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);
}
static inline void arr_store_si256(
int *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_epi8_rpl(vWH, 31);
}
if (0 <= i+1 && i+1 < s1Len && 0 <= j-1 && j-1 < s2Len) {
array[1LL*(i+1)*s2Len + (j-1)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 30);
}
if (0 <= i+2 && i+2 < s1Len && 0 <= j-2 && j-2 < s2Len) {
array[1LL*(i+2)*s2Len + (j-2)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 29);
}
if (0 <= i+3 && i+3 < s1Len && 0 <= j-3 && j-3 < s2Len) {
array[1LL*(i+3)*s2Len + (j-3)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 28);
}
if (0 <= i+4 && i+4 < s1Len && 0 <= j-4 && j-4 < s2Len) {
array[1LL*(i+4)*s2Len + (j-4)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 27);
}
if (0 <= i+5 && i+5 < s1Len && 0 <= j-5 && j-5 < s2Len) {
array[1LL*(i+5)*s2Len + (j-5)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 26);
}
if (0 <= i+6 && i+6 < s1Len && 0 <= j-6 && j-6 < s2Len) {
array[1LL*(i+6)*s2Len + (j-6)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 25);
}
if (0 <= i+7 && i+7 < s1Len && 0 <= j-7 && j-7 < s2Len) {
array[1LL*(i+7)*s2Len + (j-7)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 24);
}
if (0 <= i+8 && i+8 < s1Len && 0 <= j-8 && j-8 < s2Len) {
array[1LL*(i+8)*s2Len + (j-8)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 23);
}
if (0 <= i+9 && i+9 < s1Len && 0 <= j-9 && j-9 < s2Len) {
array[1LL*(i+9)*s2Len + (j-9)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 22);
}
if (0 <= i+10 && i+10 < s1Len && 0 <= j-10 && j-10 < s2Len) {
array[1LL*(i+10)*s2Len + (j-10)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 21);
}
if (0 <= i+11 && i+11 < s1Len && 0 <= j-11 && j-11 < s2Len) {
array[1LL*(i+11)*s2Len + (j-11)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 20);
}
if (0 <= i+12 && i+12 < s1Len && 0 <= j-12 && j-12 < s2Len) {
array[1LL*(i+12)*s2Len + (j-12)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 19);
}
if (0 <= i+13 && i+13 < s1Len && 0 <= j-13 && j-13 < s2Len) {
array[1LL*(i+13)*s2Len + (j-13)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 18);
}
if (0 <= i+14 && i+14 < s1Len && 0 <= j-14 && j-14 < s2Len) {
array[1LL*(i+14)*s2Len + (j-14)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 17);
}
if (0 <= i+15 && i+15 < s1Len && 0 <= j-15 && j-15 < s2Len) {
array[1LL*(i+15)*s2Len + (j-15)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 16);
}
if (0 <= i+16 && i+16 < s1Len && 0 <= j-16 && j-16 < s2Len) {
array[1LL*(i+16)*s2Len + (j-16)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 15);
}
if (0 <= i+17 && i+17 < s1Len && 0 <= j-17 && j-17 < s2Len) {
array[1LL*(i+17)*s2Len + (j-17)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 14);
}
if (0 <= i+18 && i+18 < s1Len && 0 <= j-18 && j-18 < s2Len) {
array[1LL*(i+18)*s2Len + (j-18)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 13);
}
if (0 <= i+19 && i+19 < s1Len && 0 <= j-19 && j-19 < s2Len) {
array[1LL*(i+19)*s2Len + (j-19)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 12);
}
if (0 <= i+20 && i+20 < s1Len && 0 <= j-20 && j-20 < s2Len) {
array[1LL*(i+20)*s2Len + (j-20)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 11);
}
if (0 <= i+21 && i+21 < s1Len && 0 <= j-21 && j-21 < s2Len) {
array[1LL*(i+21)*s2Len + (j-21)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 10);
}
if (0 <= i+22 && i+22 < s1Len && 0 <= j-22 && j-22 < s2Len) {
array[1LL*(i+22)*s2Len + (j-22)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 9);
}
if (0 <= i+23 && i+23 < s1Len && 0 <= j-23 && j-23 < s2Len) {
array[1LL*(i+23)*s2Len + (j-23)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 8);
}
if (0 <= i+24 && i+24 < s1Len && 0 <= j-24 && j-24 < s2Len) {
array[1LL*(i+24)*s2Len + (j-24)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 7);
}
if (0 <= i+25 && i+25 < s1Len && 0 <= j-25 && j-25 < s2Len) {
array[1LL*(i+25)*s2Len + (j-25)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 6);
}
if (0 <= i+26 && i+26 < s1Len && 0 <= j-26 && j-26 < s2Len) {
array[1LL*(i+26)*s2Len + (j-26)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 5);
}
if (0 <= i+27 && i+27 < s1Len && 0 <= j-27 && j-27 < s2Len) {
array[1LL*(i+27)*s2Len + (j-27)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 4);
}
if (0 <= i+28 && i+28 < s1Len && 0 <= j-28 && j-28 < s2Len) {
