifdef PARASAIL_ROWCOL
static inline void arr_store_col(
        int *col,
        __m256i vH,
        int32_t t,
        int32_t seglen)
{
    col[ 0*seglen+t] = (int16_t)_mm256_extract_epi16_rpl(vH,  0);
    col[ 1*seglen+t] = (int16_t)_mm256_extract_epi16_rpl(vH,  1);
    col[ 2*seglen+t] = (int16_t)_mm256_extract_epi16_rpl(vH,  2);
    col[ 3*seglen+t] = (int16_t)_mm256_extract_epi16_rpl(vH,  3);
    col[ 4*seglen+t] = (int16_t)_mm256_extract_epi16_rpl(vH,  4);
    col[ 5*seglen+t] = (int16_t)_mm256_extract_epi16_rpl(vH,  5);
    col[ 6*seglen+t] = (int16_t)_mm256_extract_epi16_rpl(vH,  6);
    col[ 7*seglen+t] = (int16_t)_mm256_extract_epi16_rpl(vH,  7);
    col[ 8*seglen+t] = (int16_t)_mm256_extract_epi16_rpl(vH,  8);
    col[ 9*seglen+t] = (int16_t)_mm256_extract_epi16_rpl(vH,  9);
    col[10*seglen+t] = (int16_t)_mm256_extract_epi16_rpl(vH, 10);
    col[11*seglen+t] = (int16_t)_mm256_extract_epi16_rpl(vH, 11);
    col[12*seglen+t] = (int16_t)_mm256_extract_epi16_rpl(vH, 12);
    col[13*seglen+t] = (int16_t)_mm256_extract_epi16_rpl(vH, 13);
    col[14*seglen+t] = (int16_t)_mm256_extract_epi16_rpl(vH, 14);
    col[15*seglen+t] = (int16_t)_mm256_extract_epi16_rpl(vH, 15);
}
#endif

#ifdef PARASAIL_TABLE
#define FNAME parasail_sg_flags_table_scan_avx2_256_16
#define PNAME parasail_sg_flags_table_scan_profile_avx2_256_16
#else
#ifdef PARASAIL_ROWCOL
#define FNAME parasail_sg_flags_rowcol_scan_avx2_256_16
#define PNAME parasail_sg_flags_rowcol_scan_profile_avx2_256_16
#else
#define FNAME parasail_sg_flags_scan_avx2_256_16
#define PNAME parasail_sg_flags_scan_profile_avx2_256_16
#endif
#endif

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 */
    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, s1_beg, s1_end, s2_beg, s2_end);

    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,
        int s1_beg, int s1_end, int s2_beg, int s2_end)
{
    /* declare local variables */
    int32_t i = 0;
    int32_t j = 0;
    int32_t k = 0;
    int32_t end_query = 0;
    int32_t end_ref = 0;
    int32_t s1Len = 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 pvP = NULL;
    __m256i* restrict pvE = NULL;
    int16_t* restrict boundary = NULL;
    __m256i* restrict pvHt = NULL;
    __m256i* restrict pvH = NULL;
    __m256i* restrict pvGapper = NULL;
    __m256i vGapO;
    __m256i vGapE;
    int16_t NEG_LIMIT = 0;
    int16_t POS_LIMIT = 0;
    __m256i vZero;
    int16_t score = 0;
    __m256i vNegLimit;
    __m256i vPosLimit;
    __m256i vSaturationCheckMin;
    __m256i vSaturationCheckMax;
    __m256i vMaxH;
    __m256i vPosMask;
    __m256i vNegInfFront;
    __m256i vSegLenXgap;
    parasail_result_t *result = NULL;

    /* 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;
    pvP = (__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;
    vZero = _mm256_setzero_si256();
    score = NEG_LIMIT;
    vNegLimit = _mm256_set1_epi16(NEG_LIMIT);
    vPosLimit = _mm256_set1_epi16(POS_LIMIT);
    vSaturationCheckMin = vPosLimit;
    vSaturationCheckMax = vNegLimit;
    vMaxH = vNegLimit;
    vPosMask = _mm256_cmpeq_epi16(_mm256_set1_epi16(position),
            _mm256_set_epi16(0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15));
    vNegInfFront = vZero;
    vNegInfFront = _mm256_insert_epi16_rpl(vNegInfFront, NEG_LIMIT, 0);
    vSegLenXgap = _mm256_adds_epi16(vNegInfFront,
            _mm256_slli_si256_rpl(_mm256_set1_epi16(-segLen*gap), 2));

