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"


#if HAVE_AVX2_MM256_INSERT_EPI8
#define _mm256_insert_epi8_rpl _mm256_insert_epi8
#else
static inline __m256i _mm256_insert_epi8_rpl(__m256i a, int8_t i, int imm) {
    __m256i_8_t A;
    A.m = a;
    A.v[imm] = i;
    return A.m;
}
#endif

#if HAVE_AVX2_MM256_EXTRACT_EPI8
#define _mm256_extract_epi8_rpl _mm256_extract_epi8
#else
static inline int8_t _mm256_extract_epi8_rpl(__m256i a, int imm) {
    __m256i_8_t A;
    A.m = a;
    return A.v[imm];
}
#endif

#define _mm256_cmplt_epi8_rpl(a,b) _mm256_cmpgt_epi8(b,a)

#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_scan_avx2_256_8
#define PNAME parasail_nw_trace_scan_profile_avx2_256_8

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_8(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 pvP = NULL;
    __m256i* restrict pvE = NULL;
    int8_t* restrict boundary = NULL;
    __m256i* restrict pvHt = NULL;
    __m256i* restrict pvH = NULL;
    __m256i* restrict pvGapper = NULL;
    __m256i vGapO;
    __m256i vGapE;
    int8_t NEG_LIMIT = 0;
    int8_t POS_LIMIT = 0;
    __m256i vZero;
    int8_t score = 0;
    __m256i vNegLimit;
    __m256i vPosLimit;
    __m256i vSaturationCheckMin;
    __m256i vSaturationCheckMax;
    __m256i vNegInfFront;
    __m256i vSegLenXgap;
    parasail_result_t *result = NULL;
    __m256i vTIns;
    __m256i vTDel;
    __m256i vTDiag;
    __m256i vTDiagE;
    __m256i vTInsE;
    __m256i vTDiagF;
    __m256i vTDelF;

    /* validate inputs */
    PARASAIL_CHECK_NULL(profile);
    PARASAIL_CHECK_NULL(profile->profile8.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 = 32; /* number of values in vector unit */
    segLen = (s1Len + segWidth - 1) / segWidth;
    offset = (s1Len - 1) % segLen;
    position = (segWidth - 1) - (s1Len - 1) / segLen;
    pvP = (__m256i*)profile->profile8.score;
    vGapO = _mm256_set1_epi8(open);
    vGapE = _mm256_set1_epi8(gap);
    NEG_LIMIT = (-open < matrix->min ? INT8_MIN + open : INT8_MIN - matrix->min) + 1;
    POS_LIMIT = INT8_MAX - matrix->max - 1;
    vZero = _mm256_setzero_si256();
    score = NEG_LIMIT;
    vNegLimit = _mm256_set1_epi8(NEG_LIMIT);
    vPosLimit = _mm256_set1_epi8(POS_LIMIT);
    vSaturationCheckMin = vPosLimit;
    vSaturationCheckMax = vNegLimit;
    vNegInfFront = vZero;
    vTIns  = _mm256_set1_epi8(PARASAIL_INS);
    vTDel  = _mm256_set1_epi8(PARASAIL_DEL);
    vTDiag = _mm256_set1_epi8(PARASAIL_DIAG);
    vTDiagE = _mm256_set1_epi8(PARASAIL_DIAG_E);
    vTInsE = _mm256_set1_epi8(PARASAIL_INS_E);
    vTDiagF = _mm256_set1_epi8(PARASAIL_DIAG_F);
    vTDelF = _mm256_set1_epi8(PARASAIL_DEL_F);
    vNegInfFront = _mm256_insert_epi8_rpl(vNegInfFront, NEG_LIMIT, 0);
    vSegLenXgap = _mm256_adds_epi8(vNegInfFront,
            _mm256_slli_si256_rpl(_mm256_set1_epi8(-segLen*gap), 1));

    /* 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_SCAN
        | PARASAIL_FLAG_TRACE
        | PARASAIL_FLAG_BITS_8 | PARASAIL_FLAG_LANES_32;

    /* initialize heap variables */
    pvE = parasail_memalign___m256i(32, segLen);
    boundary = parasail_memalign_int8_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_8_t h;
            __m256i_8_t e;
            for (segNum=0; segNum<segWidth; ++segNum) {
                int64_t tmp = -open-gap*(segNum*segLen+i);
                h.v[segNum] = tmp < INT8_MIN ? INT8_MIN : tmp;
                tmp = tmp - open;
                e.v[segNum] = tmp < INT8_MIN ? INT8_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 = -open-gap*(i-1);
            boundary[i] = tmp < INT8_MIN ? INT8_MIN : tmp;
        }
    }

    {
        __m256i vGapper = _mm256_subs_epi8(vZero,vGapO);
        for (i=segLen-1; i>=0; --i) {
            _mm256_store_si256(pvGapper+i, vGapper);
            vGapper = _mm256_subs_epi8(vGapper, vGapE);
        }
    }