array[1LL*(i+28)*s2Len + (j-28)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 3);
}
if (0 <= i+29 && i+29 < s1Len && 0 <= j-29 && j-29 < s2Len) {
array[1LL*(i+29)*s2Len + (j-29)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 2);
}
if (0 <= i+30 && i+30 < s1Len && 0 <= j-30 && j-30 < s2Len) {
array[1LL*(i+30)*s2Len + (j-30)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 1);
}
if (0 <= i+31 && i+31 < s1Len && 0 <= j-31 && j-31 < s2Len) {
array[1LL*(i+31)*s2Len + (j-31)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 0);
}
}
static inline void arr_store_rowcol(
int *row,
int *col,
__m256i vWH,
int32_t i,
int32_t s1Len,
int32_t j,
int32_t s2Len)
{
if (i+0 == s1Len-1 && 0 <= j-0 && j-0 < s2Len) {
row[j-0] = (int8_t)_mm256_extract_epi8_rpl(vWH, 31);
}
if (j-0 == s2Len-1 && 0 <= i+0 && i+0 < s1Len) {
col[(i+0)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 31);
}
if (i+1 == s1Len-1 && 0 <= j-1 && j-1 < s2Len) {
row[j-1] = (int8_t)_mm256_extract_epi8_rpl(vWH, 30);
}
if (j-1 == s2Len-1 && 0 <= i+1 && i+1 < s1Len) {
col[(i+1)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 30);
}
if (i+2 == s1Len-1 && 0 <= j-2 && j-2 < s2Len) {
row[j-2] = (int8_t)_mm256_extract_epi8_rpl(vWH, 29);
}
if (j-2 == s2Len-1 && 0 <= i+2 && i+2 < s1Len) {
col[(i+2)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 29);
}
if (i+3 == s1Len-1 && 0 <= j-3 && j-3 < s2Len) {
row[j-3] = (int8_t)_mm256_extract_epi8_rpl(vWH, 28);
}
if (j-3 == s2Len-1 && 0 <= i+3 && i+3 < s1Len) {
col[(i+3)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 28);
}
if (i+4 == s1Len-1 && 0 <= j-4 && j-4 < s2Len) {
row[j-4] = (int8_t)_mm256_extract_epi8_rpl(vWH, 27);
}
if (j-4 == s2Len-1 && 0 <= i+4 && i+4 < s1Len) {
col[(i+4)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 27);
}
if (i+5 == s1Len-1 && 0 <= j-5 && j-5 < s2Len) {
row[j-5] = (int8_t)_mm256_extract_epi8_rpl(vWH, 26);
}
if (j-5 == s2Len-1 && 0 <= i+5 && i+5 < s1Len) {
col[(i+5)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 26);
}
if (i+6 == s1Len-1 && 0 <= j-6 && j-6 < s2Len) {
row[j-6] = (int8_t)_mm256_extract_epi8_rpl(vWH, 25);
}
if (j-6 == s2Len-1 && 0 <= i+6 && i+6 < s1Len) {
col[(i+6)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 25);
}
if (i+7 == s1Len-1 && 0 <= j-7 && j-7 < s2Len) {
row[j-7] = (int8_t)_mm256_extract_epi8_rpl(vWH, 24);
}
if (j-7 == s2Len-1 && 0 <= i+7 && i+7 < s1Len) {
col[(i+7)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 24);
}
if (i+8 == s1Len-1 && 0 <= j-8 && j-8 < s2Len) {
row[j-8] = (int8_t)_mm256_extract_epi8_rpl(vWH, 23);
}
if (j-8 == s2Len-1 && 0 <= i+8 && i+8 < s1Len) {
col[(i+8)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 23);
}
if (i+9 == s1Len-1 && 0 <= j-9 && j-9 < s2Len) {
row[j-9] = (int8_t)_mm256_extract_epi8_rpl(vWH, 22);
}
if (j-9 == s2Len-1 && 0 <= i+9 && i+9 < s1Len) {
col[(i+9)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 22);
}
if (i+10 == s1Len-1 && 0 <= j-10 && j-10 < s2Len) {
row[j-10] = (int8_t)_mm256_extract_epi8_rpl(vWH, 21);
}
if (j-10 == s2Len-1 && 0 <= i+10 && i+10 < s1Len) {
col[(i+10)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 21);
}
if (i+11 == s1Len-1 && 0 <= j-11 && j-11 < s2Len) {
row[j-11] = (int8_t)_mm256_extract_epi8_rpl(vWH, 20);
}
if (j-11 == s2Len-1 && 0 <= i+11 && i+11 < s1Len) {
col[(i+11)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 20);
}
if (i+12 == s1Len-1 && 0 <= j-12 && j-12 < s2Len) {
row[j-12] = (int8_t)_mm256_extract_epi8_rpl(vWH, 19);
}
if (j-12 == s2Len-1 && 0 <= i+12 && i+12 < s1Len) {
col[(i+12)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 19);
}
if (i+13 == s1Len-1 && 0 <= j-13 && j-13 < s2Len) {
row[j-13] = (int8_t)_mm256_extract_epi8_rpl(vWH, 18);
}
if (j-13 == s2Len-1 && 0 <= i+13 && i+13 < s1Len) {
col[(i+13)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 18);
}
if (i+14 == s1Len-1 && 0 <= j-14 && j-14 < s2Len) {
row[j-14] = (int8_t)_mm256_extract_epi8_rpl(vWH, 17);
}
if (j-14 == s2Len-1 && 0 <= i+14 && i+14 < s1Len) {
col[(i+14)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 17);
}
if (i+15 == s1Len-1 && 0 <= j-15 && j-15 < s2Len) {
row[j-15] = (int8_t)_mm256_extract_epi8_rpl(vWH, 16);
}
if (j-15 == s2Len-1 && 0 <= i+15 && i+15 < s1Len) {
col[(i+15)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 16);
}
if (i+16 == s1Len-1 && 0 <= j-16 && j-16 < s2Len) {
