static inline void arr_store_col(
int *col,
__m128i vH,
int32_t t,
int32_t seglen)
{
col[0*seglen+t] = (int32_t)_mm_extract_epi32_rpl(vH, 0);
col[1*seglen+t] = (int32_t)_mm_extract_epi32_rpl(vH, 1);
col[2*seglen+t] = (int32_t)_mm_extract_epi32_rpl(vH, 2);
col[3*seglen+t] = (int32_t)_mm_extract_epi32_rpl(vH, 3);
}
static inline void arr_store_si128(
int *array,
__m128i vH,
int32_t t,
int32_t seglen,
int32_t d,
int32_t dlen)
{
array[1LL*(0*seglen+t)*dlen + d] = (int32_t)_mm_extract_epi32_rpl(vH, 0);
array[1LL*(1*seglen+t)*dlen + d] = (int32_t)_mm_extract_epi32_rpl(vH, 1);
array[1LL*(2*seglen+t)*dlen + d] = (int32_t)_mm_extract_epi32_rpl(vH, 2);
array[1LL*(3*seglen+t)*dlen + d] = (int32_t)_mm_extract_epi32_rpl(vH, 3);
}
static inline void arr_store_rowcol(
int *row,
int *col,
__m128i vWscore,
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] = (int32_t)_mm_extract_epi32_rpl(vWscore, 3);
}
if (j-0 == s2Len-1 && 0 <= i+0 && i+0 < s1Len) {
col[(i+0)] = (int32_t)_mm_extract_epi32_rpl(vWscore, 3);
}
if (i+1 == s1Len-1 && 0 <= j-1 && j-1 < s2Len) {
row[j-1] = (int32_t)_mm_extract_epi32_rpl(vWscore, 2);
}
if (j-1 == s2Len-1 && 0 <= i+1 && i+1 < s1Len) {
col[(i+1)] = (int32_t)_mm_extract_epi32_rpl(vWscore, 2);
}
if (i+2 == s1Len-1 && 0 <= j-2 && j-2 < s2Len) {
row[j-2] = (int32_t)_mm_extract_epi32_rpl(vWscore, 1);
}
if (j-2 == s2Len-1 && 0 <= i+2 && i+2 < s1Len) {
col[(i+2)] = (int32_t)_mm_extract_epi32_rpl(vWscore, 1);
}
if (i+3 == s1Len-1 && 0 <= j-3 && j-3 < s2Len) {
row[j-3] = (int32_t)_mm_extract_epi32_rpl(vWscore, 0);
}
if (j-3 == s2Len-1 && 0 <= i+3 && i+3 < s1Len) {
col[(i+3)] = (int32_t)_mm_extract_epi32_rpl(vWscore, 0);
}
}
static inline void arr_store_si128(
int *array,
__m128i vWscore,
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[(i+0)*s2Len + (j-0)] = (int32_t)_mm_extract_epi32_rpl(vWscore, 3);
}
if (0 <= i+1 && i+1 < s1Len && 0 <= j-1 && j-1 < s2Len) {
array[(i+1)*s2Len + (j-1)] = (int32_t)_mm_extract_epi32_rpl(vWscore, 2);
}
if (0 <= i+2 && i+2 < s1Len && 0 <= j-2 && j-2 < s2Len) {
array[(i+2)*s2Len + (j-2)] = (int32_t)_mm_extract_epi32_rpl(vWscore, 1);
}
if (0 <= i+3 && i+3 < s1Len && 0 <= j-3 && j-3 < s2Len) {
array[(i+3)*s2Len + (j-3)] = (int32_t)_mm_extract_epi32_rpl(vWscore, 0);
}
}
static inline int32_t _mm_hmax_epi32_rpl(__m128i a) {
a = _mm_max_epi32_rpl(a, _mm_srli_si128(a, 8));
a = _mm_max_epi32_rpl(a, _mm_srli_si128(a, 4));
return _mm_extract_epi32_rpl(a, 0);
}