static inline int8_t _mm_hmax_epi8_rpl(__m128i a) {
a = _mm_max_epi8_rpl(a, _mm_srli_si128(a, 8));
a = _mm_max_epi8_rpl(a, _mm_srli_si128(a, 4));
a = _mm_max_epi8_rpl(a, _mm_srli_si128(a, 2));
a = _mm_max_epi8_rpl(a, _mm_srli_si128(a, 1));
return _mm_extract_epi8_rpl(a, 0);
}
static inline void arr_store_col(
int *col,
__m128i vH,
int32_t t,
int32_t seglen)
{
col[ 0*seglen+t] = (int8_t)_mm_extract_epi8_rpl(vH, 0);
col[ 1*seglen+t] = (int8_t)_mm_extract_epi8_rpl(vH, 1);
col[ 2*seglen+t] = (int8_t)_mm_extract_epi8_rpl(vH, 2);
col[ 3*seglen+t] = (int8_t)_mm_extract_epi8_rpl(vH, 3);
col[ 4*seglen+t] = (int8_t)_mm_extract_epi8_rpl(vH, 4);
col[ 5*seglen+t] = (int8_t)_mm_extract_epi8_rpl(vH, 5);
col[ 6*seglen+t] = (int8_t)_mm_extract_epi8_rpl(vH, 6);
col[ 7*seglen+t] = (int8_t)_mm_extract_epi8_rpl(vH, 7);
col[ 8*seglen+t] = (int8_t)_mm_extract_epi8_rpl(vH, 8);
col[ 9*seglen+t] = (int8_t)_mm_extract_epi8_rpl(vH, 9);
col[10*seglen+t] = (int8_t)_mm_extract_epi8_rpl(vH, 10);
col[11*seglen+t] = (int8_t)_mm_extract_epi8_rpl(vH, 11);
col[12*seglen+t] = (int8_t)_mm_extract_epi8_rpl(vH, 12);
col[13*seglen+t] = (int8_t)_mm_extract_epi8_rpl(vH, 13);
col[14*seglen+t] = (int8_t)_mm_extract_epi8_rpl(vH, 14);
col[15*seglen+t] = (int8_t)_mm_extract_epi8_rpl(vH, 15);
}
static inline void arr_store_si128(
int *array,
__m128i vH,
int32_t t,
int32_t seglen,
int32_t d,
int32_t dlen)
{
array[( 0*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 0);
array[( 1*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 1);
array[( 2*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 2);
array[( 3*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 3);
array[( 4*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 4);
array[( 5*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 5);
array[( 6*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 6);
array[( 7*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 7);
array[( 8*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 8);
array[( 9*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 9);
array[(10*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 10);
array[(11*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 11);
array[(12*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 12);
array[(13*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 13);
array[(14*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 14);
array[(15*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 15);
}
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)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 15);
}
if (0 <= i+1 && i+1 < s1Len && 0 <= j-1 && j-1 < s2Len) {
array[(i+1)*s2Len + (j-1)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 14);
}
if (0 <= i+2 && i+2 < s1Len && 0 <= j-2 && j-2 < s2Len) {
array[(i+2)*s2Len + (j-2)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 13);
}
if (0 <= i+3 && i+3 < s1Len && 0 <= j-3 && j-3 < s2Len) {
array[(i+3)*s2Len + (j-3)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 12);
}
if (0 <= i+4 && i+4 < s1Len && 0 <= j-4 && j-4 < s2Len) {
array[(i+4)*s2Len + (j-4)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 11);
}
if (0 <= i+5 && i+5 < s1Len && 0 <= j-5 && j-5 < s2Len) {
array[(i+5)*s2Len + (j-5)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 10);
}
if (0 <= i+6 && i+6 < s1Len && 0 <= j-6 && j-6 < s2Len) {
array[(i+6)*s2Len + (j-6)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 9);
}
if (0 <= i+7 && i+7 < s1Len && 0 <= j-7 && j-7 < s2Len) {
array[(i+7)*s2Len + (j-7)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 8);
}
if (0 <= i+8 && i+8 < s1Len && 0 <= j-8 && j-8 < s2Len) {
array[(i+8)*s2Len + (j-8)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 7);
}
if (0 <= i+9 && i+9 < s1Len && 0 <= j-9 && j-9 < s2Len) {
array[(i+9)*s2Len + (j-9)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 6);
}
if (0 <= i+10 && i+10 < s1Len && 0 <= j-10 && j-10 < s2Len) {
array[(i+10)*s2Len + (j-10)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 5);
}
if (0 <= i+11 && i+11 < s1Len && 0 <= j-11 && j-11 < s2Len) {
array[(i+11)*s2Len + (j-11)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 4);
}
if (0 <= i+12 && i+12 < s1Len && 0 <= j-12 && j-12 < s2Len) {
array[(i+12)*s2Len + (j-12)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 3);
}
if (0 <= i+13 && i+13 < s1Len && 0 <= j-13 && j-13 < s2Len) {
array[(i+13)*s2Len + (j-13)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 2);
}
if (0 <= i+14 && i+14 < s1Len && 0 <= j-14 && j-14 < s2Len) {
array[(i+14)*s2Len + (j-14)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 1);
}
if (0 <= i+15 && i+15 < s1Len && 0 <= j-15 && j-15 < s2Len) {
array[(i+15)*s2Len + (j-15)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 0);
}
}
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] = (int8_t)_mm_extract_epi8_rpl(vWscore, 15);
