Update README.md

.gitmodules

@@ -1,3 +1,6 @@
 [submodule "lib/simde"]
 	path = lib/simde
 	url = https://github.com/nemequ/simde.git
+[submodule "ext/TAL"]
+	path = ext/TAL
+	url = https://github.com/IntelLabs/Trans-Omics-Acceleration-Library.git

Makefile

@@ -1,6 +1,43 @@
 CFLAGS=		-g -Wall -O2 -Wc++-compat #-Wextra
-CPPFLAGS=	-DHAVE_KALLOC
-INCLUDES=
+CPPFLAGS=	-DHAVE_KALLOC #-march=native #-DALIGN_AVX -DPARALLEL_CHAINING #-DMANUAL_PROFILING
+COMP_FLAG = -march=native
+
+ifeq ($(avx2_compile), 1)
+	COMP_FLAG = -mavx2
+endif
+
+#CPPFLAGS=	-DHAVE_KALLOC -mavx2 -DALIGN_AVX -DAPPLY_AVX2 -DPARALLEL_CHAINING #-DLISA_HASH -DUINT64 -DVECTORIZE #-DMANUAL_PROFILING
+#CPPFLAGS=	-DHAVE_KALLOC -mavx2 -DPARALLEL_CHAINING  #-DMANUAL_PROFILING
+
+OPT_FLAGS= -DPARALLEL_CHAINING -DALIGN_AVX -DAPPLY_AVX2 
+OPT_FLAGS+=$(COMP_FLAG)
+ifeq ($(lhash_index), 1)	
+	CPPFLAGS+=	-DLISA_INDEX 
+endif
+ifeq ($(lhash), 1)	
+	OPT_FLAGS+=	-DLISA_HASH -DUINT64 -DVECTORIZE 
+endif
+ifeq ($(manual_profile), 1)
+	CPPFLAGS+= -DMANUAL_PROFILING 
+endif
+
+#ifeq ($(use_avx2), 1)
+#	OPT_FLAGS+= -DAPPLY_AVX2
+#endif
+
+ifeq ($(disable_output), 1)
+	CPPFLAGS+= -DDISABLE_OUTPUT
+endif
+
+ifeq ($(no_opt),)
+	CPPFLAGS+= $(OPT_FLAGS)
+endif
+
+
+
+#INCLUDES=
+#INCLUDES=	-I./ext/TAL_offline/src/LISA-hash #-I./ext/TAL/src/dynamic-programming 
+INCLUDES=	-I./ext/TAL/src/LISA-hash -I./ext/TAL/src/dynamic-programming 
 OBJS=		kthread.o kalloc.o misc.o bseq.o sketch.o sdust.o options.o index.o \
 			lchain.o align.o hit.o seed.o map.o format.o pe.o esterr.o splitidx.o \
 			ksw2_ll_sse.o
@@ -8,9 +45,14 @@ PROG=		minimap2
 PROG_EXTRA=	sdust minimap2-lite
 LIBS=		-lm -lz -lpthread
 
+CC=$(CXX)
+ifeq ($(CC), g++)
+	CC=g++ -std=c++11
+endif
+
 ifeq ($(arm_neon),) # if arm_neon is not defined
 ifeq ($(sse2only),) # if sse2only is not defined
-	OBJS+=ksw2_extz2_sse41.o ksw2_extd2_sse41.o ksw2_exts2_sse41.o ksw2_extz2_sse2.o ksw2_extd2_sse2.o ksw2_exts2_sse2.o ksw2_dispatch.o
+	OBJS+=ksw2_extz2_sse41.o ksw2_extd2_sse41.o ksw2_exts2_sse41.o ksw2_extz2_sse2.o ksw2_extd2_sse2.o ksw2_exts2_sse2.o ksw2_dispatch.o ksw2_extd2_avx.o
 else                # if sse2only is defined
 	OBJS+=ksw2_extz2_sse.o ksw2_extd2_sse.o ksw2_exts2_sse.o
 endif
@@ -56,6 +98,17 @@ libminimap2.a:$(OBJS)
 sdust:sdust.c kalloc.o kalloc.h kdq.h kvec.h kseq.h ketopt.h sdust.h
 		$(CC) -D_SDUST_MAIN $(CFLAGS) $< kalloc.o -o $@ -lz
 
+multi:
+		$(MAKE) clean
+		$(MAKE)
+		mv minimap2 mm2-fast
+		$(MAKE) clean
+		$(MAKE) lhash=1
+		mv minimap2 mm2-fast-lhash
+		$(MAKE) clean
+		$(MAKE) no_opt=1
+		mv minimap2 mm2-fast-no-opt
+
 # SSE-specific targets on x86/x86_64
 
 ifeq ($(arm_neon),)   # if arm_neon is defined, compile this target with the default setting (i.e. no -msse2)

NEWS.md

@@ -1,35 +1,3 @@
-Release 2.24-r1122 (26 December 2021)
--------------------------------------
-
-This release improves alignment around long poorly aligned regions. Older
-minimap2 may chain through such regions in rare cases which may result in
-missing alignments later. The issue has become worse since the the change of
-the chaining algorithm in v2.19. v2.23 implements an incomplete remedy. This
-release provides a better solution with a X-drop-like heuristic and by enabling
-two-bandwidth chaining in the assembly mode.
-
-(2.24: 26 December 2021, r1122)
-
-
-
-Release 2.23-r1111 (18 November 2021)
--------------------------------------
-
-Notable changes:
-
- * Bugfix: fixed missing alignments around long inversions (#806 and #816).
-   This bug affected v2.19 through v2.22.
-
- * Improvement: avoid extremely long mapping time for pathologic reads with
-   highly repeated k-mers not in the reference (#771). Use --q-occ-frac=0
-   to disable the new heuristic.
-
- * Change: use --cap-kalloc=1g by default.
-
-(2.23: 18 November 2021, r1111)
-
-
-
 Release 2.22-r1101 (7 August 2021)
 ----------------------------------
 

README.md

@@ -1,3 +1,76 @@
+## mm2-fast
+### Introduction
+mm2-fast is an accelerated implementation of minimap2 on modern CPUs. mm2-fast accelerates all the three major modules of minimap2: (a) seeding, (b) chaining, and (c) pairwise alignment, achieving up to 1.8x speedup using AVX512 over minimap2. 
+mm2-fast is a drop-in replacement of minimap2, providing the same functionality with the exact same output.
+In the current version, all the modules are optimized using **AVX-512** and **AVX2** vectorization. Detailed benchmark results are available in our [preprint](https://doi.org/10.1101/2021.07.21.453294).
+
+### System requirement
+Operating System: Linux   
+mm2-fast was tested using g++ (GCC) 9.2.0 and icpc version 19.1.3.304   
+Architecture: x86\_64 CPUs with [AVX512, AVX2](https://en.wikipedia.org/wiki/Advanced_Vector_Extensions)   
+Memory requirement: ~30GB for human genome   
+
+### Installation
+Clone the *fast-contrib-v2.22* branch from minimap2 github page. The source code can be compiled by using *make* command. It only takes a few seconds.
+```
+git clone --recursive https://github.com/lh3/minimap2.git -b fast-contrib-v2.22 mm2-fast   
+cd mm2-fast
+make 
+```
+
+### Usage
+The usage of mm2-fast is same as minimap2. Here is an example of mapping ONT reads with test data.
+```sh
+./minimap2 -ax map-ont test/MT-human.fa test/MT-orang.fa > mm2-fast_output
+```
+
+### Accuracy evaluation
+As mm2-fast is an accelerated version of minimap2-v2.22, the output of mm2-fast can be verified against minimap2-v2.22. Note that the optimized chaining in mm2-fast is strictly required to be run with a chaining parameter *max-chain-skip=infinity*. Note that having parameter *max-chain-skip=infinity* leads to higher chaining precision. Therefore, for correctness verification, minimap2 should run with a larger value of *max-chain-skip* parameter. Follow the below steps to verify the accuracy of mm2-fast.
+```sh
+git clone --recursive https://github.com/lh3/minimap2.git -b fast-contrib-v2.22 mm2-fast   
+cd mm2-fast && make 
+./minimap2 -ax map-ont test/MT-human.fa test/MT-orang.fa --max-chain-skip=1000000 > mm2-fast_output
+```
+```sh
+git clone https://github.com/lh3/minimap2.git -b v2.22
+cd minimap2 && make
+./minimap2 -ax map-ont test/MT-human.fa test/MT-orang.fa --max-chain-skip=1000000 > minimap2_output
+```
+The output generated by minimap2 and mm2-fast should match. 
+```sh
+diff minimap2_output mm2-fast_output > diff_result
+```
+The file ```diff_result``` should be empty, meaning a difference of 0 lines.
+
+### Advanced options
+The default compilation using make applies two optimizations: vectorized chaining and sequence alignment. The learned-indexes based seeding is disabled by default as it requires availability of [Rust](https://en.wikipedia.org/wiki/Rust_(programming_language)). This is because the learned hash-table uses an external training library that runs on Rust. Rust is trivial to install, see https://rustup.rs/ and add its path to .bashrc file. Rust installation only takes a few seconds. Following are the steps to enable learned hash table optimization in mm2-fast:
+```sh
+# Start by building learned hash table index for optimized seeding module 
+./build_rmi.sh test/MT-human.fa map-ont               ##Takes two arguments: 1. path-to-reference-seq-file 2. preset. 
+						      ##For human genome, this step should take around 2-3 minutes to finish.    
+
+# Next, compile and run the mapping phase  
+make clean && make lhash=1
+./minimap2 -ax map-ont test/MT-human.fa test/MT-orang.fa > mm2-fast-lhash_output
+```
+To compile mm2-fast with all optimizations turned off and switch back to default minimap2, use the following command during compilation. This could be useful for debugging.
+```sh
+make clean && make no_opt=1
+```
+
+### Performance
+We have observed up to 1.8x speedup across datasets (please refer to the paper for more details). For example, for the randomly sampled 100K reads from ["HG002\_GM24385\_1\_2\_3\_Guppy\_3.6.0\_prom.fastq.gz"](https://precision.fda.gov/challenges/10/view), minimap2 takes 92 seconds, while mm2-fast takes 54 seconds to map against the human genome on a 28 cores Intel® Xeon® Platinum 8280 CPUs.  Our sampled datasets with 100K reads are available [here](https://drive.google.com/drive/folders/1131j7ejHdT7QZnjxLcTLi5qqwYcfFbuv).
+
+### Future Plans
+
+
+### Citations
+["Accelerating long-read analysis on modern CPUs"](https://doi.org/10.1101/2021.07.21.453294); Saurabh Kalikar, Chirag Jain, Vasimuddin Md, Sanchit Misra; BioRxiv 2021
+
+---
+The original README content of minimap2 follows.
+
+
 [![GitHub Downloads](https://img.shields.io/github/downloads/lh3/minimap2/total.svg?style=social&logo=github&label=Download)](https://github.com/lh3/minimap2/releases)
 [![BioConda Install](https://img.shields.io/conda/dn/bioconda/minimap2.svg?style=flag&label=BioConda%20install)](https://anaconda.org/bioconda/minimap2)
 [![PyPI](https://img.shields.io/pypi/v/mappy.svg?style=flat)](https://pypi.python.org/pypi/mappy)
@@ -74,8 +147,8 @@ Detailed evaluations are available from the [minimap2 paper][doi] or the
 Minimap2 is optimized for x86-64 CPUs. You can acquire precompiled binaries from
 the [release page][release] with:
 ```sh
-curl -L https://github.com/lh3/minimap2/releases/download/v2.24/minimap2-2.24_x64-linux.tar.bz2 | tar -jxvf -
-./minimap2-2.24_x64-linux/minimap2
+curl -L https://github.com/lh3/minimap2/releases/download/v2.22/minimap2-2.22_x64-linux.tar.bz2 | tar -jxvf -
+./minimap2-2.22_x64-linux/minimap2
 ```
 If you want to compile from the source, you need to have a C compiler, GNU make
 and zlib development files installed. Then type `make` in the source code

align.c

@@ -5,6 +5,13 @@
 #include "minimap.h"
 #include "mmpriv.h"
 #include "ksw2.h"
+#include "ksw2_extd2_avx.h"
+#include <x86intrin.h>
+extern uint64_t avg;
+extern uint64_t alignment_time;
+extern void *km1;
+extern uint64_t km_size;// = 500000000; // 500 MB
+extern int km_top;
 
 static void ksw_gen_simple_mat(int m, int8_t *mat, int8_t a, int8_t b, int8_t sc_ambi)
 {
@@ -313,6 +320,7 @@ static void mm_append_cigar(mm_reg1_t *r, uint32_t n_cigar, uint32_t *cigar) //
 	}
 }
 
+#if 0
 static void mm_align_pair(void *km, const mm_mapopt_t *opt, int qlen, const uint8_t *qseq, int tlen, const uint8_t *tseq, const uint8_t *junc, const int8_t *mat, int w, int end_bonus, int zdrop, int flag, ksw_extz_t *ez)
 {
 	if (mm_dbg_flag & MM_DBG_PRINT_ALN_SEQ) {
@@ -340,7 +348,65 @@ static void mm_align_pair(void *km, const mm_mapopt_t *opt, int qlen, const uint
 		fprintf(stderr, "\n");
 	}
 }
+#endif
+#if 1
+static void mm_align_pair(void *km, const mm_mapopt_t *opt, int qlen, const uint8_t *qseq, int tlen, const uint8_t *tseq, const uint8_t *junc, const int8_t *mat, int w, int end_bonus, int zdrop, int flag, ksw_extz_t *ez)
+{
+#ifdef MANUAL_PROFILING
+	uint64_t align_start = __rdtsc();
+#endif
 