row[j-16] = (int8_t)_mm256_extract_epi8_rpl(vWH, 15);
}
if (j-16 == s2Len-1 && 0 <= i+16 && i+16 < s1Len) {
col[(i+16)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 15);
}
if (i+17 == s1Len-1 && 0 <= j-17 && j-17 < s2Len) {
row[j-17] = (int8_t)_mm256_extract_epi8_rpl(vWH, 14);
}
if (j-17 == s2Len-1 && 0 <= i+17 && i+17 < s1Len) {
col[(i+17)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 14);
}
if (i+18 == s1Len-1 && 0 <= j-18 && j-18 < s2Len) {
row[j-18] = (int8_t)_mm256_extract_epi8_rpl(vWH, 13);
}
if (j-18 == s2Len-1 && 0 <= i+18 && i+18 < s1Len) {
col[(i+18)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 13);
}
if (i+19 == s1Len-1 && 0 <= j-19 && j-19 < s2Len) {
row[j-19] = (int8_t)_mm256_extract_epi8_rpl(vWH, 12);
}
if (j-19 == s2Len-1 && 0 <= i+19 && i+19 < s1Len) {
col[(i+19)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 12);
}
if (i+20 == s1Len-1 && 0 <= j-20 && j-20 < s2Len) {
row[j-20] = (int8_t)_mm256_extract_epi8_rpl(vWH, 11);
}
if (j-20 == s2Len-1 && 0 <= i+20 && i+20 < s1Len) {
col[(i+20)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 11);
}
if (i+21 == s1Len-1 && 0 <= j-21 && j-21 < s2Len) {
row[j-21] = (int8_t)_mm256_extract_epi8_rpl(vWH, 10);
}
if (j-21 == s2Len-1 && 0 <= i+21 && i+21 < s1Len) {
col[(i+21)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 10);
}
if (i+22 == s1Len-1 && 0 <= j-22 && j-22 < s2Len) {
row[j-22] = (int8_t)_mm256_extract_epi8_rpl(vWH, 9);
}
if (j-22 == s2Len-1 && 0 <= i+22 && i+22 < s1Len) {
col[(i+22)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 9);
}
if (i+23 == s1Len-1 && 0 <= j-23 && j-23 < s2Len) {
row[j-23] = (int8_t)_mm256_extract_epi8_rpl(vWH, 8);
}
if (j-23 == s2Len-1 && 0 <= i+23 && i+23 < s1Len) {
col[(i+23)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 8);
}
if (i+24 == s1Len-1 && 0 <= j-24 && j-24 < s2Len) {
row[j-24] = (int8_t)_mm256_extract_epi8_rpl(vWH, 7);
}
if (j-24 == s2Len-1 && 0 <= i+24 && i+24 < s1Len) {
col[(i+24)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 7);
}
if (i+25 == s1Len-1 && 0 <= j-25 && j-25 < s2Len) {
row[j-25] = (int8_t)_mm256_extract_epi8_rpl(vWH, 6);
}
if (j-25 == s2Len-1 && 0 <= i+25 && i+25 < s1Len) {
col[(i+25)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 6);
}
if (i+26 == s1Len-1 && 0 <= j-26 && j-26 < s2Len) {
row[j-26] = (int8_t)_mm256_extract_epi8_rpl(vWH, 5);
}
if (j-26 == s2Len-1 && 0 <= i+26 && i+26 < s1Len) {
col[(i+26)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 5);
}
if (i+27 == s1Len-1 && 0 <= j-27 && j-27 < s2Len) {
row[j-27] = (int8_t)_mm256_extract_epi8_rpl(vWH, 4);
}
if (j-27 == s2Len-1 && 0 <= i+27 && i+27 < s1Len) {
col[(i+27)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 4);
}
if (i+28 == s1Len-1 && 0 <= j-28 && j-28 < s2Len) {
row[j-28] = (int8_t)_mm256_extract_epi8_rpl(vWH, 3);
}
if (j-28 == s2Len-1 && 0 <= i+28 && i+28 < s1Len) {
col[(i+28)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 3);
}
if (i+29 == s1Len-1 && 0 <= j-29 && j-29 < s2Len) {
row[j-29] = (int8_t)_mm256_extract_epi8_rpl(vWH, 2);
}
if (j-29 == s2Len-1 && 0 <= i+29 && i+29 < s1Len) {
col[(i+29)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 2);
}
if (i+30 == s1Len-1 && 0 <= j-30 && j-30 < s2Len) {
row[j-30] = (int8_t)_mm256_extract_epi8_rpl(vWH, 1);
}
if (j-30 == s2Len-1 && 0 <= i+30 && i+30 < s1Len) {
col[(i+30)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 1);
}
if (i+31 == s1Len-1 && 0 <= j-31 && j-31 < s2Len) {
row[j-31] = (int8_t)_mm256_extract_epi8_rpl(vWH, 0);
}
if (j-31 == s2Len-1 && 0 <= i+31 && i+31 < s1Len) {
col[(i+31)] = (int8_t)_mm256_extract_epi8_rpl(vWH, 0);
}
}
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);
}
static inline void arr_store_si256(
int *array,
__m256i vH,
int32_t t,
int32_t seglen,
int32_t d,
int32_t dlen)
{
array[( 0*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 0);
array[( 1*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 1);
array[( 2*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 2);
array[( 3*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 3);
array[( 4*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 4);
array[( 5*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 5);
array[( 6*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 6);
array[( 7*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 7);
