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 <string.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 int8_t _mm256_hmax_epi8_rpl(__m256i a) {
    a = _mm256_max_epi8(a, _mm256_permute2x128_si256(a, a, _MM_SHUFFLE(0,0,0,0)));
    a = _mm256_max_epi8(a, _mm256_slli_si256(a, 8));
    a = _mm256_max_epi8(a, _mm256_slli_si256(a, 4));
    a = _mm256_max_epi8(a, _mm256_slli_si256(a, 2));
    a = _mm256_max_epi8(a, _mm256_slli_si256(a, 1));
    return _mm256_extract_epi8_rpl(a, 31);
}


#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] = (int8_t)_mm256_extract_epi8_rpl(vH,  0);
    array[1LL*( 1*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH,  1);
    array[1LL*( 2*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH,  2);
    array[1LL*( 3*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH,  3);
    array[1LL*( 4*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH,  4);
    array[1LL*( 5*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH,  5);
    array[1LL*( 6*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH,  6);
    array[1LL*( 7*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH,  7);
    array[1LL*( 8*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH,  8);
    array[1LL*( 9*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH,  9);
    array[1LL*(10*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 10);
    array[1LL*(11*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 11);
    array[1LL*(12*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 12);
    array[1LL*(13*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 13);
    array[1LL*(14*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 14);
    array[1LL*(15*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 15);
    array[1LL*(16*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 16);
    array[1LL*(17*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 17);
    array[1LL*(18*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 18);
    array[1LL*(19*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 19);
    array[1LL*(20*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 20);
    array[1LL*(21*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 21);
    array[1LL*(22*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 22);
    array[1LL*(23*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 23);
    array[1LL*(24*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 24);
    array[1LL*(25*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 25);
    array[1LL*(26*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 26);
    array[1LL*(27*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 27);
    array[1LL*(28*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 28);
    array[1LL*(29*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 29);
    array[1LL*(30*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 30);
    array[1LL*(31*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 31);
}
#endif

#ifdef PARASAIL_ROWCOL
static inline void arr_store_col(
        int *col,
        __m256i vH,
        int32_t t,
        int32_t seglen)
{
    col[ 0*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH,  0);
    col[ 1*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH,  1);
    col[ 2*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH,  2);
    col[ 3*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH,  3);
    col[ 4*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH,  4);
    col[ 5*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH,  5);
    col[ 6*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH,  6);
    col[ 7*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH,  7);
    col[ 8*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH,  8);
    col[ 9*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH,  9);
    col[10*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH, 10);
    col[11*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH, 11);
    col[12*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH, 12);
    col[13*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH, 13);
    col[14*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH, 14);
    col[15*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH, 15);
    col[16*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH, 16);
    col[17*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH, 17);
    col[18*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH, 18);
    col[19*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH, 19);
    col[20*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH, 20);
    col[21*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH, 21);
    col[22*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH, 22);
    col[23*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH, 23);
    col[24*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH, 24);
    col[25*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH, 25);
    col[26*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH, 26);
    col[27*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH, 27);
    col[28*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH, 28);
    col[29*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH, 29);
    col[30*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH, 30);
    col[31*seglen+t] = (int8_t)_mm256_extract_epi8_rpl(vH, 31);
}
#endif

#ifdef PARASAIL_TABLE
#define FNAME parasail_sw_table_scan_avx2_256_8
#define PNAME parasail_sw_table_scan_profile_avx2_256_8
#else
#ifdef PARASAIL_ROWCOL
#define FNAME parasail_sw_rowcol_scan_avx2_256_8
#define PNAME parasail_sw_rowcol_scan_profile_avx2_256_8
#else
#define FNAME parasail_sw_scan_avx2_256_8
#define PNAME parasail_sw_scan_profile_avx2_256_8
#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)
{
    /* 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 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;
#ifdef PARASAIL_ROWCOL
    int32_t offset = 0;
    int32_t position = 0;
#endif
    __m256i* restrict pvP = NULL;
    __m256i* restrict pvE = NULL;
    __m256i* restrict pvHt = NULL;
    __m256i* restrict pvH = NULL;
    __m256i* restrict pvHMax = 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 vMaxH;
    __m256i vMaxHUnit;
    __m256i vNegInfFront;
    __m256i vSegLenXgap;
    parasail_result_t *result = NULL;