+	if (mm_dbg_flag & MM_DBG_PRINT_ALN_SEQ) {
+		int i;
+		fprintf(stderr, "===> q=(%d,%d), e=(%d,%d), bw=%d, flag=%d, zdrop=%d <===\n", opt->q, opt->q2, opt->e, opt->e2, w, flag, opt->zdrop);
+		for (i = 0; i < tlen; ++i) fputc("ACGTN"[tseq[i]], stderr);
+		fputc('\n', stderr);
+		for (i = 0; i < qlen; ++i) fputc("ACGTN"[qseq[i]], stderr);
+		fputc('\n', stderr);
+	}
+	if (opt->max_sw_mat > 0 && (int64_t)tlen * qlen > opt->max_sw_mat) {
+		ksw_reset_extz(ez);
+		ez->zdropped = 1;
+	} else if (opt->flag & MM_F_SPLICE)
+		ksw_exts2_sse(km, qlen, qseq, tlen, tseq, 5, mat, opt->q, opt->e, opt->q2, opt->noncan, zdrop, opt->junc_bonus, flag, junc, ez);
+	else if (opt->q == opt->q2 && opt->e == opt->e2)
+		ksw_extz2_sse(km, qlen, qseq, tlen, tseq, 5, mat, opt->q, opt->e, w, zdrop, end_bonus, flag, ez);
+	else{
+#if defined (ALIGN_AVX) && (defined(__AVX512BW__) || (defined(__AVX2__) && defined(APPLY_AVX2)))
+#ifdef __AVX512BW__
+		
+            	ksw_extd2_avx512(km, qlen, qseq, tlen, tseq, 5, mat, opt->q, opt->e, opt->q2, opt->e2, w, zdrop, end_bonus, flag, ez);
+#elif __AVX2__
+	avg = 0;
+//	uint64_t *ptr_km = (uint64_t *) km1;
+//	for(uint64_t itr = 0; itr < km_size/512; itr++){
+//		avg+=ptr_km[itr];
+//	}
+
+//#ifdef MANUAL_PROFILING
+//	uint64_t align_start = __rdtsc();
+//#endif
+		ksw_extd2_avx2(km, qlen, qseq, tlen, tseq, 5, mat, opt->q, opt->e, opt->q2, opt->e2, w, zdrop, end_bonus, flag, ez);
+//#ifdef MANUAL_PROFILING
+//	alignment_time += (__rdtsc() - align_start);
+//#endif
+#endif
+#else
+		ksw_extd2_sse(km, qlen, qseq, tlen, tseq, 5, mat, opt->q, opt->e, opt->q2, opt->e2, w, zdrop, end_bonus, flag, ez);
+#endif
+	   }
+	if (mm_dbg_flag & MM_DBG_PRINT_ALN_SEQ) {
+		int i;
+		fprintf(stderr, "score=%d, cigar=", ez->score);
+		for (i = 0; i < ez->n_cigar; ++i)
+			fprintf(stderr, "%d%c", ez->cigar[i]>>4, "MIDN"[ez->cigar[i]&0xf]);
+		fprintf(stderr, "\n");
+	}
+#ifdef MANUAL_PROFILING
+	alignment_time += (__rdtsc() - align_start);
+#endif
+}
+#endif
 static inline int mm_get_hplen_back(const mm_idx_t *mi, uint32_t rid, uint32_t x)
 {
 	int64_t i, off0 = mi->seq[rid].offset, off = off0 + x;

build_rmi.sh

@@ -0,0 +1,16 @@
+ref_data=$1
+preset=$2
+
+make clean && make lhash_index=1
+touch temp_read.fastq
+./minimap2  -ax $2 $1 temp_read.fastq >/dev/null      
+
+kv_file=$1"_"$2"_minimizers_key_value_sorted"  
+
+full_path=`readlink -f $kv_file`
+
+cd ./ext/TAL
+make lisa_hash
+./build-lisa-hash-index $full_path
+
+rm ../../temp_read.fastq

