ksw2_exts2_sse.c

#include <string.h>
#include <stdio.h>
#include <assert.h>
#include "ksw2.h"

#if defined(__aarch64__)

#ifndef __SSE4_1__
#define __SSE4_1__
#endif

#ifndef __SSE2__
#define __SSE2__
#endif

#if !defined(USE_SIMDE)
#include <sse2neon.h>
#else
#include <simde/x86/sse2.h>
#include <simde/x86/sse4.1.h>
#endif

#else

#ifdef __SSE2__
#ifdef USE_SIMDE
#include <simde/x86/sse2.h>
#else
#include <emmintrin.h>
#endif
#ifdef KSW_SSE2_ONLY
#undef __SSE4_1__
#endif

#ifdef __SSE4_1__
#ifdef USE_SIMDE
#include <simde/x86/sse4.1.h>
#else
#include <smmintrin.h>
#endif
#endif
#endif
#endif // defined(__aarch64__) && !defined(USE_SIMDE)

#ifdef __SSE2__

#ifdef KSW_CPU_DISPATCH
#ifdef __SSE4_1__
void ksw_exts2_sse41(void *km, int qlen, const uint8_t *query, int tlen, const uint8_t *target, int8_t m, const int8_t *mat,
				   int8_t q, int8_t e, int8_t q2, int8_t noncan, int zdrop, int8_t junc_bonus, int flag, const uint8_t *junc, ksw_extz_t *ez)
#else
void ksw_exts2_sse2(void *km, int qlen, const uint8_t *query, int tlen, const uint8_t *target, int8_t m, const int8_t *mat,
				   int8_t q, int8_t e, int8_t q2, int8_t noncan, int zdrop, int8_t junc_bonus, int flag, const uint8_t *junc, ksw_extz_t *ez)
#endif
#else
void ksw_exts2_sse(void *km, int qlen, const uint8_t *query, int tlen, const uint8_t *target, int8_t m, const int8_t *mat,
				   int8_t q, int8_t e, int8_t q2, int8_t noncan, int zdrop, int8_t junc_bonus, int flag, const uint8_t *junc, ksw_extz_t *ez)
#endif // ~KSW_CPU_DISPATCH
{
#define __dp_code_block1 \
	z = _mm_load_si128(&s[t]); \
	xt1 = _mm_load_si128(&x[t]);                     /* xt1 <- x[r-1][t..t+15] */ \
	tmp = _mm_srli_si128(xt1, 15);                   /* tmp <- x[r-1][t+15] */ \
	xt1 = _mm_or_si128(_mm_slli_si128(xt1, 1), x1_); /* xt1 <- x[r-1][t-1..t+14] */ \
	x1_ = tmp; \
	vt1 = _mm_load_si128(&v[t]);                     /* vt1 <- v[r-1][t..t+15] */ \
	tmp = _mm_srli_si128(vt1, 15);                   /* tmp <- v[r-1][t+15] */ \
	vt1 = _mm_or_si128(_mm_slli_si128(vt1, 1), v1_); /* vt1 <- v[r-1][t-1..t+14] */ \
	v1_ = tmp; \
	a = _mm_add_epi8(xt1, vt1);                      /* a <- x[r-1][t-1..t+14] + v[r-1][t-1..t+14] */ \
	ut = _mm_load_si128(&u[t]);                      /* ut <- u[t..t+15] */ \
	b = _mm_add_epi8(_mm_load_si128(&y[t]), ut);     /* b <- y[r-1][t..t+15] + u[r-1][t..t+15] */ \
	x2t1= _mm_load_si128(&x2[t]); \
	tmp = _mm_srli_si128(x2t1, 15); \
	x2t1= _mm_or_si128(_mm_slli_si128(x2t1, 1), x21_); \
	x21_= tmp; \
	a2  = _mm_add_epi8(x2t1, vt1); \
	a2a = _mm_add_epi8(a2, _mm_load_si128(&acceptor[t]));