    /* initialize result */
#ifdef PARASAIL_TABLE
    result = parasail_result_new_table1(segLen*segWidth, s2Len);
#else
#ifdef PARASAIL_ROWCOL
    result = parasail_result_new_rowcol1(segLen*segWidth, s2Len);
#else
    result = parasail_result_new();
#endif
#endif
    if (!result) return NULL;

    /* set known flags */
    result->flag |= PARASAIL_FLAG_SG | PARASAIL_FLAG_SCAN
        | 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;
#ifdef PARASAIL_TABLE
    result->flag |= PARASAIL_FLAG_TABLE;
#endif
#ifdef PARASAIL_ROWCOL
    result->flag |= PARASAIL_FLAG_ROWCOL;
#endif

    /* initialize heap variables */
    pvE = parasail_memalign___m256i(32, segLen);
    boundary = parasail_memalign_int16_t(32, s2Len+1);
    pvHt= parasail_memalign___m256i(32, segLen);
    pvH = parasail_memalign___m256i(32, segLen);
    pvGapper = parasail_memalign___m256i(32, segLen);

    /* validate heap variables */
    if (!pvE) return NULL;
    if (!boundary) return NULL;
    if (!pvHt) return NULL;
    if (!pvH) return NULL;
    if (!pvGapper) return NULL;

    /* initialize H and E */
    {
        int32_t index = 0;
        for (i=0; i<segLen; ++i) {
            int32_t segNum = 0;
            __m256i_16_t h;
            __m256i_16_t e;
            for (segNum=0; segNum<segWidth; ++segNum) {
                int64_t tmp = s1_beg ? 0 : (-open-gap*(segNum*segLen+i));
                h.v[segNum] = tmp < INT16_MIN ? INT16_MIN : tmp;
                tmp = tmp - open;
                e.v[segNum] = tmp < INT16_MIN ? INT16_MIN : tmp;
            }
            _mm256_store_si256(&pvH[index], h.m);
            _mm256_store_si256(&pvE[index], e.m);
            ++index;
        }
    }

    /* initialize uppder boundary */
    {
        boundary[0] = 0;
        for (i=1; i<=s2Len; ++i) {
            int64_t tmp = s2_beg ? 0 : (-open-gap*(i-1));
            boundary[i] = tmp < INT16_MIN ? INT16_MIN : tmp;
        }
    }

    {
        __m256i vGapper = _mm256_subs_epi16(vZero,vGapO);
        for (i=segLen-1; i>=0; --i) {
            _mm256_store_si256(pvGapper+i, vGapper);
            vGapper = _mm256_subs_epi16(vGapper, vGapE);
            /* long queries and/or large penalties will break the pseudo prefix scan */
            vSaturationCheckMin = _mm256_min_epi16(vSaturationCheckMin, vGapper);
        }
    }

    /* outer loop over database sequence */
    for (j=0; j<s2Len; ++j) {
        __m256i vE;
        __m256i vHt;
        __m256i vF;
        __m256i vH;
        __m256i vHp;
        __m256i *pvW;
        __m256i vW;

        /* calculate E */
        /* calculate Ht */
        /* calculate F and H first pass */
        vHp = _mm256_load_si256(pvH+(segLen-1));
        vHp = _mm256_slli_si256_rpl(vHp, 2);
        vHp = _mm256_insert_epi16_rpl(vHp, boundary[j], 0);
        pvW = pvP + matrix->mapper[(unsigned char)s2[j]]*segLen;
        vHt = _mm256_subs_epi16(vNegLimit, pvGapper[0]);
        vF = vNegLimit;
        for (i=0; i<segLen; ++i) {
            vH = _mm256_load_si256(pvH+i);
            vE = _mm256_load_si256(pvE+i);
            vW = _mm256_load_si256(pvW+i);
            vE = _mm256_max_epi16(
                    _mm256_subs_epi16(vE, vGapE),
                    _mm256_subs_epi16(vH, vGapO));
            vHp = _mm256_adds_epi16(vHp, vW);
            vF = _mm256_max_epi16(vF, _mm256_adds_epi16(vHt, pvGapper[i]));
            vHt = _mm256_max_epi16(vE, vHp);
            _mm256_store_si256(pvE+i, vE);
            _mm256_store_si256(pvHt+i, vHt);
            vHp = vH;
        }