cookbook.md

@@ -31,8 +31,8 @@ To acquire the data used in this cookbook and to install minimap2 and paftools,
 please follow the command lines below:
 ```sh
 # install minimap2 executables
-curl -L https://github.com/lh3/minimap2/releases/download/v2.24/minimap2-2.24_x64-linux.tar.bz2 | tar jxf -
-cp minimap2-2.24_x64-linux/{minimap2,k8,paftools.js} .  # copy executables
+curl -L https://github.com/lh3/minimap2/releases/download/v2.22/minimap2-2.22_x64-linux.tar.bz2 | tar jxf -
+cp minimap2-2.22_x64-linux/{minimap2,k8,paftools.js} .  # copy executables
 export PATH="$PATH:"`pwd`                               # put the current directory on PATH
 # download example datasets
 curl -L https://github.com/lh3/minimap2/releases/download/v2.10/cookbook-data.tgz | tar zxf -

hit.c

@@ -252,7 +252,7 @@ void mm_sync_regs(void *km, int n_regs, mm_reg1_t *regs) // keep mm_reg1_t::{id,
 	mm_set_sam_pri(n_regs, regs);
 }
 
-void mm_select_sub(void *km, float pri_ratio, int min_diff, int best_n, int check_strand, int min_strand_sc, int *n_, mm_reg1_t *r)
+void mm_select_sub(void *km, float pri_ratio, int min_diff, int best_n, int *n_, mm_reg1_t *r)
 {
 	if (pri_ratio > 0.0f && *n_ > 0) {
 		int i, k, n = *n_, n_2nd = 0;
@@ -264,9 +264,6 @@ void mm_select_sub(void *km, float pri_ratio, int min_diff, int best_n, int chec
 				if (!(r[i].qs == r[p].qs && r[i].qe == r[p].qe && r[i].rid == r[p].rid && r[i].rs == r[p].rs && r[i].re == r[p].re)) // not identical hits
 					r[k++] = r[i], ++n_2nd;
 				else if (r[i].p) free(r[i].p);
-			} else if (check_strand && n_2nd < best_n && r[i].score > min_strand_sc && r[i].rev != r[p].rev) {
-				r[i].strand_retained = 1;
-				r[k++] = r[i], ++n_2nd;
 			} else if (r[i].p) free(r[i].p);
 		}
 		if (k != n) mm_sync_regs(km, k, r); // removing hits requires sync()
@@ -274,19 +271,6 @@ void mm_select_sub(void *km, float pri_ratio, int min_diff, int best_n, int chec
 	}
 }
 
-int mm_filter_strand_retained(int n_regs, mm_reg1_t *r)
-{
-	int i, k;
-	for (i = k = 0; i < n_regs; ++i) {
-		int p = r[i].parent;
-		if (!r[i].strand_retained || r[i].div < r[p].div * 5.0f || r[i].div < 0.01f) {
-			if (k < i) r[k++] = r[i];
-			else ++k;
-		}
-	}
-	return k;
-}
-
 void mm_filter_regs(const mm_mapopt_t *opt, int qlen, int *n_regs, mm_reg1_t *regs)
 { // NB: after this call, mm_reg1_t::parent can be -1 if its parent filtered out
 	int i, k;

index.c

@@ -14,6 +14,21 @@
 #include "mmpriv.h"
 #include "kvec.h"
 #include "khash.h"
+#include <map>
+#include <fstream>
+#include <vector>
+#include <algorithm>
+#include <x86intrin.h>
+using namespace std;
+
+
+extern uint64_t minimizer_lookup_time, alignment_time, dp_time, rmq_time, rmq_t1, rmq_t2, rmq_t3, rmq_t4;
+#ifdef LISA_HASH
+#include "lisa_hash.h"
+extern lisa_hash<uint64_t, uint64_t> *lh;
+#endif
+
+
 
 #define idx_hash(a) ((a)>>1)
 #define idx_eq(a, b) ((a)>>1 == (b)>>1)
@@ -52,9 +67,43 @@ mm_idx_t *mm_idx_init(int w, int k, int b, int flag)
 	if (!(mm_dbg_flag & 1)) mi->km = km_init();
 	return mi;
 }
+void mm_idx_destroy_mm_hash(mm_idx_t *mi)
+{
+	//fprintf(stderr, "mm_destroy_hash\n");
+	uint32_t i;
+	if (mi == 0) return;
+	if (mi->h) kh_destroy(str, (khash_t(str)*)mi->h);
+	if (mi->B) {
+		for (i = 0; i < 1U<<mi->b; ++i) {
+			free(mi->B[i].p);
+			free(mi->B[i].a.a);
+			kh_destroy(idx, (idxhash_t*)mi->B[i].h);
+		}
+	}
+}
+void mm_idx_destroy_seq(mm_idx_t *mi)
+{
+	//fprintf(stderr, "mm_destroy_seq\n");
+
+	uint32_t i;
+	if (mi == 0) return;
+	if (mi->I) {
+		for (i = 0; i < mi->n_seq; ++i)
+			free(mi->I[i].a);
+		free(mi->I);
+	}
+	if (!mi->km) {
+		for (i = 0; i < mi->n_seq; ++i)
+			free(mi->seq[i].name);
+		free(mi->seq);
+	} else km_destroy(mi->km);
+	free(mi->B); free(mi->S); free(mi);
+}
+
 
 void mm_idx_destroy(mm_idx_t *mi)
 {
+
 	uint32_t i;
 	if (mi == 0) return;
 	if (mi->h) kh_destroy(str, (khash_t(str)*)mi->h);
@@ -96,6 +145,317 @@ const uint64_t *mm_idx_get(const mm_idx_t *mi, uint64_t minier, int *n)
 		return &b->p[kh_val(h, k)>>32];
 	}
 }
+//Output minimap2's hash table entries
+class hash_entry {
+	public: 
+	uint64_t key;
+	uint64_t n;
+	uint64_t *p;
+	hash_entry(uint64_t k, uint64_t n_, uint64_t *p_){
+		key = k;
+		n = n_;
+		p = p_;
+	}
+
+};
+bool key_sort( hash_entry i1, hash_entry i2)
+{
+    return (i1.key < i2.key);
+} 
+
+#if 0
+void mm_idx_load_key_value_lisa(const char* f_name, const mm_idx_t *mi)  
+{
+	uint64_t tic = __rdtsc();
+	std::vector<hash_entry> v_hash;
+
+	//ofstream f(f_name);
+	fprintf(stderr, "Building sorted key-val map\n");
+
+	uint32_t i,j;
+	uint64_t num_values = 0;
+	for (i = 0; i < 1U<<mi->b; ++i) {
+		
+		
+		//fprintf(stderr, "BucketID %lu \n", i);
+		idxhash_t *h = (idxhash_t*)mi->B[i].h;
+		khint_t k;
+		if (h == 0) continue;
+		for (k = 0; k < kh_end(h); ++k){
+			if (kh_exist(h, k)) {
+				uint64_t key = kh_key(h, k), bucket_id = i;
+				key = key>>1;
+				
+				key = key<<mi->b | bucket_id;
+				
+				if(kh_key(h, k)&1)
+				{
+					//print key value
+					//fprintf(stderr, "%llu %llu %llu\n", key, kh_val(h, k), 0);
+					v_hash.push_back(hash_entry(key, kh_val(h, k), NULL));
+				}
+				else
+				{	// print key
+					uint32_t n = (uint32_t)kh_val(h, k);
+					//fprintf(stderr, "%llu %llu %llu ", key, kh_val(h, k), n);
+					// for 0 to lsb 32 val
+					//  print b->p[msb 32 of val]
+					
+					v_hash.push_back(hash_entry(key, n, &mi->B[i].p[(kh_val(h, k)>>32) + 0]));
+				}
+			}
+		}
+	}
+	sort(v_hash.begin(), v_hash.end(), key_sort);
+	fprintf(stderr, "Sorted map building time = %lld \n", __rdtsc() - tic);
+	fprintf(stderr, "Storing hash to %s \n", f_name);
+	tic = __rdtsc();
+
+	int64_t itr_p = 0;
+	for( int i = 0; i < v_hash.size(); i++){
+	
+		if(v_hash[i].p == NULL){	
+			//f<<v_hash[i].key << " "<<1<<"\n"<<v_hash[i].n<<" \n";
+			
+			lh->p[itr_p++] = v_hash[i].n;
+			continue;
+		}
+		
+		//f<<v_hash[i].key << " "<<v_hash[i].n<<endl;
+		
+		for(int j = 0; j < v_hash[i].n; j++){
+		//	f<<v_hash[i].p[j]<<" ";
+			lh->p[itr_p++] = v_hash[i].p[j];
+			num_values++;
+		}
+		//f<<endl;
+	
+	}
+
+	//f.close();	
+
+	string size_file_name = (string) f_name + "_size";
+	ofstream size_f(size_file_name);
+	size_f<<v_hash.size()<<" "<<num_values;
+	size_f.close();
+
+	string prefix = (string)f_name + "_keys";	
+	string keys_bin_file_name = prefix + ".uint64";		
+	ofstream wf(keys_bin_file_name, ios::out | ios::binary);
+	wf.write((char*)&key_list[0], (key_list.size())*sizeof(uint64_t));		
+	wf.close();	
+
+	key_list.clear();
+
+	m.clear();
+	v_hash.clear();
+
+	fprintf(stderr, "Index store File IO time %lld \n", __rdtsc() - tic);
+
+}
+#endif
+
+void mm_idx_dump_hash(const char* f_name, const mm_idx_t *mi)  
+{
+	uint64_t tic = __rdtsc();
+	//std::map<uint64_t, vector<uint64_t>> m;
+	std::vector<hash_entry> v_hash;
+
+	//ofstream f(f_name);
+	fprintf(stderr, "Building sorted key-val map\n");
+
+	uint32_t i,j;
+	uint64_t num_values = 0;
+	for (i = 0; i < 1U<<mi->b; ++i) {
+		
+		
+		//fprintf(stderr, "BucketID %lu \n", i);
+		idxhash_t *h = (idxhash_t*)mi->B[i].h;
+		khint_t k;
+		if (h == 0) continue;
+		for (k = 0; k < kh_end(h); ++k){
+			if (kh_exist(h, k)) {
+				uint64_t key = kh_key(h, k), bucket_id = i;
+				key = key>>1;
+				
+				key = key<<mi->b | bucket_id;
+				
+				if(kh_key(h, k)&1)
+				{
+					//print key value
+					//fprintf(stderr, "%llu %llu %llu\n", key, kh_val(h, k), 0);
+					//m[key].push_back(kh_val(h, k));
+					v_hash.push_back(hash_entry(key, kh_val(h, k), NULL));
+				}
+				else
+				{	// print key
+					uint32_t n = (uint32_t)kh_val(h, k);
+					//fprintf(stderr, "%llu %llu %llu ", key, kh_val(h, k), n);
+					// for 0 to lsb 32 val
+					//  print b->p[msb 32 of val]
+					
+					v_hash.push_back(hash_entry(key, n, &mi->B[i].p[(kh_val(h, k)>>32) + 0]));
+				}
+			}
+		}
+	}
+	sort(v_hash.begin(), v_hash.end(), key_sort);
+	fprintf(stderr, "Sorted map building time = %lld \n", __rdtsc() - tic);
+	fprintf(stderr, "Storing hash to %s \n", f_name);
+	tic = __rdtsc();
+
+	vector<uint64_t> key_list;
+	vector<uint64_t> val_list;
+	vector<uint64_t> p_list;
+/*
+	key_list.push_back(m.size());	
+	for(auto k : m){
+		key_list.push_back(k.first);
+		f<<k.first << " "<<k.second.size()<<endl;
+		for(int j = 0; j < k.second.size(); j++){
+			f<<k.second[j]<<" ";
+			num_values++;
+		}
+		f<<endl;
+	}
+*/
+
+
+	key_list.push_back(v_hash.size());
+	int64_t itr_p = 0;
+	uint64_t sum_pos = 0;
+	string f1_name = (string)f_name + "_pos_bin";
+	string f2_name = (string)f_name + "_val_bin";
+	ofstream f1(f1_name, ios::out | ios::binary);
+	ofstream f2(f2_name, ios::out | ios::binary);
+	for( int i = 0; i < v_hash.size(); i++){
+		key_list.push_back(v_hash[i].key);
+		if(v_hash[i].p == NULL){	
+			//f<<v_hash[i].key << " "<<1<<"\n"<<v_hash[i].n<<" \n";
+			val_list.push_back(sum_pos<<32|(uint64_t)1);
+			sum_pos+=1;
+			p_list.push_back(v_hash[i].n);
+			num_values++;
+			continue;
+		}
+		
+		//f<<v_hash[i].key << " "<<v_hash[i].n<<endl;
+		val_list.push_back(sum_pos<<32|(uint64_t)v_hash[i].n);
+		sum_pos+=v_hash[i].n;	
+			
+		num_values+=v_hash[i].n;
+
+
+		for(int j = 0; j < v_hash[i].n; j++){
+			//f<<v_hash[i].p[j]<<" ";
+			p_list.push_back(v_hash[i].p[j]);
+		}
+//		f<<endl;
+	
+	}
+	f1.write((char*)&val_list[0], (val_list.size())*sizeof(uint64_t));		
+	f2.write((char*)&p_list[0], (p_list.size())*sizeof(uint64_t));		
+	f1.close();
+	f2.close();
+	fprintf(stderr, "Index sorted SoA time %lld \n", __rdtsc() - tic);
+
+	//f.close();	
+
+	string size_file_name = (string) f_name + "_size";
+	ofstream size_f(size_file_name);
+	size_f<<v_hash.size()<<" "<<num_values;
+	size_f.close();
+
+	string prefix = (string)f_name + "_keys";	
+	string keys_bin_file_name = prefix + ".uint64";		
+	ofstream wf(keys_bin_file_name, ios::out | ios::binary);
+	wf.write((char*)&key_list[0], (key_list.size())*sizeof(uint64_t));		
+	wf.close();	
+
+	key_list.clear();
+
+	//m.clear();
+	v_hash.clear();
+
+	fprintf(stderr, "Index store File IO time %lld \n", __rdtsc() - tic);
+
+}
+void mm_idx_dump_hash_1(const char* f_name, const mm_idx_t *mi)  
+{
+	uint64_t tic = __rdtsc();
+	std::map<uint64_t, vector<uint64_t>> m;
+
+	ofstream f(f_name);
+	fprintf(stderr, "Building sorted key-val map\n");
+
+	uint32_t i,j;
+	uint64_t num_values = 0;
+	for (i = 0; i < 1U<<mi->b; ++i) {
+		
+		
+		//fprintf(stderr, "BucketID %lu \n", i);
+		idxhash_t *h = (idxhash_t*)mi->B[i].h;
+		khint_t k;
+		if (h == 0) continue;
+		for (k = 0; k < kh_end(h); ++k){
+			if (kh_exist(h, k)) {
+				uint64_t key = kh_key(h, k), bucket_id = i;
+				key = key>>1;
+				
+				key = key<<mi->b | bucket_id;
+				
+				if(kh_key(h, k)&1)
+				{
+					//print key value
+					//fprintf(stderr, "%llu %llu %llu\n", key, kh_val(h, k), 0);
+					m[key].push_back(kh_val(h, k));
+				}
+				else
+				{	// print key
+					uint32_t n = (uint32_t)kh_val(h, k);
+					//fprintf(stderr, "%llu %llu %llu ", key, kh_val(h, k), n);
+					// for 0 to lsb 32 val
+					//  print b->p[msb 32 of val]
+					for(j = 0; j < n; j++)
+					{
+						//fprintf(stderr, "%llu ", mi->B[i].p[(kh_val(h, k)>>32) + j]);
+						m[key].push_back(mi->B[i].p[(kh_val(h, k)>>32) + j]);
+					}
+				}
+			}
+		}
+	}
+	fprintf(stderr, "Sorted map building time = %lld \n", __rdtsc() - tic);
+	fprintf(stderr, "Storing hash to %s \n", f_name);
+	tic = __rdtsc();
+	vector<uint64_t> key_list;
+	key_list.push_back(m.size());	
+	for(auto k : m){
+		key_list.push_back(k.first);
+		f<<k.first << " "<<k.second.size()<<endl;
+		for(int j = 0; j < k.second.size(); j++){
+			f<<k.second[j]<<" ";
+			num_values++;
+		}
+		f<<endl;
+	}
+	f.close();	
+	string size_file_name = (string) f_name + "_size";
+	ofstream size_f(size_file_name);
+	size_f<<m.size()<<" "<<num_values;
+	size_f.close();
+
+	string prefix = (string)f_name + "_keys";	
+	string keys_bin_file_name = prefix + ".uint64";		
+	ofstream wf(keys_bin_file_name, ios::out | ios::binary);
+	wf.write((char*)&key_list[0], (key_list.size())*sizeof(uint64_t));		
+	wf.close();	
+
+	key_list.clear();
+	m.clear();
+	fprintf(stderr, "Index store File IO time %lld \n", __rdtsc() - tic);
+
+}
 
 void mm_idx_stat(const mm_idx_t *mi)
 {
@@ -119,6 +479,7 @@ void mm_idx_stat(const mm_idx_t *mi)
 	}
 	fprintf(stderr, "[M::%s::%.3f*%.2f] distinct minimizers: %d (%.2f%% are singletons); average occurrences: %.3lf; average spacing: %.3lf; total length: %ld\n",
 			__func__, realtime() - mm_realtime0, cputime() / (realtime() - mm_realtime0), n, 100.0*n1/n, (double)sum / n, (double)len / sum, (long)len);
+	fprintf(stderr, "minimizer-lookup: %lld dp: %lld rmq: %lld rmq_t1: %lld rmq_t2: %lld rmq_t3: %lld rmq_t4: %lld alignment: %lld \n", minimizer_lookup_time, dp_time, rmq_time, rmq_t1, rmq_t2, rmq_t3, rmq_t4, alignment_time);
 }
 
 int mm_idx_index_name(mm_idx_t *mi)

ksw2_extd2_avx.h

@@ -0,0 +1,42 @@
+/* The MIT License
+
+Copyright (c) 2018-     Dana-Farber Cancer Institute
+              2017-2018 Broad Institute, Inc.
+
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of this software and associated documentation files (the
+"Software"), to deal in the Software without restriction, including