#define __dp_code_block2 \
	_mm_store_si128(&u[t], _mm_sub_epi8(z, vt1));    /* u[r][t..t+15] <- z - v[r-1][t-1..t+14] */ \
	_mm_store_si128(&v[t], _mm_sub_epi8(z, ut));     /* v[r][t..t+15] <- z - u[r-1][t..t+15] */ \
	tmp = _mm_sub_epi8(z, q_); \
	a = _mm_sub_epi8(a, tmp); \
	b = _mm_sub_epi8(b, tmp); \
	a2= _mm_sub_epi8(a2, _mm_sub_epi8(z, q2_));

	int r, t, qe = q + e, n_col_, *off = 0, *off_end = 0, tlen_, qlen_, last_st, last_en, max_sc, min_sc, long_thres, long_diff;
	int with_cigar = !(flag&KSW_EZ_SCORE_ONLY), approx_max = !!(flag&KSW_EZ_APPROX_MAX);
	int32_t *H = 0, H0 = 0, last_H0_t = 0;
	uint8_t *qr, *sf, *mem, *mem2 = 0;
	__m128i q_, q2_, qe_, zero_, sc_mch_, sc_mis_, sc_N_, m1_;
	__m128i *u, *v, *x, *y, *x2, *s, *p = 0, *donor, *acceptor;

	ksw_reset_extz(ez);
	if (m <= 1 || qlen <= 0 || tlen <= 0 || q2 <= q + e) return;

	zero_   = _mm_set1_epi8(0);
	q_      = _mm_set1_epi8(q);
	q2_     = _mm_set1_epi8(q2);
	qe_     = _mm_set1_epi8(q + e);
	sc_mch_ = _mm_set1_epi8(mat[0]);
	sc_mis_ = _mm_set1_epi8(mat[1]);
	sc_N_   = mat[m*m-1] == 0? _mm_set1_epi8(-e) : _mm_set1_epi8(mat[m*m-1]);
	m1_     = _mm_set1_epi8(m - 1); // wildcard

	tlen_ = (tlen + 15) / 16;
	n_col_ = ((qlen < tlen? qlen : tlen) + 15) / 16 + 1;
	qlen_ = (qlen + 15) / 16;
	for (t = 1, max_sc = mat[0], min_sc = mat[1]; t < m * m; ++t) {
		max_sc = max_sc > mat[t]? max_sc : mat[t];
		min_sc = min_sc < mat[t]? min_sc : mat[t];
	}
	if (-min_sc > 2 * (q + e)) return; // otherwise, we won't see any mismatches

	long_thres = (q2 - q) / e - 1;
	if (q2 > q + e + long_thres * e)
		++long_thres;
	long_diff = long_thres * e - (q2 - q);

	mem = (uint8_t*)kcalloc(km, tlen_ * 9 + qlen_ + 1, 16);
	u = (__m128i*)(((size_t)mem + 15) >> 4 << 4); // 16-byte aligned
	v = u + tlen_, x = v + tlen_, y = x + tlen_, x2 = y + tlen_;
	donor = x2 + tlen_, acceptor = donor + tlen_;
	s = acceptor + tlen_, sf = (uint8_t*)(s + tlen_), qr = sf + tlen_ * 16;
	memset(u,  -q - e,  tlen_ * 16 * 4); // this set u, v, x, y (because they are in the same array)
	memset(x2, -q2,     tlen_ * 16);
	if (!approx_max) {
		H = (int32_t*)kmalloc(km, tlen_ * 16 * 4);
		for (t = 0; t < tlen_ * 16; ++t) H[t] = KSW_NEG_INF;
	}
	if (with_cigar) {
		mem2 = (uint8_t*)kmalloc(km, ((size_t)(qlen + tlen - 1) * n_col_ + 1) * 16);
		p = (__m128i*)(((size_t)mem2 + 15) >> 4 << 4);
		off = (int*)kmalloc(km, (qlen + tlen - 1) * sizeof(int) * 2);
		off_end = off + qlen + tlen - 1;
	}

	for (t = 0; t < qlen; ++t) qr[t] = query[qlen - 1 - t];
	memcpy(sf, target, tlen);