array[( 8*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 8);
array[( 9*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 9);
array[(10*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 10);
array[(11*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 11);
array[(12*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 12);
array[(13*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 13);
array[(14*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 14);
array[(15*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 15);
array[(16*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 16);
array[(17*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 17);
array[(18*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 18);
array[(19*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 19);
array[(20*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 20);
array[(21*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 21);
array[(22*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 22);
array[(23*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 23);
array[(24*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 24);
array[(25*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 25);
array[(26*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 26);
array[(27*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 27);
array[(28*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 28);
array[(29*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 29);
array[(30*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 30);
array[(31*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 31);
}
static inline void arr_store(
int *array,
__m256i vH,
int32_t t,
int32_t seglen,
int32_t d,
int32_t dlen,
int32_t bias)
{
array[1LL*( 0*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 0) - bias;
array[1LL*( 1*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 1) - bias;
array[1LL*( 2*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 2) - bias;
array[1LL*( 3*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 3) - bias;
array[1LL*( 4*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 4) - bias;
array[1LL*( 5*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 5) - bias;
array[1LL*( 6*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 6) - bias;
array[1LL*( 7*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 7) - bias;
array[1LL*( 8*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 8) - bias;
array[1LL*( 9*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 9) - bias;
array[1LL*(10*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 10) - bias;
array[1LL*(11*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 11) - bias;
array[1LL*(12*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 12) - bias;
array[1LL*(13*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 13) - bias;
array[1LL*(14*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 14) - bias;
array[1LL*(15*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 15) - bias;
array[1LL*(16*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 16) - bias;
array[1LL*(17*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 17) - bias;
array[1LL*(18*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 18) - bias;
array[1LL*(19*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 19) - bias;
array[1LL*(20*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 20) - bias;
array[1LL*(21*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 21) - bias;
array[1LL*(22*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 22) - bias;
array[1LL*(23*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 23) - bias;
array[1LL*(24*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 24) - bias;
array[1LL*(25*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 25) - bias;
array[1LL*(26*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 26) - bias;
array[1LL*(27*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 27) - bias;
array[1LL*(28*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 28) - bias;
array[1LL*(29*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 29) - bias;
array[1LL*(30*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 30) - bias;
array[1LL*(31*seglen+t)*dlen + d] = (int8_t)_mm256_extract_epi8_rpl(vH, 31) - bias;
}