+without limitation the rights to use, copy, modify, merge, publish,
+distribute, sublicense, and/or sell copies of the Software, and to
+permit persons to whom the Software is furnished to do so, subject to
+the following conditions:
+
+The above copyright notice and this permission notice shall be
+included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+Modified Copyright (C) 2021 Intel Corporation
+   Contacts: Saurabh Kalikar <saurabh.kalikar@intel.com>; 
+	Vasimuddin Md <vasimuddin.md@intel.com>; Sanchit Misra <sanchit.misra@intel.com>; 
+	Chirag Jain <chirag@iisc.ac.in>; Heng Li <hli@jimmy.harvard.edu>
+*/
+#include <string.h>
+#include <stdio.h>
+#include <assert.h>
+#include "ksw2.h"
+#include <immintrin.h>
+#include <x86intrin.h>
+#include <smmintrin.h>
+#include <emmintrin.h>
+void ksw_extd2_avx512(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 e2, int w, int zdrop, int end_bonus, int flag, ksw_extz_t *ez);
+
+void ksw_extd2_avx2(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 e2, int w, int zdrop, int end_bonus, int flag, ksw_extz_t *ez);

lchain.c

@@ -5,26 +5,17 @@
 #include "mmpriv.h"
 #include "kalloc.h"
 #include "krmq.h"
+#include <x86intrin.h>
+//#include "simd_chain.h"
+//#include "parallel_chaining_32_bit.h"
+#include "parallel_chaining_v2_22.h"
 
-static int64_t mg_chain_bk_end(int32_t max_drop, const mm128_t *z, const int32_t *f, const int64_t *p, int32_t *t, int64_t k)
-{
-	int64_t i = z[k].y, end_i = -1, max_i = i;
-	int32_t max_s = 0;
-	if (i < 0 || t[i] != 0) return i;
-	do {
-		int32_t s;
-		t[i] = 2;
-		end_i = i = p[i];
-		s = i < 0? z[k].x : (int32_t)z[k].x - f[i];
-		if (s > max_s) max_s = s, max_i = i;
-		else if (max_s - s > max_drop) break;
-	} while (i >= 0 && t[i] == 0);
-	for (i = z[k].y; i >= 0 && i != end_i; i = p[i]) // reset modified t[]
-		t[i] = 0;
-	return max_i;
-}
+#ifdef MANUAL_PROFILING
+extern uint64_t dp_time, rmq_time, rmq_t1, rmq_t2, rmq_t3, rmq_t4;
+#endif
 
-uint64_t *mg_chain_backtrack(void *km, int64_t n, const int32_t *f, const int64_t *p, int32_t *v, int32_t *t, int32_t min_cnt, int32_t min_sc, int32_t max_drop, int32_t *n_u_, int32_t *n_v_)
+extern bool enable_vect_dp_chaining;
+uint64_t *mg_chain_backtrack(void *km, int64_t n, const int32_t *f, const int64_t *p, int32_t *v, int32_t *t, int32_t min_cnt, int32_t min_sc, int32_t *n_u_, int32_t *n_v_)
 {
 	mm128_t *z;
 	uint64_t *u;
@@ -42,32 +33,26 @@ uint64_t *mg_chain_backtrack(void *km, int64_t n, const int32_t *f, const int64_
 
 	memset(t, 0, n * 4);
 	for (k = n_z - 1, n_v = n_u = 0; k >= 0; --k) { // precompute n_u
-		if (t[z[k].y] == 0) {
-			int64_t n_v0 = n_v, end_i;
-			int32_t sc;
-			end_i = mg_chain_bk_end(max_drop, z, f, p, t, k);
-			for (i = z[k].y; i != end_i; i = p[i])
-				++n_v, t[i] = 1;
-			sc = i < 0? z[k].x : (int32_t)z[k].x - f[i];
-			if (sc >= min_sc && n_v > n_v0 && n_v - n_v0 >= min_cnt)
-				++n_u;
-			else n_v = n_v0;
-		}
+		int64_t n_v0 = n_v;
+		int32_t sc;
+		for (i = z[k].y; i >= 0 && t[i] == 0; i = p[i])
+			++n_v, t[i] = 1;
+		sc = i < 0? z[k].x : (int32_t)z[k].x - f[i];
+		if (sc >= min_sc && n_v > n_v0 && n_v - n_v0 >= min_cnt)
+			++n_u;
+		else n_v = n_v0;
 	}
 	KMALLOC(km, u, n_u);
 	memset(t, 0, n * 4);
 	for (k = n_z - 1, n_v = n_u = 0; k >= 0; --k) { // populate u[]
-		if (t[z[k].y] == 0) {
-			int64_t n_v0 = n_v, end_i;
-			int32_t sc;
-			end_i = mg_chain_bk_end(max_drop, z, f, p, t, k);
-			for (i = z[k].y; i != end_i; i = p[i])
-				v[n_v++] = i, t[i] = 1;
-			sc = i < 0? z[k].x : (int32_t)z[k].x - f[i];
-			if (sc >= min_sc && n_v > n_v0 && n_v - n_v0 >= min_cnt)
-				u[n_u++] = (uint64_t)sc << 32 | (n_v - n_v0);
-			else n_v = n_v0;
-		}
+		int64_t n_v0 = n_v;
+		int32_t sc;
+		for (i = z[k].y; i >= 0 && t[i] == 0; i = p[i])
+			v[n_v++] = i, t[i] = 1;
+		sc = i < 0? z[k].x : (int32_t)z[k].x - f[i];
+		if (sc >= min_sc && n_v > n_v0 && n_v - n_v0 >= min_cnt)
+			u[n_u++] = (uint64_t)sc << 32 | (n_v - n_v0);
+		else n_v = n_v0;
 	}
 	kfree(km, z);
 	assert(n_v < INT32_MAX);
@@ -112,15 +97,65 @@ static mm128_t *compact_a(void *km, int32_t n_u, uint64_t *u, int32_t n_v, int32
 
 static inline int32_t comput_sc(const mm128_t *ai, const mm128_t *aj, int32_t max_dist_x, int32_t max_dist_y, int32_t bw, float chn_pen_gap, float chn_pen_skip, int is_cdna, int n_seg)
 {
-	int32_t dq = (int32_t)ai->y - (int32_t)aj->y, dr, dd, dg, q_span, sc;
-	int32_t sidi = (ai->y & MM_SEED_SEG_MASK) >> MM_SEED_SEG_SHIFT;
-	int32_t sidj = (aj->y & MM_SEED_SEG_MASK) >> MM_SEED_SEG_SHIFT;
-	if (dq <= 0 || dq > max_dist_x) return INT32_MIN;
-	dr = (int32_t)(ai->x - aj->x);
-	if (sidi == sidj && (dr == 0 || dq > max_dist_y)) return INT32_MIN;
+
+	uint64_t ai_x, ai_y, aj_x, aj_y;
+	ai_x = ai->x; ai_y = ai->y; aj_x = aj->x; aj_y = aj->y;
+
+#ifdef CHAIN_DEBUG
+	int32_t sc_vect = obj.comput_sc_vectorized_avx2_caller(ai_x, ai_y, aj_x, aj_y, aj->y>>32&0xff);
+#endif
+
+	//if (sc_vect == 0) return INT32_MIN;
+	//else 
+	//return sc_vect;
+
+	//fprintf(stderr, "%lld %lld %lld %lld \n", ai_x, ai_y, aj_x, aj_y); 
+	//fprintf(stderr, "%lld %lld %lld %f %f %d %d\n", max_dist_x, max_dist_y, bw, chn_pen_gap, chn_pen_skip, is_cdna, n_seg);
+	int32_t dq = (int32_t)ai_y - (int32_t)aj_y, dr, dd, dg, q_span, sc;
+	int32_t sidi = (ai_y & MM_SEED_SEG_MASK) >> MM_SEED_SEG_SHIFT;
+	int32_t sidj = (aj_y & MM_SEED_SEG_MASK) >> MM_SEED_SEG_SHIFT;
+	if (dq <= 0 || dq > max_dist_x) { 
+
+#ifdef CHAIN_DEBUG
+		if(INT32_MIN != sc_vect){
+		//fprintf(stderr, "score mismatch %d -- %d", sc , sc_vect);
+			fprintf(stderr, "int-min exit: %llu, %llu, %llu, %llu : %d -- %d\n", ai_x, ai_y, aj_x, aj_y, sc, sc_vect);
+		}
+#endif
+	  return INT32_MIN; 
+	}
+	dr = (int32_t)(ai_x - aj_x);
+	if (sidi == sidj && (dr == 0 || dq > max_dist_y)) { 
+
+#ifdef CHAIN_DEBUG
+		if(INT32_MIN != sc_vect){
+		//fprintf(stderr, "score mismatch %d -- %d", sc , sc_vect);
+			fprintf(stderr, "int-min exit: %llu, %llu, %llu, %llu : %d -- %d\n", ai_x, ai_y, aj_x, aj_y, sc, sc_vect);
+		}
+#endif
+		return INT32_MIN;
+	}
 	dd = dr > dq? dr - dq : dq - dr;
-	if (sidi == sidj && dd > bw) return INT32_MIN;
-	if (n_seg > 1 && !is_cdna && sidi == sidj && dr > max_dist_y) return INT32_MIN;
+	if (sidi == sidj && dd > bw) {
+
+#ifdef CHAIN_DEBUG
+		if(INT32_MIN != sc_vect){
+		//fprintf(stderr, "score mismatch %d -- %d", sc , sc_vect);
+			fprintf(stderr, "int-min exit: %llu, %llu, %llu, %llu : %d -- %d\n", ai_x, ai_y, aj_x, aj_y, sc, sc_vect);
+		}
+#endif
+		return INT32_MIN; 
+	}
+	if (n_seg > 1 && !is_cdna && sidi == sidj && dr > max_dist_y) { 
+
+#ifdef CHAIN_DEBUG
+		if(INT32_MIN != sc_vect){
+		//fprintf(stderr, "score mismatch %d -- %d", sc , sc_vect);
+			fprintf(stderr, "int-min exit: %llu, %llu, %llu, %llu : %d -- %d\n", ai_x, ai_y, aj_x, aj_y, sc, sc_vect);
+		}
+#endif
+		return INT32_MIN;
+	}
 	dg = dr < dq? dr : dq;
 	q_span = aj->y>>32&0xff;
 	sc = q_span < dg? q_span : dg;
@@ -134,6 +169,13 @@ static inline int32_t comput_sc(const mm128_t *ai, const mm128_t *aj, int32_t ma
 			else sc -= (int)(lin_pen + .5f * log_pen);
 		} else sc -= (int)(lin_pen + .5f * log_pen);
 	}
+#ifdef CHAIN_DEBUG
+
+	if(sc != sc_vect ){
+		//fprintf(stderr, "score mismatch %d -- %d", sc , sc_vect);
+		fprintf(stderr, "outer: %llu, %llu, %llu, %llu : %d -- %d\n", ai_x, ai_y, aj_x, aj_y, sc, sc_vect);
+	}
+#endif
 	return sc;
 }
 
@@ -148,10 +190,18 @@ static inline int32_t comput_sc(const mm128_t *ai, const mm128_t *aj, int32_t ma
 mm128_t *mg_lchain_dp(int max_dist_x, int max_dist_y, int bw, int max_skip, int max_iter, int min_cnt, int min_sc, float chn_pen_gap, float chn_pen_skip,
 					  int is_cdna, int n_seg, int64_t n, mm128_t *a, int *n_u_, uint64_t **_u, void *km)
 { // TODO: make sure this works when n has more than 32 bits
-	int32_t *f, *t, *v, n_u, n_v, mmax_f = 0, max_drop = bw;
+	///fprintf(stderr, "chaining called\n");
+
+
+
+#ifdef MANUAL_PROFILING
+	uint64_t align_start = __rdtsc();
+#endif
+
+	int32_t *f, *t, *v, *v_1, *p_1, n_u, n_v, mmax_f = 0;
 	int64_t *p, i, j, max_ii, st = 0, n_iter = 0;
 	uint64_t *u;
-
+	uint32_t* f_1;
 	if (_u) *_u = 0, *n_u_ = 0;
 	if (n == 0 || a == 0) {
 		kfree(km, a);
@@ -159,18 +209,57 @@ mm128_t *mg_lchain_dp(int max_dist_x, int max_dist_y, int bw, int max_skip, int
 	}
 	if (max_dist_x < bw) max_dist_x = bw;
 	if (max_dist_y < bw && !is_cdna) max_dist_y = bw;
-	if (is_cdna) max_drop = INT32_MAX;
 	KMALLOC(km, p, n);
+	KMALLOC(km, p_1, n);
 	KMALLOC(km, f, n);
+	KMALLOC(km, f_1, n);
 	KMALLOC(km, v, n);
+	KMALLOC(km, v_1, n);
 	KCALLOC(km, t, n);
 
+//#ifdef PARALLEL_CHAINING
+if(enable_vect_dp_chaining){
+	// Parallel chaining data-structures
+	anchor_t* anchors = (anchor_t*)malloc(n* sizeof(anchor_t));
+	for (i = 0; i < n; ++i) {
+		uint64_t ri = a[i].x;
+		int32_t qi = (int32_t)a[i].y, q_span = a[i].y>>32&0xff; // NB: only 8 bits of span is used!!!
+		anchors[i].r = ri;
+		anchors[i].q = qi;
+		anchors[i].l = q_span;
+	}
+	num_bits_t *anchor_r, *anchor_q, *anchor_l;
+	create_SoA_Anchors_32_bit(anchors, n, anchor_r, anchor_q, anchor_l);
+	dp_chain obj(max_dist_x, max_dist_y, bw, max_skip, max_iter, min_cnt, min_sc, chn_pen_gap, chn_pen_skip, is_cdna, n_seg);
+
+#ifdef PARALLEL_CHAINING
+	obj.mm_dp_vectorized(n, &anchors[0], anchor_r, anchor_q, anchor_l, f_1, p_1, v_1, max_dist_x, max_dist_y, NULL, NULL);
+#endif
+	// -16 is due to extra padding at the start of arrays
+	anchor_r -= 16; anchor_q -= 16; anchor_l -= 16;
+	free(anchor_r); 
+	free(anchor_q); 
+	free(anchor_l);
+	free(anchors);
+	for(int i = 0; i < n; i++){
+#if 1
+			f[i] = f_1[i];
+			p[i] = p_1[i];
+			v[i] = v_1[i];
+#endif
+	}
+
+//
+} else {
+//#else
+
 	// fill the score and backtrack arrays
 	for (i = 0, max_ii = -1; i < n; ++i) {
 		int64_t max_j = -1, end_j;
 		int32_t max_f = a[i].y>>32&0xff, n_skip = 0;
 		while (st < i && (a[i].x>>32 != a[st].x>>32 || a[i].x > a[st].x + max_dist_x)) ++st;
 		if (i - st > max_iter) st = i - max_iter;
+		int my_cnt = 0;
 		for (j = i - 1; j >= st; --j) {
 			int32_t sc;
 			sc = comput_sc(&a[i], &a[j], max_dist_x, max_dist_y, bw, chn_pen_gap, chn_pen_skip, is_cdna, n_seg);
@@ -187,32 +276,62 @@ mm128_t *mg_lchain_dp(int max_dist_x, int max_dist_y, int bw, int max_skip, int
 			if (p[j] >= 0) t[p[j]] = i;
 		}
 		end_j = j;
+		int debug_iter = 2057329;
+
 		if (max_ii < 0 || a[i].x - a[max_ii].x > (int64_t)max_dist_x) {
 			int32_t max = INT32_MIN;
 			max_ii = -1;
-			for (j = i - 1; j >= st; --j)
-				if (max < f[j]) max = f[j], max_ii = j;
+			for (j = i - 1; j >= st; --j) {
+				if (max < (int32_t)f[j]) max = f[j], max_ii = j;
+			}
 		}
 		if (max_ii >= 0 && max_ii < end_j) {
 			int32_t tmp;
 			tmp = comput_sc(&a[i], &a[max_ii], max_dist_x, max_dist_y, bw, chn_pen_gap, chn_pen_skip, is_cdna, n_seg);
-			if (tmp != INT32_MIN && max_f < tmp + f[max_ii])
+		//	if (i == debug_iter) fprintf(stderr, "mm2: endj: %d max_ii: %d max_f: %d tmp_score: %d \n", end_j, max_ii, max_f, tmp);	
+		
+
+			if (tmp != INT32_MIN && max_f < tmp + f[max_ii]){
 				max_f = tmp + f[max_ii], max_j = max_ii;
+			}
 		}
 		f[i] = max_f, p[i] = max_j;
 		v[i] = max_j >= 0 && v[max_j] > max_f? v[max_j] : max_f; // v[] keeps the peak score up to i; f[] is the score ending at i, not always the peak
+
 		if (max_ii < 0 || (a[i].x - a[max_ii].x <= (int64_t)max_dist_x && f[max_ii] < f[i]))
 			max_ii = i;
 		if (mmax_f < max_f) mmax_f = max_f;
 	}
 
-	u = mg_chain_backtrack(km, n, f, p, v, t, min_cnt, min_sc, max_drop, &n_u, &n_v);
+}
+//#endif
+
+#ifdef CHAIN_DEBUG
+
+		for(int i = 0; i < n; i++){
+			if(f[i] != f_1[i] || p[i] != p_1[i] || v[i] !=v_1[i])
+			{
+				fprintf(stderr, "i:%d %d %d %d %d %d %d\n",i, f[i], f_1[i], p[i], p_1[i], v[i], v_1[i] );
+			}
+#if 0
+			f[i] = f_1[i];
+			p[i] = p_1[i];
+			v[i] = v_1[i];
+#endif
+		}
+#endif
+	u = mg_chain_backtrack(km, n, f, p, v, t, min_cnt, min_sc, &n_u, &n_v);
 	*n_u_ = n_u, *_u = u; // NB: note that u[] may not be sorted by score here
-	kfree(km, p); kfree(km, f); kfree(km, t);
+	kfree(km, p); kfree(km, p_1); kfree(km, f); kfree(km, f_1); kfree(km, t); kfree(km, v_1);
 	if (n_u == 0) {
 		kfree(km, a); kfree(km, v);
 		return 0;
 	}
+
+
+#ifdef MANUAL_PROFILING
+	dp_time += __rdtsc() - align_start;
+#endif
 	return compact_a(km, n_u, u, n_v, v, a);
 }
 
@@ -250,7 +369,12 @@ static inline int32_t comput_sc_simple(const mm128_t *ai, const mm128_t *aj, flo
 mm128_t *mg_lchain_rmq(int max_dist, int max_dist_inner, int bw, int max_chn_skip, int cap_rmq_size, int min_cnt, int min_sc, float chn_pen_gap, float chn_pen_skip,
 					   int64_t n, mm128_t *a, int *n_u_, uint64_t **_u, void *km)
 {
-	int32_t *f,*t, *v, n_u, n_v, mmax_f = 0, max_rmq_size = 0, max_drop = bw;
+#ifdef MANUAL_PROFILING
+	uint64_t start = __rdtsc();
+#endif
+	uint64_t tim;
+	//fprintf(stderr, "rmq call \n");
+	int32_t *f,*t, *v, n_u, n_v, mmax_f = 0, max_rmq_size = 0;
 	int64_t *p, i, i0, st = 0, st_inner = 0, n_iter = 0;
 	uint64_t *u;
 	lc_elem_t *root = 0, *root_inner = 0;
@@ -277,6 +401,9 @@ mm128_t *mg_lchain_rmq(int max_dist, int max_dist_inner, int bw, int max_chn_ski
 		int32_t q_span = a[i].y>>32&0xff, max_f = q_span;