	// set the donor and acceptor arrays. TODO: this assumes 0/1/2/3 encoding!
	if (flag & (KSW_EZ_SPLICE_FOR|KSW_EZ_SPLICE_REV)) {
		int semi_cost = flag&KSW_EZ_SPLICE_FLANK? -noncan/2 : 0; // GTr or yAG is worth 0.5 bit; see PMID:18688272
		memset(donor, -noncan, tlen_ * 16);
		memset(acceptor, -noncan, tlen_ * 16);
		if (!(flag & KSW_EZ_REV_CIGAR)) {
			for (t = 0; t < tlen - 4; ++t) {
				int can_type = 0; // type of canonical site: 0=none, 1=GT/AG only, 2=GTr/yAG
				if ((flag & KSW_EZ_SPLICE_FOR) && target[t+1] == 2 && target[t+2] == 3) can_type = 1; // GTr...
				if ((flag & KSW_EZ_SPLICE_REV) && target[t+1] == 1 && target[t+2] == 3) can_type = 1; // CTr...
				if (can_type && (target[t+3] == 0 || target[t+3] == 2)) can_type = 2;
				if (can_type) ((int8_t*)donor)[t] = can_type == 2? 0 : semi_cost;
			}
			if (junc)
				for (t = 0; t < tlen - 1; ++t)
					if (((flag & KSW_EZ_SPLICE_FOR) && (junc[t+1]&1)) || ((flag & KSW_EZ_SPLICE_REV) && (junc[t+1]&8)))
						((int8_t*)donor)[t] += junc_bonus;
			for (t = 2; t < tlen; ++t) {
				int can_type = 0;
				if ((flag & KSW_EZ_SPLICE_FOR) && target[t-1] == 0 && target[t] == 2) can_type = 1; // ...yAG
				if ((flag & KSW_EZ_SPLICE_REV) && target[t-1] == 0 && target[t] == 1) can_type = 1; // ...yAC
				if (can_type && (target[t-2] == 1 || target[t-2] == 3)) can_type = 2;
				if (can_type) ((int8_t*)acceptor)[t] = can_type == 2? 0 : semi_cost;
			}
			if (junc)
				for (t = 0; t < tlen; ++t)
					if (((flag & KSW_EZ_SPLICE_FOR) && (junc[t]&2)) || ((flag & KSW_EZ_SPLICE_REV) && (junc[t]&4)))
						((int8_t*)acceptor)[t] += junc_bonus;
		} else {
			for (t = 0; t < tlen - 4; ++t) {
				int can_type = 0; // type of canonical site: 0=none, 1=GT/AG only, 2=GTr/yAG
				if ((flag & KSW_EZ_SPLICE_FOR) && target[t+1] == 2 && target[t+2] == 0) can_type = 1; // GAy...
				if ((flag & KSW_EZ_SPLICE_REV) && target[t+1] == 1 && target[t+2] == 0) can_type = 1; // CAy...
				if (can_type && (target[t+3] == 1 || target[t+3] == 3)) can_type = 2;
				if (can_type) ((int8_t*)donor)[t] = can_type == 2? 0 : semi_cost;
			}
			if (junc)
				for (t = 0; t < tlen - 1; ++t)
					if (((flag & KSW_EZ_SPLICE_FOR) && (junc[t+1]&2)) || ((flag & KSW_EZ_SPLICE_REV) && (junc[t+1]&4)))
						((int8_t*)donor)[t] += junc_bonus;
			for (t = 2; t < tlen; ++t) {
				int can_type = 0;
				if ((flag & KSW_EZ_SPLICE_FOR) && target[t-1] == 3 && target[t] == 2) can_type = 1; // ...rTG
				if ((flag & KSW_EZ_SPLICE_REV) && target[t-1] == 3 && target[t] == 1) can_type = 1; // ...rTC
				if (can_type && (target[t-2] == 0 || target[t-2] == 2)) can_type = 2;
				if (can_type) ((int8_t*)acceptor)[t] = can_type == 2? 0 : semi_cost;
			}
			if (junc)
				for (t = 0; t < tlen; ++t)
					if (((flag & KSW_EZ_SPLICE_FOR) && (junc[t]&1)) || ((flag & KSW_EZ_SPLICE_REV) && (junc[t]&8)))
						((int8_t*)acceptor)[t] += junc_bonus;
		}
	}