    /* 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;
    end_query = 0;
    end_ref = 0;
    s1Len = profile->s1Len;
    matrix = profile->matrix;
    segWidth = 32; /* number of values in vector unit */
    segLen = (s1Len + segWidth - 1) / segWidth;
#ifdef PARASAIL_ROWCOL
    offset = (s1Len - 1) % segLen;
    position = (segWidth - 1) - (s1Len - 1) / segLen;
#endif
    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;
    vMaxH = vNegLimit;
    vMaxHUnit = vNegLimit;
    vNegInfFront = vZero;
    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 */
#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_SW | PARASAIL_FLAG_SCAN
        | PARASAIL_FLAG_BITS_8 | PARASAIL_FLAG_LANES_32;
#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);
    pvHt= parasail_memalign___m256i(32, segLen);
    pvH = parasail_memalign___m256i(32, segLen);
    pvHMax = parasail_memalign___m256i(32, segLen);
    pvGapper = parasail_memalign___m256i(32, segLen);

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

    parasail_memset___m256i(pvH, vZero, segLen);
    parasail_memset___m256i(pvE, vNegLimit, segLen);
    {
        __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 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, 1);
        pvW = pvP + matrix->mapper[(unsigned char)s2[j]]*segLen;
        vHt = vZero;
        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_epi8(
                    _mm256_subs_epi8(vE, vGapE),
                    _mm256_subs_epi8(vH, vGapO));
            vHp = _mm256_adds_epi8(vHp, vW);
            vF = _mm256_max_epi8(vF, _mm256_adds_epi8(vHt, pvGapper[i]));
            vHt = _mm256_max_epi8(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, 1);
        vF = _mm256_max_epi8(vF, _mm256_adds_epi8(vHt, pvGapper[0]));
        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(vHt, vF);
        vH = _mm256_max_epi8(vH, vZero);
        for (i=0; i<segLen; ++i) {
            vHt = _mm256_load_si256(pvHt+i);
            vF = _mm256_max_epi8(
                    _mm256_subs_epi8(vF, vGapE),
                    _mm256_subs_epi8(vH, vGapO));
            vH = _mm256_max_epi8(vHt, vF);
            vH = _mm256_max_epi8(vH, vZero);
            _mm256_store_si256(pvH+i, vH);
            vSaturationCheckMin = _mm256_min_epi8(vSaturationCheckMin, vH);
            vSaturationCheckMax = _mm256_max_epi8(vSaturationCheckMax, vH);
#ifdef PARASAIL_TABLE
            arr_store_si256(result->tables->score_table, vH, i, segLen, j, s2Len);
#endif
            vMaxH = _mm256_max_epi8(vH, vMaxH);
        } 

        {
            __m256i vCompare = _mm256_cmpgt_epi8(vMaxH, vMaxHUnit);
            if (_mm256_movemask_epi8(vCompare)) {
                score = _mm256_hmax_epi8_rpl(vMaxH);
                vMaxHUnit = _mm256_set1_epi8(score);
                end_ref = j;
                (void)memcpy(pvHMax, pvH, sizeof(__m256i)*segLen);
            }
        }

#ifdef PARASAIL_ROWCOL
        /* extract last value from the column */
        {
            int32_t k = 0;
            __m256i vH = _mm256_load_si256(pvH + offset);
            for (k=0; k<position; ++k) {
                vH = _mm256_slli_si256_rpl(vH, 1);
            }
            result->rowcols->score_row[j] = (int8_t) _mm256_extract_epi8_rpl (vH, 31);
        }
#endif
    }

    /* Trace the alignment ending position on read. */
    {
        int8_t *t = (int8_t*)pvHMax;
        int32_t column_len = segLen * segWidth;
        end_query = s1Len;
        for (i = 0; i<column_len; ++i, ++t) {
            if (*t == score) {
                int32_t temp = i / segWidth + i % segWidth * segLen;
                if (temp < end_query) {
                    end_query = temp;
                }
            }
        }
    }

#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

    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(pvHMax);
    parasail_free(pvH);
    parasail_free(pvHt);
    parasail_free(pvE);

    return result;
}