 		lc_elem_t s, *q, *r, lo, hi;
 		// add in-range anchors
+#ifdef MANUAL_PROFILING_RMQ
+	tim = __rdtsc();
+#endif
 		if (i0 < i && a[i0].x != a[i].x) {
 			int64_t j;
 			for (j = i0; j < i; ++j) {
@@ -291,7 +418,13 @@ mm128_t *mg_lchain_rmq(int max_dist, int max_dist_inner, int bw, int max_chn_ski
 			}
 			i0 = i;
 		}
+#ifdef MANUAL_PROFILING_RMQ
+	rmq_t1 += __rdtsc() - tim;
+#endif
 		// get rid of active chains out of range
+#ifdef MANUAL_PROFILING_RMQ
+	tim = __rdtsc();
+#endif
 		while (st < i && (a[i].x>>32 != a[st].x>>32 || a[i].x > a[st].x + max_dist || krmq_size(head, root) > cap_rmq_size)) {
 			s.y = (int32_t)a[st].y, s.i = st;
 			if ((q = krmq_find(lc_elem, root, &s, 0)) != 0) {
@@ -300,6 +433,12 @@ mm128_t *mg_lchain_rmq(int max_dist, int max_dist_inner, int bw, int max_chn_ski
 			}
 			++st;
 		}
+#ifdef MANUAL_PROFILING_RMQ
+	rmq_t2 += __rdtsc() - tim;
+#endif
+#ifdef MANUAL_PROFILING_RMQ
+	tim = __rdtsc();
+#endif
 		if (max_dist_inner > 0)  { // similar to the block above, but applied to the inner tree
 			while (st_inner < i && (a[i].x>>32 != a[st_inner].x>>32 || a[i].x > a[st_inner].x + max_dist_inner || krmq_size(head, root_inner) > cap_rmq_size)) {
 				s.y = (int32_t)a[st_inner].y, s.i = st_inner;
@@ -310,6 +449,9 @@ mm128_t *mg_lchain_rmq(int max_dist, int max_dist_inner, int bw, int max_chn_ski
 				++st_inner;
 			}
 		}
+#ifdef MANUAL_PROFILING_RMQ
+	rmq_t3 += __rdtsc() - tim;
+#endif
 		// RMQ
 		lo.i = INT32_MAX, lo.y = (int32_t)a[i].y - max_dist;
 		hi.i = 0, hi.y = (int32_t)a[i].y;
@@ -329,6 +471,9 @@ mm128_t *mg_lchain_rmq(int max_dist, int max_dist_inner, int bw, int max_chn_ski
 					krmq_itr_t(lc_elem) itr;
 					krmq_itr_find(lc_elem, root_inner, lo, &itr);
 					while ((q = krmq_at(&itr)) != 0) {
+#ifdef MANUAL_PROFILING_RMQ
+	tim = __rdtsc();
+#endif
 						if (q->y < (int32_t)a[i].y - max_dist_inner) break;
 						++n_rmq_iter;
 						j = q->i;
@@ -344,11 +489,15 @@ mm128_t *mg_lchain_rmq(int max_dist, int max_dist_inner, int bw, int max_chn_ski
 							if (p[j] >= 0) t[p[j]] = i;
 						}
 						if (!krmq_itr_prev(lc_elem, &itr)) break;
+#ifdef MANUAL_PROFILING_RMQ
+	rmq_t4 += __rdtsc() - tim;
+#endif
 					}
 					n_iter += n_rmq_iter;
 				}
 			}
 		}
+
 		// set max
 		assert(max_j < 0 || (a[max_j].x < a[i].x && (int32_t)a[max_j].y < (int32_t)a[i].y));
 		f[i] = max_f, p[i] = max_j;
@@ -358,12 +507,15 @@ mm128_t *mg_lchain_rmq(int max_dist, int max_dist_inner, int bw, int max_chn_ski
 	}
 	km_destroy(mem_mp);
 
-	u = mg_chain_backtrack(km, n, f, p, v, t, min_cnt, min_sc, max_drop, &n_u, &n_v);
+	u = mg_chain_backtrack(km, n, f, p, v, t, min_cnt, min_sc, &n_u, &n_v);
 	*n_u_ = n_u, *_u = u; // NB: note that u[] may not be sorted by score here
 	kfree(km, p); kfree(km, f); kfree(km, t);
 	if (n_u == 0) {
 		kfree(km, a); kfree(km, v);
 		return 0;
 	}
+#ifdef MANUAL_PROFILING
+	rmq_time += __rdtsc() -  start;
+#endif
 	return compact_a(km, n_u, u, n_v, v, a);
 }

main.c

@@ -6,8 +6,77 @@
 #include "minimap.h"
 #include "mmpriv.h"
 #include "ketopt.h"
+#include <x86intrin.h>
+#include <immintrin.h>
+#include <sys/time.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+#include <string>
+#include <map>
+#include <errno.h>
+#include "bseq.h"
+#include "minimap.h"
+#include "mmpriv.h"
+#include "ketopt.h"
 
-#define MM_VERSION "2.24-r1122"
+//#include "profile.h"
+#include <stdint.h>
+#include <unistd.h>
+#include <x86intrin.h>
+
+using namespace std;
+uint64_t avg;
+uint64_t minimizer_lookup_time, alignment_time, dp_time, rmq_time, rmq_t1, rmq_t2, rmq_t3, rmq_t4;
+
+bool enable_vect_dp_chaining = false;
+
+#ifdef LISA_HASH
+#include "lisa_hash.h"
+lisa_hash<uint64_t, uint64_t> *lh;
+#endif
+
+// New memory allocation approach for alignment optimizations
+//
+void *km1;
+uint64_t km_size = 500000000; // 500 MB
+int km_top;
+/*
+void *kcalloc_(void* km, int count, int size)
+{
+    assert(count*size < km_size);
+    km_top += count*size + 1024;
+    memset(km, 0, count * size);
+        
+    // printf("km_top: %d\n", km_top);
+    return km;
+}
+
+void *kmalloc_(void* km, int count) {
+    if(km_top + count >= km_size)
+        printf("count: %d\n", count);
+    assert(km_top + count < km_size);
+    void *mem = (void*) ((int8_t*) km + km_top);
+    km_top += count + 1024;
+    // printf("km_top: %d\n", km_top);
+    return mem;
+}
+
+void kfree_all() { km_top = 0;}
+  */  
+
+// Memory for alignment end
+
+
+#ifndef __rdtsc
+#ifdef _rdtsc
+#define __rdtsc _rdtsc
+#else
+#define __rdtsc __builtin_ia32_rdtsc
+#endif
+#endif
+
+#define MM_VERSION "2.22-r1101"
 
 #ifdef __linux__
 #include <sys/resource.h>
@@ -74,10 +143,6 @@ static ko_longopt_t long_options[] = {
 	{ "rmq",            ko_optional_argument, 347 },
 	{ "qstrand",        ko_no_argument,       348 },
 	{ "cap-kalloc",     ko_required_argument, 349 },
-	{ "q-occ-frac",     ko_required_argument, 350 },
-	{ "chain-skip-scale",ko_required_argument,351 },
-	{ "print-chains",   ko_no_argument,       352 },
-	{ "no-hash-name",   ko_no_argument,       353 },
 	{ "help",           ko_no_argument,       'h' },
 	{ "max-intron-len", ko_required_argument, 'G' },
 	{ "version",        ko_no_argument,       'V' },
@@ -121,6 +186,12 @@ static inline void yes_or_no(mm_mapopt_t *opt, int64_t flag, int long_idx, const
 
 int main(int argc, char *argv[])
 {
+	// Memory allocation for alignment optimizations
+    //km1 = calloc(km_size, 1); // 10 MB init contg. alloc
+#ifdef PARALLEL_CHAINING
+	enable_vect_dp_chaining = true;
+#endif
+
 	const char *opt_str = "2aSDw:k:K:t:r:f:Vv:g:G:I:d:XT:s:x:Hcp:M:n:z:A:B:O:E:m:N:Qu:R:hF:LC:yYPo:e:U:";
 	ketopt_t o = KETOPT_INIT;
 	mm_mapopt_t opt;
@@ -135,9 +206,11 @@ int main(int argc, char *argv[])
 	liftrlimit();
 	mm_realtime0 = realtime();
 	mm_set_opt(0, &ipt, &opt);
+	string preset_arg = "";
 
 	while ((c = ketopt(&o, argc, argv, 1, opt_str, long_options)) >= 0) { // test command line options and apply option -x/preset first
 		if (c == 'x') {
+			preset_arg +=  (string) o.arg;
 			if (mm_set_opt(o.arg, &ipt, &opt) < 0) {
 				fprintf(stderr, "[ERROR] unknown preset '%s'\n", o.arg);
 				return 1;
@@ -228,15 +301,11 @@ int main(int argc, char *argv[])
 		else if (c == 341) opt.junc_bonus = atoi(o.arg); // --junc-bonus
 		else if (c == 342) opt.flag |= MM_F_SAM_HIT_ONLY; // --sam-hit-only
 		else if (c == 343) opt.chain_gap_scale = atof(o.arg); // --chain-gap-scale
-		else if (c == 351) opt.chain_skip_scale = atof(o.arg); // --chain-skip-scale
 		else if (c == 344) alt_list = o.arg; // --alt
 		else if (c == 345) opt.alt_drop = atof(o.arg); // --alt-drop
 		else if (c == 346) opt.mask_len = mm_parse_num(o.arg); // --mask-len
 		else if (c == 348) opt.flag |= MM_F_QSTRAND | MM_F_NO_INV; // --qstrand
 		else if (c == 349) opt.cap_kalloc = mm_parse_num(o.arg); // --cap-kalloc
-		else if (c == 350) opt.q_occ_frac = atof(o.arg); // --q-occ-frac
-		else if (c == 352) mm_dbg_flag |= MM_DBG_PRINT_CHAIN; // --print-chains
-		else if (c == 353) opt.flag |= MM_F_NO_HASH_NAME; // --no-hash-name
 		else if (c == 330) {
 			fprintf(stderr, "[WARNING] \033[1;31m --lj-min-ratio has been deprecated.\033[0m\n");
 		} else if (c == 314) { // --frag
@@ -374,6 +443,7 @@ int main(int argc, char *argv[])
 		fprintf(stderr, "[ERROR] incorrect input: in the sr mode, please specify no more than two query files.\n");
 		return 1;
 	}
+	preset_arg = (string)argv[o.ind] + "_" + preset_arg + "_minimizers_key_value_sorted";
 	idx_rdr = mm_idx_reader_open(argv[o.ind], &ipt, fnw);
 	if (idx_rdr == 0) {
 		fprintf(stderr, "[ERROR] failed to open file '%s': %s\n", argv[o.ind], strerror(errno));
@@ -416,6 +486,9 @@ int main(int argc, char *argv[])
 					__func__, realtime() - mm_realtime0, cputime() / (realtime() - mm_realtime0), mi->n_seq);
 		if (argc != o.ind + 1) mm_mapopt_update(&opt, mi);
 		if (mm_verbose >= 3) mm_idx_stat(mi);
+#ifdef LISA_INDEX
+		mm_idx_dump_hash(preset_arg.c_str(), mi); 
+#endif	
 		if (junc_bed) mm_idx_bed_read(mi, junc_bed, 1);
 		if (alt_list) mm_idx_alt_read(mi, alt_list);
 		if (argc - (o.ind + 1) == 0) {
@@ -423,6 +496,16 @@ int main(int argc, char *argv[])
 			continue; // no query files
 		}
 		ret = 0;
+#ifdef LISA_HASH
+	fprintf(stderr, "Using LISA_HASH..\n");
+	mm_idx_destroy_mm_hash(mi);
+	char* prefix;
+	lh = new lisa_hash<uint64_t, uint64_t>(preset_arg, prefix);
+	fprintf(stderr, "Loading done.\n");
+//	total_time =  __rdtsc();
+//	fprintf(stderr, "\nIndexing Real time: %.3f sec;\n", realtime() - mapping_time);
+#endif
+	mm_realtime0 = realtime();
 		if (!(opt.flag & MM_F_FRAG_MODE)) {
 			for (i = o.ind + 1; i < argc; ++i) {
 				ret = mm_map_file(mi, argv[i], &opt, n_threads);
@@ -431,12 +514,17 @@ int main(int argc, char *argv[])
 		} else {
 			ret = mm_map_file_frag(mi, argc - (o.ind + 1), (const char**)&argv[o.ind + 1], &opt, n_threads);
 		}
-		mm_idx_destroy(mi);
+		//mm_idx_destroy(mi);
 		if (ret < 0) {
 			fprintf(stderr, "ERROR: failed to map the query file\n");
 			exit(EXIT_FAILURE);
 		}
 	}
+#ifdef LISA_HASH
+		mm_idx_destroy_seq(mi);
+#else
+		mm_idx_destroy(mi);
+#endif
 	n_parts = idx_rdr->n_parts;
 	mm_idx_reader_close(idx_rdr);
 
@@ -455,5 +543,10 @@ int main(int argc, char *argv[])
 			fprintf(stderr, " %s", argv[i]);
 		fprintf(stderr, "\n[M::%s] Real time: %.3f sec; CPU: %.3f sec; Peak RSS: %.3f GB\n", __func__, realtime() - mm_realtime0, cputime(), peakrss() / 1024.0 / 1024.0 / 1024.0);
 	}
+
+	fprintf(stderr, "minimizer-lookup: %lld dp: %lld rmq: %lld rmq_t1: %lld rmq_t2: %lld rmq_t3: %lld rmq_t4: %lld alignment: %lld %lld\n", minimizer_lookup_time, dp_time, rmq_time, rmq_t1, rmq_t2, rmq_t3, rmq_t4, alignment_time, avg);
+#ifdef LISA_HASH
+	delete lh;
+#endif	
 	return 0;
 }

map.c

@@ -9,6 +9,12 @@
 #include "mmpriv.h"
 #include "bseq.h"
 #include "khash.h"
+#include <x86intrin.h>
+
+#ifdef MANUAL_PROFILING
+extern uint64_t minimizer_lookup_time;
+extern uint64_t rmq_time;
+#endif
 
 struct mm_tbuf_s {
 	void *km;
@@ -173,6 +179,9 @@ static mm128_t *collect_seed_hits_heap(void *km, const mm_mapopt_t *opt, int max
 static mm128_t *collect_seed_hits(void *km, const mm_mapopt_t *opt, int max_occ, const mm_idx_t *mi, const char *qname, const mm128_v *mv, int qlen, int64_t *n_a, int *rep_len,
 								  int *n_mini_pos, uint64_t **mini_pos)
 {
+#ifdef MANUAL_PROFILING
+	uint64_t lookup_start = __rdtsc();
+#endif
 	int i, n_m;
 	mm_seed_t *m;
 	mm128_t *a;
@@ -205,6 +214,9 @@ static mm128_t *collect_seed_hits(void *km, const mm_mapopt_t *opt, int max_occ,
 	}
 	kfree(km, m);
 	radix_sort_128x(a, a + (*n_a));
+#ifdef MANUAL_PROFILING
+	minimizer_lookup_time += __rdtsc() - lookup_start;
+#endif
 	return a;
 }
 
@@ -212,7 +224,7 @@ static void chain_post(const mm_mapopt_t *opt, int max_chain_gap_ref, const mm_i
 {
 	if (!(opt->flag & MM_F_ALL_CHAINS)) { // don't choose primary mapping(s)
 		mm_set_parent(km, opt->mask_level, opt->mask_len, *n_regs, regs, opt->a * 2 + opt->b, opt->flag&MM_F_HARD_MLEVEL, opt->alt_drop);
-		if (n_segs <= 1) mm_select_sub(km, opt->pri_ratio, mi->k*2, opt->best_n, 1, opt->max_gap * 0.8, n_regs, regs);
+		if (n_segs <= 1) mm_select_sub(km, opt->pri_ratio, mi->k*2, opt->best_n, n_regs, regs);
 		else mm_select_sub_multi(km, opt->pri_ratio, 0.2f, 0.7f, max_chain_gap_ref, mi->k*2, opt->best_n, n_segs, qlens, n_regs, regs);
 	}
 }
@@ -223,7 +235,7 @@ static mm_reg1_t *align_regs(const mm_mapopt_t *opt, const mm_idx_t *mi, void *k
 	regs = mm_align_skeleton(km, opt, mi, qlen, seq, n_regs, regs, a); // this calls mm_filter_regs()
 	if (!(opt->flag & MM_F_ALL_CHAINS)) { // don't choose primary mapping(s)
 		mm_set_parent(km, opt->mask_level, opt->mask_len, *n_regs, regs, opt->a * 2 + opt->b, opt->flag&MM_F_HARD_MLEVEL, opt->alt_drop);
-		mm_select_sub(km, opt->pri_ratio, mi->k*2, opt->best_n, 0, opt->max_gap * 0.8, n_regs, regs);
+		mm_select_sub(km, opt->pri_ratio, mi->k*2, opt->best_n, n_regs, regs);
 		mm_set_sam_pri(*n_regs, regs);
 	}
 	return regs;
@@ -240,7 +252,6 @@ void mm_map_frag(const mm_idx_t *mi, int n_segs, const int *qlens, const char **
 	mm128_v mv = {0,0,0};
 	mm_reg1_t *regs0;
 	km_stat_t kmst;
-	float chn_pen_gap, chn_pen_skip;
 
 	for (i = 0, qlen_sum = 0; i < n_segs; ++i)
 		qlen_sum += qlens[i], n_regs[i] = 0, regs[i] = 0;
@@ -248,12 +259,11 @@ void mm_map_frag(const mm_idx_t *mi, int n_segs, const int *qlens, const char **
 	if (qlen_sum == 0 || n_segs <= 0 || n_segs > MM_MAX_SEG) return;
 	if (opt->max_qlen > 0 && qlen_sum > opt->max_qlen) return;
 
-	hash  = qname && !(opt->flag & MM_F_NO_HASH_NAME)? __ac_X31_hash_string(qname) : 0;
+	hash  = qname? __ac_X31_hash_string(qname) : 0;
 	hash ^= __ac_Wang_hash(qlen_sum) + __ac_Wang_hash(opt->seed);
 	hash  = __ac_Wang_hash(hash);
 
 	collect_minimizers(b->km, opt, mi, n_segs, qlens, seqs, &mv);
-	if (opt->q_occ_frac > 0.0f) mm_seed_mz_flt(b->km, &mv, opt->mid_occ, opt->q_occ_frac);
 	if (opt->flag & MM_F_HEAP_SORT) a = collect_seed_hits_heap(b->km, opt, opt->mid_occ, mi, qname, &mv, qlen_sum, &n_a, &rep_len, &n_mini_pos, &mini_pos);
 	else a = collect_seed_hits(b->km, opt, opt->mid_occ, mi, qname, &mv, qlen_sum, &n_a, &rep_len, &n_mini_pos, &mini_pos);
 