}
if (j-0 == s2Len-1 && 0 <= i+0 && i+0 < s1Len) {
col[(i+0)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 15);
}
if (i+1 == s1Len-1 && 0 <= j-1 && j-1 < s2Len) {
row[j-1] = (int8_t)_mm_extract_epi8_rpl(vWscore, 14);
}
if (j-1 == s2Len-1 && 0 <= i+1 && i+1 < s1Len) {
col[(i+1)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 14);
}
if (i+2 == s1Len-1 && 0 <= j-2 && j-2 < s2Len) {
row[j-2] = (int8_t)_mm_extract_epi8_rpl(vWscore, 13);
}
if (j-2 == s2Len-1 && 0 <= i+2 && i+2 < s1Len) {
col[(i+2)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 13);
}
if (i+3 == s1Len-1 && 0 <= j-3 && j-3 < s2Len) {
row[j-3] = (int8_t)_mm_extract_epi8_rpl(vWscore, 12);
}
if (j-3 == s2Len-1 && 0 <= i+3 && i+3 < s1Len) {
col[(i+3)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 12);
}
if (i+4 == s1Len-1 && 0 <= j-4 && j-4 < s2Len) {
row[j-4] = (int8_t)_mm_extract_epi8_rpl(vWscore, 11);
}
if (j-4 == s2Len-1 && 0 <= i+4 && i+4 < s1Len) {
col[(i+4)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 11);
}
if (i+5 == s1Len-1 && 0 <= j-5 && j-5 < s2Len) {
row[j-5] = (int8_t)_mm_extract_epi8_rpl(vWscore, 10);
}
if (j-5 == s2Len-1 && 0 <= i+5 && i+5 < s1Len) {
col[(i+5)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 10);
}
if (i+6 == s1Len-1 && 0 <= j-6 && j-6 < s2Len) {
row[j-6] = (int8_t)_mm_extract_epi8_rpl(vWscore, 9);
}
if (j-6 == s2Len-1 && 0 <= i+6 && i+6 < s1Len) {
col[(i+6)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 9);
}
if (i+7 == s1Len-1 && 0 <= j-7 && j-7 < s2Len) {
row[j-7] = (int8_t)_mm_extract_epi8_rpl(vWscore, 8);
}
if (j-7 == s2Len-1 && 0 <= i+7 && i+7 < s1Len) {
col[(i+7)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 8);
}
if (i+8 == s1Len-1 && 0 <= j-8 && j-8 < s2Len) {
row[j-8] = (int8_t)_mm_extract_epi8_rpl(vWscore, 7);
}
if (j-8 == s2Len-1 && 0 <= i+8 && i+8 < s1Len) {
col[(i+8)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 7);
}
if (i+9 == s1Len-1 && 0 <= j-9 && j-9 < s2Len) {
row[j-9] = (int8_t)_mm_extract_epi8_rpl(vWscore, 6);
}
if (j-9 == s2Len-1 && 0 <= i+9 && i+9 < s1Len) {
col[(i+9)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 6);
}
if (i+10 == s1Len-1 && 0 <= j-10 && j-10 < s2Len) {
row[j-10] = (int8_t)_mm_extract_epi8_rpl(vWscore, 5);
}
if (j-10 == s2Len-1 && 0 <= i+10 && i+10 < s1Len) {
col[(i+10)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 5);
}
if (i+11 == s1Len-1 && 0 <= j-11 && j-11 < s2Len) {
row[j-11] = (int8_t)_mm_extract_epi8_rpl(vWscore, 4);
}
if (j-11 == s2Len-1 && 0 <= i+11 && i+11 < s1Len) {
col[(i+11)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 4);
}
if (i+12 == s1Len-1 && 0 <= j-12 && j-12 < s2Len) {
row[j-12] = (int8_t)_mm_extract_epi8_rpl(vWscore, 3);
}
if (j-12 == s2Len-1 && 0 <= i+12 && i+12 < s1Len) {
col[(i+12)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 3);
}
if (i+13 == s1Len-1 && 0 <= j-13 && j-13 < s2Len) {
row[j-13] = (int8_t)_mm_extract_epi8_rpl(vWscore, 2);
}
if (j-13 == s2Len-1 && 0 <= i+13 && i+13 < s1Len) {
col[(i+13)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 2);
}
if (i+14 == s1Len-1 && 0 <= j-14 && j-14 < s2Len) {
row[j-14] = (int8_t)_mm_extract_epi8_rpl(vWscore, 1);
}
if (j-14 == s2Len-1 && 0 <= i+14 && i+14 < s1Len) {
col[(i+14)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 1);
}
if (i+15 == s1Len-1 && 0 <= j-15 && j-15 < s2Len) {
row[j-15] = (int8_t)_mm_extract_epi8_rpl(vWscore, 0);
}
if (j-15 == s2Len-1 && 0 <= i+15 && i+15 < s1Len) {
col[(i+15)] = (int8_t)_mm_extract_epi8_rpl(vWscore, 0);
}
}
static inline void arr_store(
int *array,
__m128i 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)_mm_extract_epi8_rpl(vH, 0) - bias;
array[1LL*( 1*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 1) - bias;
array[1LL*( 2*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 2) - bias;
array[1LL*( 3*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 3) - bias;
array[1LL*( 4*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 4) - bias;
array[1LL*( 5*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 5) - bias;
array[1LL*( 6*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 6) - bias;
array[1LL*( 7*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 7) - bias;
array[1LL*( 8*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 8) - bias;
array[1LL*( 9*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 9) - bias;
array[1LL*(10*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 10) - bias;
array[1LL*(11*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 11) - bias;
array[1LL*(12*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 12) - bias;
array[1LL*(13*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 13) - bias;
array[1LL*(14*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 14) - bias;
array[1LL*(15*seglen+t)*dlen + d] = (int8_t)_mm_extract_epi8_rpl(vH, 15) - bias;
}