	for (r = 0, last_st = last_en = -1; r < qlen + tlen - 1; ++r) {
		int st = 0, en = tlen - 1, st0, en0, st_, en_;
		int8_t x1, x21, v1, *u8 = (int8_t*)u, *v8 = (int8_t*)v;
		uint8_t *qrr = qr + (qlen - 1 - r);
		__m128i x1_, x21_, v1_;
		// find the boundaries
		if (st < r - qlen + 1) st = r - qlen + 1;
		if (en > r) en = r;
		st0 = st, en0 = en;
		st = st / 16 * 16, en = (en + 16) / 16 * 16 - 1;
		// set boundary conditions
		if (st > 0) {
			if (st - 1 >= last_st && st - 1 <= last_en)
				x1 = ((int8_t*)x)[st - 1], x21 = ((int8_t*)x2)[st - 1], v1 = v8[st - 1]; // (r-1,s-1) calculated in the last round
			else x1 = -q - e, x21 = -q2, v1 = -q - e;
		} else {
			x1 = -q - e, x21 = -q2;
			v1 = r == 0? -q - e : r < long_thres? -e : r == long_thres? long_diff : 0;
		}
		if (en >= r) {
			((int8_t*)y)[r] = -q - e;
			u8[r] = r == 0? -q - e : r < long_thres? -e : r == long_thres? long_diff : 0;
		}
		// loop fission: set scores first
		if (!(flag & KSW_EZ_GENERIC_SC)) {
			for (t = st0; t <= en0; t += 16) {
				__m128i sq, st, tmp, mask;
				sq = _mm_loadu_si128((__m128i*)&sf[t]);
				st = _mm_loadu_si128((__m128i*)&qrr[t]);
				mask = _mm_or_si128(_mm_cmpeq_epi8(sq, m1_), _mm_cmpeq_epi8(st, m1_));
				tmp = _mm_cmpeq_epi8(sq, st);
#ifdef __SSE4_1__
				tmp = _mm_blendv_epi8(sc_mis_, sc_mch_, tmp);
				tmp = _mm_blendv_epi8(tmp,     sc_N_,   mask);
#else
				tmp = _mm_or_si128(_mm_andnot_si128(tmp,  sc_mis_), _mm_and_si128(tmp,  sc_mch_));
				tmp = _mm_or_si128(_mm_andnot_si128(mask, tmp),     _mm_and_si128(mask, sc_N_));
#endif
				_mm_storeu_si128((__m128i*)((int8_t*)s + t), tmp);
			}
		} else {
			for (t = st0; t <= en0; ++t)
				((uint8_t*)s)[t] = mat[sf[t] * m + qrr[t]];
		}
		// core loop
		x1_  = _mm_cvtsi32_si128((uint8_t)x1);
		x21_ = _mm_cvtsi32_si128((uint8_t)x21);
		v1_  = _mm_cvtsi32_si128((uint8_t)v1);
		st_ = st / 16, en_ = en / 16;
		assert(en_ - st_ + 1 <= n_col_);
		if (!with_cigar) { // score only
			for (t = st_; t <= en_; ++t) {
				__m128i z, a, b, a2, a2a, xt1, x2t1, vt1, ut, tmp;
				__dp_code_block1;
#ifdef __SSE4_1__
				z = _mm_max_epi8(z, a);
				z = _mm_max_epi8(z, b);
				z = _mm_max_epi8(z, a2a);
				__dp_code_block2; // save u[] and v[]; update a, b and a2
				_mm_store_si128(&x[t],  _mm_sub_epi8(_mm_max_epi8(a,  zero_), qe_));
				_mm_store_si128(&y[t],  _mm_sub_epi8(_mm_max_epi8(b,  zero_), qe_));
				tmp = _mm_load_si128(&donor[t]);
				_mm_store_si128(&x2[t], _mm_sub_epi8(_mm_max_epi8(a2, tmp), q2_));