@@ -275,25 +285,39 @@ void mm_map_frag(const mm_idx_t *mi, int n_segs, const int *qlens, const char **
 		if (max_chain_gap_ref < opt->max_gap) max_chain_gap_ref = opt->max_gap;
 	} else max_chain_gap_ref = opt->max_gap;
 
-	chn_pen_gap  = opt->chain_gap_scale * 0.01 * mi->k;
-	chn_pen_skip = opt->chain_skip_scale * 0.01 * mi->k;
 	if (opt->flag & MM_F_RMQ) {
 		a = mg_lchain_rmq(opt->max_gap, opt->rmq_inner_dist, opt->bw, opt->max_chain_skip, opt->rmq_size_cap, opt->min_cnt, opt->min_chain_score,
-						  chn_pen_gap, chn_pen_skip, n_a, a, &n_regs0, &u, b->km);
+						  opt->chain_gap_scale * 0.01 * mi->k, 0.0f, n_a, a, &n_regs0, &u, b->km);
+//		a = mg_lchain_dp(opt->max_gap, opt->rmq_inner_dist, opt->bw, opt->max_chain_skip, opt->rmq_size_cap, opt->min_cnt, opt->min_chain_score,
+//						 	opt->chain_gap_scale * 0.01 * mi->k, 0.0f, is_splice, n_segs, n_a, a, &n_regs0, &u, b->km);
+
+
 	} else {
+		//fprintf(stderr, "dp call - n_a = %lld\n", n_a);
 		a = mg_lchain_dp(max_chain_gap_ref, max_chain_gap_qry, opt->bw, opt->max_chain_skip, opt->max_chain_iter, opt->min_cnt, opt->min_chain_score,
-						 chn_pen_gap, chn_pen_skip, is_splice, n_segs, n_a, a, &n_regs0, &u, b->km);
+						 opt->chain_gap_scale * 0.01 * mi->k, 0.0f, is_splice, n_segs, n_a, a, &n_regs0, &u, b->km);
 	}
 
 	if (opt->bw_long > opt->bw && (opt->flag & (MM_F_SPLICE|MM_F_SR|MM_F_NO_LJOIN)) == 0 && n_segs == 1 && n_regs0 > 1) { // re-chain/long-join for long sequences
 		int32_t st = (int32_t)a[0].y, en = (int32_t)a[(int32_t)u[0] - 1].y;
 		if (qlen_sum - (en - st) > opt->rmq_rescue_size || en - st > qlen_sum * opt->rmq_rescue_ratio) {
+#ifdef MANUAL_PROFILING
+//			uint64_t tim = __rdtsc();
+#endif
+	//		fprintf(stderr, "pre: rmq rechain call - n_a = %lld n_regs = %lld\n",n_a, n_regs0);
 			int32_t i;
+			int64_t prev_n_a = n_a;
 			for (i = 0, n_a = 0; i < n_regs0; ++i) n_a += (int32_t)u[i];
 			kfree(b->km, u);
 			radix_sort_128x(a, a + n_a);
+//			fprintf(stderr, "post: rmq rechain call - prev_n_a = %lld n_a = %lld n_regs = %lld\n",prev_n_a, n_a, n_regs0);
+//			a = mg_lchain_dp(opt->max_gap, opt->rmq_inner_dist, opt->bw_long, opt->max_chain_skip, opt->rmq_size_cap, opt->min_cnt, opt->min_chain_score,
+//						 	opt->chain_gap_scale * 0.01 * mi->k, 0.0f, is_splice, n_segs, n_a, a, &n_regs0, &u, b->km);
 			a = mg_lchain_rmq(opt->max_gap, opt->rmq_inner_dist, opt->bw_long, opt->max_chain_skip, opt->rmq_size_cap, opt->min_cnt, opt->min_chain_score,
-							  chn_pen_gap, chn_pen_skip, n_a, a, &n_regs0, &u, b->km);
+							  opt->chain_gap_scale * 0.01 * mi->k, 0.0f, n_a, a, &n_regs0, &u, b->km);
+#ifdef MANUAL_PROFILING
+//			rmq_time += __rdtsc() - tim;
+#endif
 		}
 	} else if (opt->max_occ > opt->mid_occ && rep_len > 0 && !(opt->flag & MM_F_RMQ)) { // re-chain, mostly for short reads
 		int rechain = 0;
@@ -316,7 +340,7 @@ void mm_map_frag(const mm_idx_t *mi, int n_segs, const int *qlens, const char **
 			if (opt->flag & MM_F_HEAP_SORT) a = collect_seed_hits_heap(b->km, opt, opt->max_occ, mi, qname, &mv, qlen_sum, &n_a, &rep_len, &n_mini_pos, &mini_pos);
 			else a = collect_seed_hits(b->km, opt, opt->max_occ, mi, qname, &mv, qlen_sum, &n_a, &rep_len, &n_mini_pos, &mini_pos);
 			a = mg_lchain_dp(max_chain_gap_ref, max_chain_gap_qry, opt->bw, opt->max_chain_skip, opt->max_chain_iter, opt->min_cnt, opt->min_chain_score,
-							 chn_pen_gap, chn_pen_skip, is_splice, n_segs, n_a, a, &n_regs0, &u, b->km);
+							 opt->chain_gap_scale * 0.01 * mi->k, 0.0f, is_splice, n_segs, n_a, a, &n_regs0, &u, b->km);
 		}
 	}
 	b->frag_gap = max_chain_gap_ref;
@@ -328,17 +352,15 @@ void mm_map_frag(const mm_idx_t *mi, int n_segs, const int *qlens, const char **
 		mm_hit_sort(b->km, &n_regs0, regs0, opt->alt_drop); // this step can be merged into mm_gen_regs(); will do if this shows up in profile
 	}
 
-	if (mm_dbg_flag & (MM_DBG_PRINT_SEED|MM_DBG_PRINT_CHAIN))
+	if (mm_dbg_flag & MM_DBG_PRINT_SEED)
 		for (j = 0; j < n_regs0; ++j)
 			for (i = regs0[j].as; i < regs0[j].as + regs0[j].cnt; ++i)
 				fprintf(stderr, "CN\t%d\t%s\t%d\t%c\t%d\t%d\t%d\n", j, mi->seq[a[i].x<<1>>33].name, (int32_t)a[i].x, "+-"[a[i].x>>63], (int32_t)a[i].y, (int32_t)(a[i].y>>32&0xff),
 						i == regs0[j].as? 0 : ((int32_t)a[i].y - (int32_t)a[i-1].y) - ((int32_t)a[i].x - (int32_t)a[i-1].x));
 
 	chain_post(opt, max_chain_gap_ref, mi, b->km, qlen_sum, n_segs, qlens, &n_regs0, regs0, a);
-	if (!is_sr && !(opt->flag&MM_F_QSTRAND)) {
+	if (!is_sr && !(opt->flag&MM_F_QSTRAND))
 		mm_est_err(mi, qlen_sum, n_regs0, regs0, a, n_mini_pos, mini_pos);
-		n_regs0 = mm_filter_strand_retained(n_regs0, regs0);
-	}
 
 	if (n_segs == 1) { // uni-segment
 		regs0 = align_regs(opt, mi, b->km, qlens[0], seqs[0], &n_regs0, regs0, a);
@@ -511,7 +533,7 @@ static void merge_hits(step_t *s)
 			mm_hit_sort(km, &s->n_reg[k], s->reg[k], opt->alt_drop);
 			mm_set_parent(km, opt->mask_level, opt->mask_len, s->n_reg[k], s->reg[k], opt->a * 2 + opt->b, opt->flag&MM_F_HARD_MLEVEL, opt->alt_drop);
 			if (!(opt->flag & MM_F_ALL_CHAINS)) {
-				mm_select_sub(km, opt->pri_ratio, s->p->mi->k*2, opt->best_n, 0, opt->max_gap * 0.8, &s->n_reg[k], s->reg[k]);
+				mm_select_sub(km, opt->pri_ratio, s->p->mi->k*2, opt->best_n, &s->n_reg[k], s->reg[k]);
 				mm_set_sam_pri(s->n_reg[k], s->reg[k]);
 			}
 			mm_set_mapq(km, s->n_reg[k], s->reg[k], opt->min_chain_score, opt->a, rep_len, !!(opt->flag & MM_F_SR));
@@ -569,6 +591,7 @@ static void *worker_pipeline(void *shared, int step, void *in)
 		if ((p->opt->flag & MM_F_OUT_CS) && !(mm_dbg_flag & MM_DBG_NO_KALLOC)) km = km_init();
 		for (k = 0; k < s->n_frag; ++k) {
 			int seg_st = s->seg_off[k], seg_en = s->seg_off[k] + s->n_seg[k];
+#ifndef DISABLE_OUTPUT
 			for (i = seg_st; i < seg_en; ++i) {
 				mm_bseq1_t *t = &s->seq[i];
 				if (p->opt->split_prefix && p->n_parts == 0) { // then write to temporary files
@@ -603,6 +626,7 @@ static void *worker_pipeline(void *shared, int step, void *in)
 					mm_err_puts(p->str.s);
 				}
 			}
+#endif
 			for (i = seg_st; i < seg_en; ++i) {
 				for (j = 0; j < s->n_reg[i]; ++j) free(s->reg[i][j].p);
 				free(s->reg[i]);

minimap.h

@@ -39,7 +39,6 @@
 #define MM_F_RMQ           (0x80000000LL)
 #define MM_F_QSTRAND       (0x100000000LL)
 #define MM_F_NO_INV        (0x200000000LL)
-#define MM_F_NO_HASH_NAME  (0x400000000LL)
 
 #define MM_I_HPC          0x1
 #define MM_I_NO_SEQ       0x2
@@ -109,7 +108,7 @@ typedef struct {
 	int32_t mlen, blen;     // seeded exact match length; seeded alignment block length
 	int32_t n_sub;          // number of suboptimal mappings
 	int32_t score0;         // initial chaining score (before chain merging/spliting)
-	uint32_t mapq:8, split:2, rev:1, inv:1, sam_pri:1, proper_frag:1, pe_thru:1, seg_split:1, seg_id:8, split_inv:1, is_alt:1, strand_retained:1, dummy:5;
+	uint32_t mapq:8, split:2, rev:1, inv:1, sam_pri:1, proper_frag:1, pe_thru:1, seg_split:1, seg_id:8, split_inv:1, is_alt:1, dummy:6;
 	uint32_t hash;
 	float div;
 	mm_extra_t *p;
@@ -136,7 +135,6 @@ typedef struct {
 	int min_cnt;         // min number of minimizers on each chain
 	int min_chain_score; // min chaining score
 	float chain_gap_scale;
-	float chain_skip_scale;
 	int rmq_size_cap, rmq_inner_dist;
 	int rmq_rescue_size;
 	float rmq_rescue_ratio;
@@ -165,7 +163,6 @@ typedef struct {
 	int pe_ori, pe_bonus;
 
 	float mid_occ_frac;  // only used by mm_mapopt_update(); see below
-	float q_occ_frac;
 	int32_t min_mid_occ, max_mid_occ;
 	int32_t mid_occ;     // ignore seeds with occurrences above this threshold
 	int32_t max_occ, max_max_occ, occ_dist;
@@ -288,6 +285,13 @@ mm_idx_t *mm_idx_load(FILE *fp);
  */
 void mm_idx_dump(FILE *fp, const mm_idx_t *mi);
 
+/**
+ * Store hash table from minimap2 index into a file
+ * @param f_name     File name for output file
+ * @param mi 	     minimap2 index 	
+ */
+void mm_idx_dump_hash(const char* f_name, const mm_idx_t *mi);
+
 /**
  * Create an index from strings in memory
  *
@@ -316,6 +320,19 @@ void mm_idx_stat(const mm_idx_t *idx);
  * @param r          minimap2 index
  */
 void mm_idx_destroy(mm_idx_t *mi);
+/**
+ * Destroy/deallocate an hash table index
+ *
+ * @param r          minimap2 index
+ */
+void mm_idx_destroy_mm_hash(mm_idx_t *mi);
+
+/**
+ * Destroy/deallocate target sequences
+ *
+ * @param r          minimap2 index
+ */
+void mm_idx_destroy_seq(mm_idx_t *mi);
 
 /**
  * Initialize a thread-local buffer for mapping

minimap2.1

@@ -1,4 +1,4 @@
-.TH minimap2 1 "18 December 2021" "minimap2-2.24 (r1122)" "Bioinformatics tools"
+.TH minimap2 1 "7 August 2021" "minimap2-2.22 (r1101)" "Bioinformatics tools"
 .SH NAME
 .PP
 minimap2 - mapping and alignment between collections of DNA sequences
@@ -77,7 +77,7 @@ SAM format.
 Minimizer k-mer length [15]
 .TP
 .BI -w \ INT
-Minimizer window size [10]. A minimizer is the smallest k-mer
+Minimizer window size [2/3 of k-mer length]. A minimizer is the smallest k-mer
 in a window of w consecutive k-mers.
 .TP
 .B -H
@@ -151,16 +151,10 @@ Lower and upper bounds of k-mer occurrences [10,1000000]. The final k-mer occurr
 .BR -f }}.
 This option prevents excessively small or large
 .B -f
-estimated from the input reference. Available since r1034 and deprecating
+estimated from the input reference. It deprecates
 .B --min-occ-floor
 in earlier versions of minimap2.
 .TP
-.BI --q-occ-frac \ FLOAT
-Discard a query minimizer if its occurrence is higher than
-.I FLOAT
-fraction of query minimizers and than the reference occurrence threshold
-[0.01]. Set 0 to disable. Available since r1105.
-.TP
 .BI -e \ INT
 Sample a high-frequency minimizer every
 .I INT
@@ -318,9 +312,6 @@ faster for short reads, but slower for long reads. [no]
 .B --no-pairing
 Treat two reads in a pair as independent reads. The mate related fields in SAM
 are still properly populated.
-.TP
-.B --no-hash-name
-Produce the same alignment for identical sequences regardless of their sequence names.
 .SS Alignment options
 .TP 10
 .BI -A \ INT
@@ -565,7 +556,7 @@ Align older PacBio continuous long (CLR) reads to a reference genome
 .B asm5
 Long assembly to reference mapping
 .RB ( -k19
-.B -w19 -U50,500 --rmq -r1k,100k -g10k -A1 -B19 -O39,81 -E3,1 -s200 -z200
+.B -w19 -U50,500 --rmq -r100k -g10k -A1 -B19 -O39,81 -E3,1 -s200 -z200
 .BR -N50 ).
 Typically, the alignment will not extend to regions with 5% or higher sequence
 divergence. Only use this preset if the average divergence is far below 5%.
@@ -573,14 +564,14 @@ divergence. Only use this preset if the average divergence is far below 5%.
 .B asm10
 Long assembly to reference mapping
 .RB ( -k19
-.B -w19 -U50,500 --rmq -r1k,100k -g10k -A1 -B9 -O16,41 -E2,1 -s200 -z200
+.B -w19 -U50,500 --rmq -r100k -g10k -A1 -B9 -O16,41 -E2,1 -s200 -z200
 .BR -N50 ).
 Up to 10% sequence divergence.
 .TP
 .B asm20
 Long assembly to reference mapping
 .RB ( -k19
-.B -w10 -U50,500 --rmq -r1k,100k -g10k -A1 -B4 -O6,26 -E2,1 -s200 -z200
+.B -w10 -U50,500 --rmq -r100k -g10k -A1 -B4 -O6,26 -E2,1 -s200 -z200
 .BR -N50 ).
 Up to 20% sequence divergence.
 .TP
@@ -606,7 +597,7 @@ Long-read splice alignment for PacBio CCS reads
 .B sr
 Short single-end reads without splicing
 .RB ( -k21
-.B -w11 --sr --frag=yes -A2 -B8 -O12,32 -E2,1 -b0 -r100 -p.5 -N20 -f1000,5000 -n2 -m25
+.B -w11 --sr --frag=yes -A2 -B8 -O12,32 -E2,1 -b0 -r100 -p.5 -N20 -f1000,5000 -n2 -m20
 .B -s40 -g100 -2K50m --heap-sort=yes
 .BR --secondary=no ).
 .TP

misc/paftools.js

@@ -1,6 +1,6 @@
 #!/usr/bin/env k8
 
-var paftools_version = '2.24-r1122';
+var paftools_version = '2.22-r1101';
 
 /*****************************
  ***** Library functions *****
@@ -1532,13 +1532,12 @@ function paf_view(args)
 
 function paf_gff2bed(args)
 {
-	var c, fn_ucsc_fai = null, is_short = false, keep_gff = false, print_junc = false, output_gene = false;
-	while ((c = getopt(args, "u:sgjG")) != null) {
+	var c, fn_ucsc_fai = null, is_short = false, keep_gff = false, print_junc = false;
+	while ((c = getopt(args, "u:sgj")) != null) {
 		if (c == 'u') fn_ucsc_fai = getopt.arg;
 		else if (c == 's') is_short = true;
 		else if (c == 'g') keep_gff = true;
 		else if (c == 'j') print_junc = true;
-		else if (c == 'G') output_gene = true;
 	}
 
 	if (getopt.ind == args.length) {
@@ -1606,10 +1605,8 @@ function paf_gff2bed(args)
 		print(a[0][0], st, en, name, 1000, a[0][3], cds_st, cds_en, color, a.length, sizes.join(",") + ",", starts.join(",") + ",");
 	}
 
-	var re_gtf  = /\b(transcript_id|transcript_type|transcript_biotype|gene_name|gene_id|gbkey|transcript_name) "([^"]+)";/g;
+	var re_gtf = /\b(transcript_id|transcript_type|transcript_biotype|gene_name|gene_id|gbkey|transcript_name) "([^"]+)";/g;
 	var re_gff3 = /\b(transcript_id|transcript_type|transcript_biotype|gene_name|gene_id|gbkey|transcript_name)=([^;]+)/g;
-	var re_gtf_gene  = /\b(gene_id|gene_type|gene_name) "([^;]+)";/g;
-	var re_gff3_gene = /\b(gene_id|gene_type|source_gene|gene_biotype|gene_name)=([^;]+);/g;
 	var buf = new Bytes();
 	var file = args[getopt.ind] == '-'? new File() : new File(args[getopt.ind]);
 