    /* outer loop over database sequence */
    for (j=0; j<s2Len; ++j) {
        __m256i vE;
        __m256i vE_ext;
        __m256i vE_opn;
        __m256i vHt;
        __m256i vF;
        __m256i vF_ext;
        __m256i vF_opn;
        __m256i vH;
        __m256i vHp;
        __m256i *pvW;
        __m256i vW;
        __m256i case1;
        __m256i case2;
        __m256i vGapper;
        __m256i vT;
        __m256i vET;
        __m256i vFT;

        /* calculate E */
        /* calculate Ht */
        /* calculate F and H first pass */
        vHp = _mm256_load_si256(pvH+(segLen-1));
        vHp = _mm256_slli_si256_rpl(vHp, 1);
        vHp = _mm256_insert_epi8_rpl(vHp, boundary[j], 0);
        pvW = pvP + matrix->mapper[(unsigned char)s2[j]]*segLen;
        vHt = _mm256_subs_epi8(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);
            vGapper = _mm256_load_si256(pvGapper+i);
            vE_opn = _mm256_subs_epi8(vH, vGapO);
            vE_ext = _mm256_subs_epi8(vE, vGapE);
            case1 = _mm256_cmpgt_epi8(vE_opn, vE_ext);
            vET = _mm256_blendv_epi8(vTInsE, vTDiagE, case1);
            arr_store(result->trace->trace_table, vET, i, segLen, j);
            vE = _mm256_max_epi8(vE_opn, vE_ext);
            vSaturationCheckMin = _mm256_min_epi8(vSaturationCheckMin, vE);
            vGapper = _mm256_adds_epi8(vHt, vGapper);
            vF = _mm256_max_epi8(vF, vGapper);
            vHp = _mm256_adds_epi8(vHp, vW);
            vHt = _mm256_max_epi8(vE, vHp);
            _mm256_store_si256(pvE+i, vE);
            _mm256_store_si256(pvHt+i, vHt);
            _mm256_store_si256(pvH+i, vHp);
            vHp = vH;
        }

        /* pseudo prefix scan on F and H */
        vHt = _mm256_slli_si256_rpl(vHt, 1);
        vHt = _mm256_insert_epi8_rpl(vHt, boundary[j+1], 0);
        vGapper = _mm256_load_si256(pvGapper);
        vGapper = _mm256_adds_epi8(vHt, vGapper);
        vF = _mm256_max_epi8(vF, vGapper);
        for (i=0; i<segWidth-2; ++i) {
            __m256i vFt = _mm256_slli_si256_rpl(vF, 1);
            vFt = _mm256_adds_epi8(vFt, vSegLenXgap);
            vF = _mm256_max_epi8(vF, vFt);
        }

        /* calculate final H */
        vF = _mm256_slli_si256_rpl(vF, 1);
        vF = _mm256_adds_epi8(vF, vNegInfFront);
        vH = _mm256_max_epi8(vF, vHt);
        for (i=0; i<segLen; ++i) {
            vET = arr_load(result->trace->trace_table, i, segLen, j);
            vHp = _mm256_load_si256(pvH+i);
            vHt = _mm256_load_si256(pvHt+i);
            vF_opn = _mm256_subs_epi8(vH, vGapO);
            vF_ext = _mm256_subs_epi8(vF, vGapE);
            vF = _mm256_max_epi8(vF_opn, vF_ext);
            case1 = _mm256_cmpgt_epi8(vF_opn, vF_ext);
            vFT = _mm256_blendv_epi8(vTDelF, vTDiagF, case1);
            vH = _mm256_max_epi8(vHt, vF);
            case1 = _mm256_cmpeq_epi8(vH, vHp);
            case2 = _mm256_cmpeq_epi8(vH, vF);
            vT = _mm256_blendv_epi8(
                    _mm256_blendv_epi8(vTIns, vTDel, case2),
                    vTDiag, case1);
            vT = _mm256_or_si256(vT, vET);
            vT = _mm256_or_si256(vT, vFT);
            arr_store(result->trace->trace_table, vT, i, segLen, j);
            _mm256_store_si256(pvH+i, vH);
            vSaturationCheckMin = _mm256_min_epi8(vSaturationCheckMin, vH);
            vSaturationCheckMin = _mm256_min_epi8(vSaturationCheckMin, vF);
            vSaturationCheckMax = _mm256_max_epi8(vSaturationCheckMax, vH);
        }
    }

    /* 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, 1);
        }
        score = (int8_t) _mm256_extract_epi8_rpl (vH, 31);
    }

    if (_mm256_movemask_epi8(_mm256_or_si256(
            _mm256_cmplt_epi8_rpl(vSaturationCheckMin, vNegLimit),
            _mm256_cmpgt_epi8(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;
}