#else
				tmp = _mm_cmpgt_epi8(a,  z);
				z = _mm_or_si128(_mm_andnot_si128(tmp, z), _mm_and_si128(tmp, a));
				tmp = _mm_cmpgt_epi8(b,  z);
				z = _mm_or_si128(_mm_andnot_si128(tmp, z), _mm_and_si128(tmp, b));
				tmp = _mm_cmpgt_epi8(a2a, z);
				z = _mm_or_si128(_mm_andnot_si128(tmp, z), _mm_and_si128(tmp, a2a));
				__dp_code_block2;
				tmp = _mm_cmpgt_epi8(a, zero_);
				_mm_store_si128(&x[t],  _mm_sub_epi8(_mm_and_si128(tmp, a),  qe_));
				tmp = _mm_cmpgt_epi8(b, zero_);
				_mm_store_si128(&y[t],  _mm_sub_epi8(_mm_and_si128(tmp, b),  qe_));
				tmp = _mm_load_si128(&donor[t]); // TODO: check if this is correct
				tmp = _mm_cmpgt_epi8(a2, tmp);
				tmp = _mm_or_si128(_mm_andnot_si128(tmp, tmp), _mm_and_si128(tmp, a2));
				_mm_store_si128(&x2[t], _mm_sub_epi8(tmp, q2_));
#endif
			}
		} else if (!(flag&KSW_EZ_RIGHT)) { // gap left-alignment
			__m128i *pr = p + r * n_col_ - st_;
			off[r] = st, off_end[r] = en;
			for (t = st_; t <= en_; ++t) {
				__m128i d, z, a, b, a2, a2a, xt1, x2t1, vt1, ut, tmp, tmp2;
				__dp_code_block1;
#ifdef __SSE4_1__
				d = _mm_and_si128(_mm_cmpgt_epi8(a, z), _mm_set1_epi8(1));       // d = a  > z? 1 : 0
				z = _mm_max_epi8(z, a);
				d = _mm_blendv_epi8(d, _mm_set1_epi8(2), _mm_cmpgt_epi8(b,  z)); // d = b  > z? 2 : d
				z = _mm_max_epi8(z, b);
				d = _mm_blendv_epi8(d, _mm_set1_epi8(3), _mm_cmpgt_epi8(a2a, z)); // d = a2 > z? 3 : d
				z = _mm_max_epi8(z, a2a);
#else // we need to emulate SSE4.1 intrinsics _mm_max_epi8() and _mm_blendv_epi8()
				tmp = _mm_cmpgt_epi8(a,  z);
				d = _mm_and_si128(tmp, _mm_set1_epi8(1));
				z = _mm_or_si128(_mm_andnot_si128(tmp, z), _mm_and_si128(tmp, a));
				tmp = _mm_cmpgt_epi8(b,  z);
				d = _mm_or_si128(_mm_andnot_si128(tmp, d), _mm_and_si128(tmp, _mm_set1_epi8(2)));
				z = _mm_or_si128(_mm_andnot_si128(tmp, z), _mm_and_si128(tmp, b));
				tmp = _mm_cmpgt_epi8(a2a, z);
				d = _mm_or_si128(_mm_andnot_si128(tmp, d), _mm_and_si128(tmp, _mm_set1_epi8(3)));
				z = _mm_or_si128(_mm_andnot_si128(tmp, z), _mm_and_si128(tmp, a2a));
#endif
				__dp_code_block2;
				tmp = _mm_cmpgt_epi8(a, zero_);
				_mm_store_si128(&x[t],  _mm_sub_epi8(_mm_and_si128(tmp, a),  qe_));
				d = _mm_or_si128(d, _mm_and_si128(tmp, _mm_set1_epi8(0x08))); // d = a > 0? 1<<3 : 0
				tmp = _mm_cmpgt_epi8(b, zero_);
				_mm_store_si128(&y[t],  _mm_sub_epi8(_mm_and_si128(tmp, b),  qe_));
				d = _mm_or_si128(d, _mm_and_si128(tmp, _mm_set1_epi8(0x10))); // d = b > 0? 1<<4 : 0