@@ -1623,26 +1620,6 @@ function paf_gff2bed(args)
 			continue;
 		}
 		if (t[0].charAt(0) == '#') continue;
-		if (output_gene) {
-			var id = null, src = null, biotype = null, type = "", name = "N/A";
-			if (t[2] != "gene") continue;
-			while ((m = re_gtf_gene.exec(t[8])) != null) {
-				if (m[1] == "gene_id") id = m[2];
-				else if (m[1] == "gene_type") type = m[2];
-				else if (m[1] == "gene_name") name = m[2];
-			}
-			while ((m = re_gff3_gene.exec(t[8])) != null) {
-				if (m[1] == "gene_id") id = m[2];
-				else if (m[1] == "source_gene") src = m[2];
-				else if (m[1] == "gene_type") type = m[2];
-				else if (m[1] == "gene_biotype") biotype = m[2];
-				else if (m[1] == "gene_name") name = m[2];
-			}
-			if (src != null) id = src;
-			if (type == "" && biotype != null) type = biotype;
-			print(t[0], parseInt(t[3]) - 1, t[4], [id, type, name].join("|"), 1000, t[6]);
-			continue;
-		}
 		if (t[2] != "CDS" && t[2] != "exon") continue;
 		t[3] = parseInt(t[3]) - 1;
 		t[4] = parseInt(t[4]);
@@ -2974,7 +2951,7 @@ function paf_pafcmp(args)
 {
 	var c, opt = { min_len:5000, min_mapq:10, min_ovlp:0.5 };
 	while ((c = getopt(args, "q:")) != null) {
-		if (c == 'q') opt.min_mapq = parseInt(getopt.arg);
+		if (c == 'q') opt.min_mapq = parseInt(opt.arg);
 	}
 
 	var buf = new Bytes();

mmpriv.h

@@ -13,7 +13,6 @@
 #define MM_DBG_PRINT_QNAME   0x2
 #define MM_DBG_PRINT_SEED    0x4
 #define MM_DBG_PRINT_ALN_SEQ 0x8
-#define MM_DBG_PRINT_CHAIN   0x10
 
 #define MM_SEED_LONG_JOIN  (1ULL<<40)
 #define MM_SEED_IGNORE     (1ULL<<41)
@@ -62,7 +61,6 @@ uint32_t ks_ksmall_uint32_t(size_t n, uint32_t arr[], size_t kk);
 void mm_sketch(void *km, const char *str, int len, int w, int k, uint32_t rid, int is_hpc, mm128_v *p);
 
 mm_seed_t *mm_collect_matches(void *km, int *_n_m, int qlen, int max_occ, int max_max_occ, int dist, const mm_idx_t *mi, const mm128_v *mv, int64_t *n_a, int *rep_len, int *n_mini_pos, uint64_t **mini_pos);
-void mm_seed_mz_flt(void *km, mm128_v *mv, int32_t q_occ_max, float q_occ_frac);
 
 double mm_event_identity(const mm_reg1_t *r);
 int mm_write_sam_hdr(const mm_idx_t *mi, const char *rg, const char *ver, int argc, char *argv[]);
@@ -92,9 +90,8 @@ void mm_sync_regs(void *km, int n_regs, mm_reg1_t *regs);
 int mm_squeeze_a(void *km, int n_regs, mm_reg1_t *regs, mm128_t *a);
 int mm_set_sam_pri(int n, mm_reg1_t *r);
 void mm_set_parent(void *km, float mask_level, int mask_len, int n, mm_reg1_t *r, int sub_diff, int hard_mask_level, float alt_diff_frac);
-void mm_select_sub(void *km, float pri_ratio, int min_diff, int best_n, int check_strand, int min_strand_sc, int *n_, mm_reg1_t *r);
+void mm_select_sub(void *km, float pri_ratio, int min_diff, int best_n, int *n_, mm_reg1_t *r);
 void mm_select_sub_multi(void *km, float pri_ratio, float pri1, float pri2, int max_gap_ref, int min_diff, int best_n, int n_segs, const int *qlens, int *n_, mm_reg1_t *r);
-int mm_filter_strand_retained(int n_regs, mm_reg1_t *r);
 void mm_filter_regs(const mm_mapopt_t *opt, int qlen, int *n_regs, mm_reg1_t *regs);
 void mm_hit_sort(void *km, int *n_regs, mm_reg1_t *r, float alt_diff_frac);
 void mm_set_mapq(void *km, int n_regs, mm_reg1_t *regs, int min_chain_sc, int match_sc, int rep_len, int is_sr);

options.c

@@ -1,7 +1,7 @@
 #include <stdio.h>
 #include <limits.h>
 #include "mmpriv.h"
-
+extern bool enable_vect_dp_chaining;
 void mm_idxopt_init(mm_idxopt_t *opt)
 {
 	memset(opt, 0, sizeof(mm_idxopt_t));
@@ -19,7 +19,6 @@ void mm_mapopt_init(mm_mapopt_t *opt)
 	opt->min_mid_occ = 10;
 	opt->max_mid_occ = 1000000;
 	opt->sdust_thres = 0; // no SDUST masking
-	opt->q_occ_frac = 0.01f;
 
 	opt->min_cnt = 3;
 	opt->min_chain_score = 40;
@@ -33,7 +32,6 @@ void mm_mapopt_init(mm_mapopt_t *opt)
 	opt->rmq_rescue_size = 1000;
 	opt->rmq_rescue_ratio = 0.1f;
 	opt->chain_gap_scale = 0.8f;
-	opt->chain_skip_scale = 0.0f;
 	opt->max_max_occ = 4095;
 	opt->occ_dist = 500;
 
@@ -54,7 +52,6 @@ void mm_mapopt_init(mm_mapopt_t *opt)
 	opt->max_clip_ratio = 1.0f;
 	opt->mini_batch_size = 500000000;
 	opt->max_sw_mat = 100000000;
-	opt->cap_kalloc = 1000000000;
 
 	opt->rank_min_len = 500;
 	opt->rank_frac = 0.9f;
@@ -74,7 +71,6 @@ void mm_mapopt_update(mm_mapopt_t *opt, const mm_idx_t *mi)
 		if (opt->max_mid_occ > opt->min_mid_occ && opt->mid_occ > opt->max_mid_occ)
 			opt->mid_occ = opt->max_mid_occ;
 	}
-	if (opt->bw_long < opt->bw) opt->bw_long = opt->bw;
 	if (mm_verbose >= 3)
 		fprintf(stderr, "[M::%s::%.3f*%.2f] mid_occ = %d\n", __func__, realtime() - mm_realtime0, cputime() / (realtime() - mm_realtime0), opt->mid_occ);
 }
@@ -98,6 +94,9 @@ int mm_set_opt(const char *preset, mm_idxopt_t *io, mm_mapopt_t *mo)
 		mo->bw = mo->bw_long = 2000;
 		mo->occ_dist = 0;
 	} else if (strcmp(preset, "map10k") == 0 || strcmp(preset, "map-pb") == 0) {
+		#if defined (PARALLEL_CHAINING) && (defined(__AVX2__)) && (!defined(__AVX512BW__))
+		enable_vect_dp_chaining = false;
+		#endif
 		io->flag |= MM_I_HPC, io->k = 19;
 	} else if (strcmp(preset, "ava-pb") == 0) {
 		io->flag |= MM_I_HPC, io->k = 19, io->w = 5;
@@ -106,6 +105,9 @@ int mm_set_opt(const char *preset, mm_idxopt_t *io, mm_mapopt_t *mo)
 		mo->bw_long = mo->bw;
 		mo->occ_dist = 0;
 	} else if (strcmp(preset, "map-hifi") == 0 || strcmp(preset, "map-ccs") == 0) {
+		#if defined (PARALLEL_CHAINING) && (defined(__AVX2__)) && (!defined(__AVX512BW__))
+		enable_vect_dp_chaining = false;
+		#endif
 		io->flag = 0, io->k = 19, io->w = 19;
 		mo->max_gap = 10000;
 		mo->a = 1, mo->b = 4, mo->q = 6, mo->q2 = 26, mo->e = 2, mo->e2 = 1;
@@ -114,7 +116,7 @@ int mm_set_opt(const char *preset, mm_idxopt_t *io, mm_mapopt_t *mo)
 		mo->min_dp_max = 200;
 	} else if (strncmp(preset, "asm", 3) == 0) {
 		io->flag = 0, io->k = 19, io->w = 19;
-		mo->bw = 1000, mo->bw_long = 100000;
+		mo->bw = mo->bw_long = 100000;
 		mo->max_gap = 10000;
 		mo->flag |= MM_F_RMQ;
 		mo->min_mid_occ = 50, mo->max_mid_occ = 500;

python/cmappy.pxd

@@ -23,7 +23,6 @@ cdef extern from "minimap.h":
 		int min_cnt
 		int min_chain_score
 		float chain_gap_scale
-		float chain_skip_scale
 		int rmq_size_cap, rmq_inner_dist
 		int rmq_rescue_size
 		float rmq_rescue_ratio
@@ -52,7 +51,6 @@ cdef extern from "minimap.h":
 		int pe_ori, pe_bonus
 
 		float mid_occ_frac
-		float q_occ_frac
 		int32_t min_mid_occ
 		int32_t mid_occ
 		int32_t max_occ

python/mappy.pyx

@@ -3,7 +3,7 @@ from libc.stdlib cimport free
 cimport cmappy
 import sys
 
-__version__ = '2.24'
+__version__ = '2.22'
 
 cmappy.mm_reset_timer()
 

seed.c

@@ -1,34 +1,40 @@
 #include "mmpriv.h"
 #include "kalloc.h"
 #include "ksort.h"
+#include <stdlib.h>
+#include<algorithm>
+#include <x86intrin.h>
 
-void mm_seed_mz_flt(void *km, mm128_v *mv, int32_t q_occ_max, float q_occ_frac)
-{
-	mm128_t *a;
-	size_t i, j, st;
-	if (mv->n <= q_occ_max || q_occ_frac <= 0.0f || q_occ_max <= 0) return;
-	KMALLOC(km, a, mv->n);
-	for (i = 0; i < mv->n; ++i)
-		a[i].x = mv->a[i].x, a[i].y = i;
-	radix_sort_128x(a, a + mv->n);
-	for (st = 0, i = 1; i <= mv->n; ++i) {
-		if (i == mv->n || a[i].x != a[st].x) {
-			int32_t cnt = i - st;
-			if (cnt > q_occ_max && cnt > mv->n * q_occ_frac)
-				for (j = st; j < i; ++j)
-					mv->a[a[j].y].x = 0;
-			st = i;
-		}
-	}
-	kfree(km, a);
-	for (i = j = 0; i < mv->n; ++i)
-		if (mv->a[i].x != 0)
-			mv->a[j++] = mv->a[i];
-	mv->n = j;
-}
+#ifdef LISA_HASH
+#include "lisa_hash.h"
+extern lisa_hash<uint64_t, uint64_t> *lh;
+#endif
+extern uint64_t minimizer_lookup_time;
 
 mm_seed_t *mm_seed_collect_all(void *km, const mm_idx_t *mi, const mm128_v *mv, int32_t *n_m_)
 {
+//#ifdef MANUAL_PROFILING
+//	uint64_t lookup_start = __rdtsc();
+//#endif
+
+#ifdef LISA_HASH
+//-----------------------------------
+	uint64_t** cr_batch = (uint64_t**) malloc((mv->n)*sizeof(uint64_t*));
+	int* t_batch = (int*)malloc((mv->n)*sizeof(int));
+	uint64_t* minimizers = (uint64_t*) malloc((mv->n)*sizeof(uint64_t));
+	int64_t* lisa_pos = (int64_t*) malloc((max(32, (int)mv->n))* sizeof(int64_t));
+
+	for (size_t i = 0; i < mv->n; i++) {
+		mm128_t *p = &mv->a[i];
+		minimizers[i] = p->x>>8;
+	}
+	
+	lh->mm_idx_get_batched(minimizers, mv->n, lisa_pos, cr_batch, t_batch); 
+//-----------------------------------
+
+#endif
+
+
 	mm_seed_t *m;
 	size_t i;
 	int32_t k;
@@ -39,7 +45,12 @@ mm_seed_t *mm_seed_collect_all(void *km, const mm_idx_t *mi, const mm128_v *mv,
 		mm128_t *p = &mv->a[i];
 		uint32_t q_pos = (uint32_t)p->y, q_span = p->x & 0xff;
 		int t;
+#ifdef LISA_HASH
+		t = t_batch[i];
+		cr = cr_batch[i];
+#else		
 		cr = mm_idx_get(mi, p->x>>8, &t);
+#endif
 		if (t == 0) continue;
 		q = &m[k++];
 		q->q_pos = q_pos, q->q_span = q_span, q->cr = cr, q->n = t, q->seg_id = p->y >> 32;
@@ -47,7 +58,16 @@ mm_seed_t *mm_seed_collect_all(void *km, const mm_idx_t *mi, const mm128_v *mv,
 		if (i > 0 && p->x>>8 == mv->a[i - 1].x>>8) q->is_tandem = 1;
 		if (i < mv->n - 1 && p->x>>8 == mv->a[i + 1].x>>8) q->is_tandem = 1;
 	}
+#ifdef LISA_HASH
+	free(cr_batch);
+	free(t_batch);
+	free(minimizers);
+	free(lisa_pos);
+#endif
 	*n_m_ = k;
+//#ifdef MANUAL_PROFILING
+//	minimizer_lookup_time += __rdtsc() - lookup_start;
+//#endif
 	return m;
 }
 

setup.py

@@ -23,7 +23,7 @@ def readme():
 
 setup(
 	name = 'mappy',
-	version = '2.24',
+	version = '2.22',
 	url = 'https://github.com/lh3/minimap2',
 	description = 'Minimap2 python binding',
 	long_description = readme(),

tex/minimap2.bib

@@ -370,6 +370,7 @@
 	year = {2020},
 	doi = {10.1101/2020.11.01.363887},
 	publisher = {Cold Spring Harbor Laboratory},
+	abstract = {About 5-10\% of the human genome remains inaccessible for functional analysis due to the presence of repetitive sequences such as segmental duplications and tandem repeat arrays. To enable high-quality resequencing of personal genomes, it is crucial to support end-to-end genome variant discovery using repeat-aware read mapping methods. In this study, we highlight the fact that existing long read mappers often yield incorrect alignments and variant calls within long, near-identical repeats, as they remain vulnerable to allelic bias. In the presence of a non-reference allele within a repeat, a read sampled from that region could be mapped to an incorrect repeat copy because the standard pairwise sequence alignment scoring system penalizes true variants.To address the above problem, we propose a novel, long read mapping method that addresses allelic bias by making use of minimal confidently alignable substrings (MCASs). MCASs are formulated as minimal length substrings of a read that have unique alignments to a reference locus with sufficient mapping confidence (i.e., a mapping quality score above a user-specified threshold). This approach treats each read mapping as a collection of confident sub-alignments, which is more tolerant of structural variation and more sensitive to paralog-specific variants (PSVs) within repeats. We mathematically define MCASs and discuss an exact algorithm as well as a practical heuristic to compute them. The proposed method, referred to as Winnowmap2, is evaluated using simulated as well as real long read benchmarks using the recently completed gapless assemblies of human chromosomes X and 8 as a reference. We show that Winnowmap2 successfully addresses the issue of allelic bias, enabling more accurate downstream variant calls in repetitive sequences. As an example, using simulated PacBio HiFi reads and structural variants in chromosome 8, Winnowmap2 alignments achieved the lowest false-negative and false-positive rates (1.89\%, 1.89\%) for calling structural variants within near-identical repeats compared to minimap2 (39.62\%, 5.88\%) and NGMLR (56.60\%, 36.11\%) respectively.Winnowmap2 code is accessible at https://github.com/marbl/WinnowmapCompeting Interest StatementThe authors have declared no competing interest.},
 	URL = {https://www.biorxiv.org/content/early/2020/11/02/2020.11.01.363887},
 	eprint = {https://www.biorxiv.org/content/early/2020/11/02/2020.11.01.363887.full.pdf},
 	journal = {bioRxiv}
@@ -448,13 +449,3 @@
 	Title = {A synthetic-diploid benchmark for accurate variant-calling evaluation},
 	Volume = {15},
 	Year = {2018}}
-
-@article{Gu:1995wt,
-	author = {Gu, X and Li, W H},
-	journal = {J Mol Evol},
-	month = {Apr},
-	number = {4},
-	pages = {464-73},
-	title = {The size distribution of insertions and deletions in human and rodent pseudogenes suggests the logarithmic gap penalty for sequence alignment},
-	volume = {40},
-	year = {1995}}

tex/mm2-update.tex

@@ -57,8 +57,8 @@ in v2.19 through v2.22 to improve mapping results.
 \begin{methods}
 \section{Methods}
 
-\subsection{Rescuing high-occurrence $k$-mers}\label{sec:high-occ}
-Minimap2 keeps all $k$-mer minimizers~\citep{Roberts:2004fv} during indexing. Its original
+\subsection{Rescuing high-occurrence $k$-mers}
+Minimap2 keeps all $k$-mer minimizers during indexing. Its original
 implementation only selected low-occurrence minimizers during mapping. The
 cutoff is a few hundred for mapping long reads against a human genome. If a
 read habors only a few or even no low-occurrence minimizers, it will fail