        /* pseudo prefix scan on F and H */
        vHt = _mm256_slli_si256_rpl(vHt, 2);
        vHt = _mm256_insert_epi16_rpl(vHt, boundary[j+1], 0);
        vF = _mm256_max_epi16(vF, _mm256_adds_epi16(vHt, pvGapper[0]));
        for (i=0; i<segWidth-2; ++i) {
            __m256i vFt = _mm256_slli_si256_rpl(vF, 2);
            vFt = _mm256_adds_epi16(vFt, vSegLenXgap);
            vF = _mm256_max_epi16(vF, vFt);
        }

        /* calculate final H */
        vF = _mm256_slli_si256_rpl(vF, 2);
        vF = _mm256_adds_epi16(vF, vNegInfFront);
        vH = _mm256_max_epi16(vHt, vF);
        for (i=0; i<segLen; ++i) {
            vHt = _mm256_load_si256(pvHt+i);
            vF = _mm256_max_epi16(
                    _mm256_subs_epi16(vF, vGapE),
                    _mm256_subs_epi16(vH, vGapO));
            vH = _mm256_max_epi16(vHt, vF);
            _mm256_store_si256(pvH+i, vH);
            vSaturationCheckMin = _mm256_min_epi16(vSaturationCheckMin, vH);
            vSaturationCheckMin = _mm256_min_epi16(vSaturationCheckMin, vF);
            vSaturationCheckMax = _mm256_max_epi16(vSaturationCheckMax, vH);
#ifdef PARASAIL_TABLE
            arr_store_si256(result->tables->score_table, vH, i, segLen, j, s2Len);
#endif
        } 

        /* extract vector containing last value from column */
        {
            __m256i vCompare;
            vH = _mm256_load_si256(pvH + offset);
            vCompare = _mm256_and_si256(vPosMask, _mm256_cmpgt_epi16(vH, vMaxH));
            vMaxH = _mm256_max_epi16(vH, vMaxH);
            if (_mm256_movemask_epi8(vCompare)) {
                end_ref = j;
            }
#ifdef PARASAIL_ROWCOL
            for (k=0; k<position; ++k) {
                vH = _mm256_slli_si256_rpl(vH, 2);
            }
            result->rowcols->score_row[j] = (int16_t) _mm256_extract_epi16_rpl (vH, 15);
#endif
        }
    }

#ifdef PARASAIL_ROWCOL
    for (i=0; i<segLen; ++i) {
        __m256i vH = _mm256_load_si256(pvH + i);
        arr_store_col(result->rowcols->score_col, vH, i, segLen);
    }
#endif

    /* max last value from all columns */
    if (s2_end)
    {
        for (k=0; k<position; ++k) {
            vMaxH = _mm256_slli_si256_rpl(vMaxH, 2);
        }
        score = (int16_t) _mm256_extract_epi16_rpl(vMaxH, 15);
        end_query = s1Len-1;
    }

    /* max of last column */
    if (s1_end)
    {
        /* Trace the alignment ending position on read. */
        int16_t *t = (int16_t*)pvH;
        int32_t column_len = segLen * segWidth;
        for (i = 0; i<column_len; ++i, ++t) {
            int32_t temp = i / segWidth + i % segWidth * segLen;
            if (temp >= s1Len) continue;
            if (*t > score) {
                score = *t;
                end_query = temp;
                end_ref = s2Len-1;
            }
            else if (*t == score && end_ref == s2Len-1 && temp < end_query) {
                end_query = temp;
            }
        }
    }

    if (!s1_end && !s2_end) {
        /* extract last value from the last column */
        {
            __m256i vH = _mm256_load_si256(pvH + offset);
            for (k=0; k<position; ++k) {
                vH = _mm256_slli_si256_rpl(vH, 2);
            }
            score = (int16_t) _mm256_extract_epi16_rpl (vH, 15);
            end_ref = s2Len - 1;
            end_query = s1Len - 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(pvGapper);
    parasail_free(pvH);
    parasail_free(pvHt);
    parasail_free(boundary);
    parasail_free(pvE);

    return result;
}

SG_IMPL_ALL
SG_IMPL_PROF_ALL