				tmp2 = _mm_load_si128(&donor[t]);
				tmp = _mm_cmpgt_epi8(a2, tmp2);
#ifdef __SSE4_1__
				tmp2 = _mm_max_epi8(a2, tmp2);
#else
				tmp2 = _mm_or_si128(_mm_andnot_si128(tmp, tmp2), _mm_and_si128(tmp, a2));
#endif
				_mm_store_si128(&x2[t], _mm_sub_epi8(tmp2, q2_));
				d = _mm_or_si128(d, _mm_and_si128(tmp, _mm_set1_epi8(0x20)));
				_mm_store_si128(&pr[t], d);
			}
		} else { // gap right-alignment
			__m128i *pr = p + r * n_col_ - st_;
			off[r] = st, off_end[r] = en;
			for (t = st_; t <= en_; ++t) {
				__m128i d, z, a, b, a2, a2a, xt1, x2t1, vt1, ut, tmp, tmp2;
				__dp_code_block1;
#ifdef __SSE4_1__
				d = _mm_andnot_si128(_mm_cmpgt_epi8(z, a), _mm_set1_epi8(1));    // d = z > a?  0 : 1
				z = _mm_max_epi8(z, a);
				d = _mm_blendv_epi8(_mm_set1_epi8(2), d, _mm_cmpgt_epi8(z, b));  // d = z > b?  d : 2
				z = _mm_max_epi8(z, b);
				d = _mm_blendv_epi8(_mm_set1_epi8(3), d, _mm_cmpgt_epi8(z, a2a)); // d = z > a2? d : 3
				z = _mm_max_epi8(z, a2a);
#else // we need to emulate SSE4.1 intrinsics _mm_max_epi8() and _mm_blendv_epi8()
				tmp = _mm_cmpgt_epi8(z, a);
				d = _mm_andnot_si128(tmp, _mm_set1_epi8(1));
				z = _mm_or_si128(_mm_and_si128(tmp, z), _mm_andnot_si128(tmp, a));
				tmp = _mm_cmpgt_epi8(z, b);
				d = _mm_or_si128(_mm_and_si128(tmp, d), _mm_andnot_si128(tmp, _mm_set1_epi8(2)));
				z = _mm_or_si128(_mm_and_si128(tmp, z), _mm_andnot_si128(tmp, b));
				tmp = _mm_cmpgt_epi8(z, a2a);
				d = _mm_or_si128(_mm_and_si128(tmp, d), _mm_andnot_si128(tmp, _mm_set1_epi8(3)));
				z = _mm_or_si128(_mm_and_si128(tmp, z), _mm_andnot_si128(tmp, a2a));
#endif
				__dp_code_block2;
				tmp = _mm_cmpgt_epi8(zero_, a);
				_mm_store_si128(&x[t],  _mm_sub_epi8(_mm_andnot_si128(tmp, a),  qe_));
				d = _mm_or_si128(d, _mm_andnot_si128(tmp, _mm_set1_epi8(0x08))); // d = a > 0? 1<<3 : 0
				tmp = _mm_cmpgt_epi8(zero_, b);
				_mm_store_si128(&y[t],  _mm_sub_epi8(_mm_andnot_si128(tmp, b),  qe_));
				d = _mm_or_si128(d, _mm_andnot_si128(tmp, _mm_set1_epi8(0x10))); // d = b > 0? 1<<4 : 0