@@ -66,24 +66,23 @@ chaining due to insufficient anchors.
 
 To resolve this issue, we implemented a new heuristic to add additional
 minimizers. Suppose we are looking at two adjacent low-occurence $k$-mers
-located at position $x_1$ and $x_2$, respectively. If $|x_1-x_2|\ge L$,
-minimap2 v2.22 additionally selects $\lfloor|x_1-x_2|/L\rfloor$ minimizers
-of the lowest occurrence among minimizers between $x_1$ and $x_2$. Here
-parameter $L$ controls the frequency of sampling. It defaults to 500.
+located at position $x_1$ and $x_2$, respectively. If $|x_1-x_2|\ge500$,
+minimap2 v2.22 additionally selects $\lfloor|x_1-x_2|/500\rfloor$ minimizers
+of the lowest occurrence among minimizers between $x_1$ and $x_2$.
+We use a binary heap data
+structure to select minimizers of the lowest occurrence in this interval.
 This strategy adds necessary anchors at the cost of increasing total alignment
 time by a few percent on real data.
 
 \subsection{Aligning through longer INDELs}
 The original minimap2 may fail to align long INDELs due to its chaining
 heuristics. Briefly, minimap2 applies dynamic programming (DP) to chain
-minimizer anchors. This is a quadratic algorithm, slow for chaining
+minimizer anchors. This is a quadratic algorithm, which is slow for chaining
 contigs. For acceptable performance, the original minimap2 uses a 500bp band by
-default, which means a gap longer than 500bp will stop chaining.
-To align through longer gaps, older minimap2 implemented a long-join heurstic as follows.
-If there is an INDEL longer than 500bp and the two chains around the INDEL
+default. If there is an INDEL longer than 500bp and the two chains around the INDEL
 have no overlaps on either the query or the reference sequence, minimap2 may
-join the two short chains later.
-This heuristic may fail around VNTRs because short chains
+join the two short chains later at a later step. We call it the
+long-join heuristic. This heuristic may fail around VNTRs because short chains
 often have overlaps in VNTRs. More subtly, minimap2 may escape the inner DP
 loop early, again for performance, if the chaining result is not improved for
 50 iterations. When there is a copy number change in a long segmental
@@ -91,13 +90,13 @@ duplication, the early escape may break around the event even if users
 specify a large band.
 
 In minigraph~\citep{Li:2020aa}, we developed a new chaining algorithm that
-finds up to 1kb INDELs with DP-based chaining and goes through longer INDELs with a
+finds short INDELs with DP-based chaining and goes through long INDELs with a
 subquadratic algorithm~\citep{DBLP:conf/wabi/AbouelhodaO03}. We ported the same
 algorithm to minimap2 for contig mapping. For long-read mapping, the minigraph
-algorithm is slower. Minimap2 v2.22 still uses the DP-based algorithm to
+algorithm is slower. Minimap2 v2.22 now still uses the DP-based algorithm to
 find short chains and then invokes the minigraph algorithm to rechain anchors in
 these short chains. The rechaining step achieves the same goal as long-join
-but is more reliable because it can resolve overlaps between short chains. The old
+but is more reliable as it can resolve overlaps between short chains. The old
 long-join heuristic has since been removed.
 
 \subsection{Properly mapping long reads with SVs}
@@ -107,25 +106,25 @@ the best scoring alignment is sometimes not the correct alignment.
 \citet{Jain2020.11.01.363887} resolved this dilemma by altering the mapping
 algorithm.
 
-In our view, this problem is rooted in inapropriate scoring: affine-gap penalty
+In our view, this problem is rooted in impropriate scoring: affine-gap penalty
 over-penalizes a long INDEL that was often evolutionarily created in one event.
-We should not penalize a SV by a function linear in the SV length. Minimap2 v2.22 instead rescores
+We should not penalize a SV linearly in its length. Minimap2 v2.22 rescores
 an alignment with the following scoring function. Suppose an alignment consists
 of $M$ matching bases, $N$ substitutions and $G$ gap opens, we empirically
 score the alignment with
 $$
-S=M-\frac{N+G}{2d}-\sum_{i=1}^G\log_2(1+g_i)
+M-\frac{N+G}{2d}-\sum_{i=1}^G\log_2(1+g_i)
 $$
 where $g_i\ge1$ is the length of the $i$-th gap and
 $$
 d=\max\left\{\frac{N+G}{M+N+G},0.02\right\}
 $$
-It approximates per-base sequence divergence except with the smallest value set
+Here $d$ approximates per-base sequence divergence with the smallest value set
 to 2\%. As an analogy to affine-gap scoring, the matching score in our scheme
 is 1, the mismatch and gap open penalties are both $1/2d$ and the gap extension
-penalty is a logarithm function of the gap length~\citep{Gu:1995wt}. Our scoring gives a long SV
+penalty is a logarithm function of the gap length. Our scoring gives a long SV
 a much milder penalty. In terms of time complexity, scoring an alignment is
-linear in the length of the alignment. The time spent on rescoring is negligible in
+linear in the length of the alignment. Time spent on rescoring is negligible in
 practice.
 
 %If we assume sequences evolve under a duplication-mutation model, we may have a
@@ -145,15 +144,13 @@ practice.
 \toprule
 $[$Benchmark$]$ Metric & v2.22 & v2.18 & Winno & lra \\
 \midrule
-$[$sim-map$]$ \% mapped reads at Q10      & 97.9       & 97.6      & {\bf 99.0}& 97.3 \\
-$[$sim-map$]$ err. rate at Q10 (phredQ)   & {\bf 52}   & {\bf 52}  & 38        & 24 \\
-$[$winno-cmp$]$ rate of diff. (phredQ)    & {\bf 41}   & 37        & truth     & 18 \\
-$[$winno-cmp$]$ CPU time (hour)           & {\bf 5.0}  & 5.3       & 71.8      & 13.1 \\
-$[$winno-cmp$]$ peak RAM (Gb)             & 17.1       & 14.4      & {\bf 9.6} & 12.4 \\
-$[$sim-sv$]$  \% false negative rate      & {\bf 0.5}  & 2.0       & {\bf 0.5} & 1.4  \\
-$[$sim-sv$]$  \% false discovery rate     & {\bf 0.0}  & 0.1       & {\bf 0.0} & 0.1  \\
-$[$real-sv-1k$]$ \% false negative rate   & {\bf 7.3}  & 20.0      & 13.0      & N/A \\
-$[$real-sv-1k$]$ \% false discovery rate  & 2.7        & {\bf 2.4} & 2.7       & N/A \\
+$[$sim-map$]$ \% mapped reads at Q10      & 97.9 & 97.6 & {\bf 99.0} & 97.3 \\
+$[$sim-map$]$ err. rate at Q10 (phredQ)   & {\bf 52}   & {\bf 52}   & 38   & 24 \\
+$[$winno-cmp$]$  rate of diff. (phredQ)   & {\bf 41}   & 37   & N/A  & 18 \\
+$[$sim-sv$]$  \% false negative rate      & {\bf 0.5}  & 2.0  & {\bf 0.5}  & 1.4  \\
+$[$sim-sv$]$  \% false discovery rate     & {\bf 0.0}  & 0.1  & {\bf 0.0}  & 0.1  \\
+$[$real-sv-1k$]$ \% false negative rate   & {\bf 7.3}  & 20.0 & 13.0 & N/A \\
+$[$real-sv-1k$]$ \% false discovery rate  & 2.7  & {\bf 2.4}  & 2.7  & N/A \\
 \botrule
 \end{tabular}}
 {In $[$sim-map$]$, 152,713 reads were simulated from the CHM13 telomere-to-telomere assembly v1.1
@@ -162,11 +159,11 @@ $[$real-sv-1k$]$ \% false discovery rate  & 2.7        & {\bf 2.4} & 2.7       &
 10 or higher were evaluated by ``paftools.js mapeval''. The mapping error rate
 is measured in the phred scale: if the error rate is $e$, $-10\log_{10}e$ is
 reported in the table. In $[$winno-cmp$]$, 1.39 million CHM13 HiFi reads from
-SRR11292121 were mapped against the same CHM13 assembly. 99.3\% of them were mapped by Winnowmap2
+SRR11292121 were mapped against CHM13. 99.3\% of them were mapped by Winnowmap2
 at mapping quality 10 or higher and were taken as ground truth to evaluate
 minimap2 and lra with ``paftools.js pafcmp''. $[$sim-sv$]$ simulated 1,000
 50bp to 1000bp INDELs from chr8 in CHM13 using SURVIVOR~\citep{Jeffares:2017aa} and simulated Nanopore
-reads at 30-fold coverage with the same pbsim2 command line. SVs were called with
+reads at 30 folds with the same pbsim2 command line. SVs were called with
 ``sniffles -q 10''~\citep{Sedlazeck:2018ab} and compared to the simulated truth with ``SURVIVOR eval
 call.vcf truth.bed 50''. In $[$real-sv-1k$]$, small and long variants were
 called by dipcall-0.3~\citep{Li:2018aa} for HG002 assemblies (AC: GCA\_018852605.1 and
@@ -175,44 +172,31 @@ GCA\_018852615.1) and compared to the GIAB truth~\citep{Zook:2020aa} using ``tru
 \end{table}
 
 We evaluated minimap2 v2.22 along with v2.18, Winnowmap2 v2.03 and lra v1.3.2
-(Table~\ref{tab:1}), using the default setting of each mapper according to the input data types.
-Both versions of minimap2 achieved high mapping accuracy on
+(Table~\ref{tab:1}). Both versions of minimap2 achieved high mapping accuracy on
 simulated Nanopore reads (sim-map). Winnowmap2 aligned more reads at mapping
 quality 10 or higher (mapQ10). However, it may occasionally assign a high mapping
 quality to a read with multiple identical best alignments. This reduced its
 mapping accuracy.
 
-In lack of groud truth for real data, we took Winnowmap2 mapping as ground
-truth to evaluate other mappers (winno-cmp in Table~\ref{tab:1}). Out of 1,378,092 reads with mapQ10
+In lack of groud truth for real data, so we took Winnowmap2 mapping as ground
+truth to evaluate other mappers (winno-cmp). Out of 1,378,092 reads with mapQ10
 alignments by Winnowmap2, minimap2 v2.22 could map all of them. 118 reads, less
 than 0.01\% of all reads, were mapped differently by v2.22. 51 of them have
 multiple identical best alignments. We believe these are more likely to be
 Winnowmap2 errors. Most of the remaining 67 (=118-51) reads have multiple
-highly similar but not identical alignments.
-Minimap2 v2.18 is less consistent with 275 differences including 30 unmapped
-reads mappable by both Winnowmap2 and v2.22.
+highly similar but not identical alignments. We are not sure what are real
+mapping errors.
 
-For the minimizer rescuing parameter $L$ in Section~\ref{sec:high-occ},
-we set its default to 500 such that v2.22 has comparable performance to v2.18 given simulated PacBio and Nanopore human reads.
-To see the effect of this parameter on real data, we tried several different $L$ values.
-v2.22 gave 99 mapping differences at $L=200$,
-118 at $L=500$ (default), 167 at $L=750$ and 224 differences at $L=1000$ in comparison to Winnowmap2.
-$L=200$ is 28\% slower than the default while $L=1000$ is 9\% faster.
-Changing the default minimizer window size (option ``-w'')
-and the initial minimizer occurrence cutoff (option ``-f'')
-also affects performance and accuracy to a similar magnitude.
-
-The two benchmarks above only evaluate read mappings when there are no variations between the reads and the reference.
+The two benchmarks above only evaluate read mappings without variations.
 To measure the mapping accuracy in the presence of SVs (sim-sv), we reproduced
 the results by~\citep{Jain2020.11.01.363887}. Minimap2 v2.22 is as good as
 Winnowmap2 now. Note that we were setting the Sniffles mapping quality
 threshold to 10 in consistent with the benchmarks above. If we used the
-default threshold 20, v2.22 would miss additional five SVs (accounting for
-0.5\% of simulated SVs). For four out of these five missing SVs, minimap2 v2.22
-mapped more variant reads than Winnowmap2. Sniffles did not call these SVs
-because minimap2 tended to give them conservative mapping quality. It is worth
-noting that the simulation here only considers a simple scenario in evolution.
-Non-allelic gene conversions, which happen often in segmental
+default threshold 20, v2.22 would miss additional 0.5\% SVs, suggesting
+minimap2 v2.22 could map variant reads correctly but with conservative mapping
+quality. This observation is more about the interaction between mappers and
+callers. Furthermore, the simulation here only considers a simple scenario in
+evolution. Non-allelic gene conversions, which happen often in segmental
 duplications~\citep{Harpak:2017aa}, would obscure the optimal mapping
 strategies. How much such simple SV simulation informs real-world SV calling
 remains a question.
@@ -220,17 +204,18 @@ remains a question.
 To see if minimap2 v2.22 could improve long INDEL alignment, we ran dipcall on
 contig-to-reference alignments and focused on INDELs longer than 1kb
 (real-sv-1k). v2.22 is more sensitive at comparable specificity, confirming its
-advantage in more contiguous alignment. We could not get dipcall to work well with lra,
-so did not report the numbers.
+advantage in more contiguous alignment. lra is supposed to handle long INDELs
+better, too. However, we could not get lra to work well with dipcall, so did
+not report the numbers.
 
 Minimap2 spends most computing time on base alignment. As recent improvements
 in v2.22 incur little additional computing and do not change the base alignment
-algorithm, the new version has similar performance to older versions. It is
+algorithm, the new version has similar performance to older verions. It is
 consistently faster than Winnowmap2 by several times. Sometimes simple
 heuristics can be as effective as more sophisticated yet slower solutions.
 
 \section*{Acknowledgements}
-We thank Arang Rhie and Chirag Jain for providing motivating examples for which
+We thank Arang Rhie and Chirag Jain for providing motivating examples where
 older minimap2 underperforms.
 
 \paragraph{Funding\textcolon} This work is funded by NHGRI grant R01HG010040.