				tmp2 = _mm_load_si128(&donor[t]);
				tmp = _mm_cmpgt_epi8(tmp2, a2);
#ifdef __SSE4_1__
				tmp2 = _mm_max_epi8(tmp2, a2);
#else
				tmp2 = _mm_or_si128(_mm_andnot_si128(tmp, a2), _mm_and_si128(tmp, tmp2));
#endif
				_mm_store_si128(&x2[t], _mm_sub_epi8(tmp2, q2_));
				d = _mm_or_si128(d, _mm_andnot_si128(tmp, _mm_set1_epi8(0x20))); // d = a > 0? 1<<5 : 0
				_mm_store_si128(&pr[t], d);
			}
		}
		if (!approx_max) { // find the exact max with a 32-bit score array
			int32_t max_H, max_t;
			// compute H[], max_H and max_t
			if (r > 0) {
				int32_t HH[4], tt[4], en1 = st0 + (en0 - st0) / 4 * 4, i;
				__m128i max_H_, max_t_;
				max_H = H[en0] = en0 > 0? H[en0-1] + u8[en0] : H[en0] + v8[en0]; // special casing the last element
				max_t = en0;
				max_H_ = _mm_set1_epi32(max_H);
				max_t_ = _mm_set1_epi32(max_t);
				for (t = st0; t < en1; t += 4) { // this implements: H[t]+=v8[t]-qe; if(H[t]>max_H) max_H=H[t],max_t=t;
					__m128i H1, tmp, t_;
					H1 = _mm_loadu_si128((__m128i*)&H[t]);
					t_ = _mm_setr_epi32(v8[t], v8[t+1], v8[t+2], v8[t+3]);
					H1 = _mm_add_epi32(H1, t_);
					_mm_storeu_si128((__m128i*)&H[t], H1);
					t_ = _mm_set1_epi32(t);
					tmp = _mm_cmpgt_epi32(H1, max_H_);
#ifdef __SSE4_1__
					max_H_ = _mm_blendv_epi8(max_H_, H1, tmp);
					max_t_ = _mm_blendv_epi8(max_t_, t_, tmp);
#else
					max_H_ = _mm_or_si128(_mm_and_si128(tmp, H1), _mm_andnot_si128(tmp, max_H_));
					max_t_ = _mm_or_si128(_mm_and_si128(tmp, t_), _mm_andnot_si128(tmp, max_t_));
#endif
				}
				_mm_storeu_si128((__m128i*)HH, max_H_);
				_mm_storeu_si128((__m128i*)tt, max_t_);
				for (i = 0; i < 4; ++i)
					if (max_H < HH[i]) max_H = HH[i], max_t = tt[i] + i;
				for (; t < en0; ++t) { // for the rest of values that haven't been computed with SSE
					H[t] += (int32_t)v8[t];
					if (H[t] > max_H)
						max_H = H[t], max_t = t;
				}
			} else H[0] = v8[0] - qe, max_H = H[0], max_t = 0; // special casing r==0
			// update ez
			if (en0 == tlen - 1 && H[en0] > ez->mte)
				ez->mte = H[en0], ez->mte_q = r - en;
			if (r - st0 == qlen - 1 && H[st0] > ez->mqe)
				ez->mqe = H[st0], ez->mqe_t = st0;
			if (ksw_apply_zdrop(ez, 1, max_H, r, max_t, zdrop, 0)) break;
			if (r == qlen + tlen - 2 && en0 == tlen - 1)
				ez->score = H[tlen - 1];
		} else { // find approximate max; Z-drop might be inaccurate, too.
			if (r > 0) {
				if (last_H0_t >= st0 && last_H0_t <= en0 && last_H0_t + 1 >= st0 && last_H0_t + 1 <= en0) {
					int32_t d0 = v8[last_H0_t];
					int32_t d1 = u8[last_H0_t + 1];
					if (d0 > d1) H0 += d0;
					else H0 += d1, ++last_H0_t;
				} else if (last_H0_t >= st0 && last_H0_t <= en0) {
					H0 += v8[last_H0_t];
				} else {
					++last_H0_t, H0 += u8[last_H0_t];
				}
			} else H0 = v8[0] - qe, last_H0_t = 0;
			if ((flag & KSW_EZ_APPROX_DROP) && ksw_apply_zdrop(ez, 1, H0, r, last_H0_t, zdrop, 0)) break;
			if (r == qlen + tlen - 2 && en0 == tlen - 1)
				ez->score = H0;
		}
		last_st = st, last_en = en;
		//for (t = st0; t <= en0; ++t) printf("(%d,%d)\t(%d,%d,%d,%d)\t%d\n", r, t, ((int8_t*)u)[t], ((int8_t*)v)[t], ((int8_t*)x)[t], ((int8_t*)y)[t], H[t]); // for debugging
	}
	kfree(km, mem);
	if (!approx_max) kfree(km, H);
	if (with_cigar) { // backtrack
		int rev_cigar = !!(flag & KSW_EZ_REV_CIGAR);
		if (!ez->zdropped && !(flag&KSW_EZ_EXTZ_ONLY))
			ksw_backtrack(km, 1, rev_cigar, long_thres, (uint8_t*)p, off, off_end, n_col_*16, tlen-1, qlen-1, &ez->m_cigar, &ez->n_cigar, &ez->cigar);
		else if (ez->max_t >= 0 && ez->max_q >= 0)
			ksw_backtrack(km, 1, rev_cigar, long_thres, (uint8_t*)p, off, off_end, n_col_*16, ez->max_t, ez->max_q, &ez->m_cigar, &ez->n_cigar, &ez->cigar);
		kfree(km, mem2); kfree(km, off);
	}
}
#endif // __SSE2__