[auto-optimize]: Speed up alignment about 3x with byte-identical output

Makefile

@@ -85,7 +85,9 @@ ifneq (,$(findstring $(ARCH), x86_64 amd64))
   POPCNT_CAPABILITY ?= 1
   ifneq (0, $(SSE_AVX2))
     ENABLE_V256 := 1
-    V256_FLAG := -march=x86-64-v3 -mtune=znver3 -faligned-new -DSSE_AVX2 -DPOPCNT_CAPABILITY
+    # -ffp-contract=off: no fused multiply-adds, so floating-point results (e.g.
+    # --score-min, MAPQ) are rounded as in the 128-bit build and outputs match it
+    V256_FLAG := -march=x86-64-v3 -mtune=znver3 -ffp-contract=off -faligned-new -DSSE_AVX2 -DPOPCNT_CAPABILITY
     # tell 128-bit version to check for v256 capability and use it, if found
     SSE_FLAG += -DENABLE_x86_64_v3
   endif

aligner_bt.h

@@ -546,7 +546,8 @@ class BtBranchTracer {
 public:
 
 	explicit BtBranchTracer() :
-		prob_(), bs_(), seenPaths_(DP_CAT), sawcell_(DP_CAT), doTri_() { }
+		prob_(), bs_(), seenPaths_(DP_CAT), sawcell_(DP_CAT), doTri_(),
+		resetPending_(false), resetNdiag_(0), resetRflen_(0) { }
 
 	/**
 	 * Add a branch to the queue.
@@ -642,7 +643,10 @@ class BtBranchTracer {
 
 	/**
 	 * Initialize the tracer with respect to a new read.  This involves
-	 * resetting all the state relating to the set of cells already visited
+	 * resetting all the state relating to the set of cells already visited.
+	 * The reset itself is deferred to the next initBt(), since most
+	 * reference windows are never backtraced with this tracer;
+	 * seenPaths_ and sawcell_ are read only from initBt() on.
 	 */
 	void initRef(
 		const char*         rd,     // in: read sequence
@@ -660,7 +664,22 @@ class BtBranchTracer {
 		size_t              nceil)  // in: N ceiling
 	{
 		prob_.initRef(rd, qu, rdlen, rf, rflen, trflen, refid, refoff, fw, rect, cper, &sc, nceil);
-		const size_t ndiag = rflen + rdlen - 1;
+		resetNdiag_ = rflen + rdlen - 1;
+		// An eager reset clears every per-column set below the largest rflen
+		// seen since the last backtrace, so keep the largest.
+		if(!resetPending_ || rflen > resetRflen_) {
+			resetRflen_ = rflen;
+		}
+		resetPending_ = true;
+	}
+
+	/**
+	 * Perform the reset of seenPaths_ and sawcell_ that the last initRef()
+	 * (and any initRef() calls before it since the last initBt()) deferred.
+	 */
+	void resetSeen() {
+		const size_t ndiag = resetNdiag_;
+		const size_t rflen = resetRflen_;
 		seenPaths_.resize(ndiag);
 		for(size_t i = 0; i < ndiag; i++) {
 			seenPaths_[i].clear();
@@ -677,6 +696,7 @@ class BtBranchTracer {
 			sawcell_[i].setCat(DP_CAT);
 			sawcell_[i].clear(); // clear the set
 		}
+		resetPending_ = false;
 	}
 
 	/**
@@ -691,6 +711,9 @@ class BtBranchTracer {
 		bool           doTri,  // in: triangle-shaped mini-fills?
 		RandomSource&  rnd)    // in: random gen, to choose among equal paths
 	{
+		if(resetPending_) {
+			resetSeen();
+		}
 		prob_.initBt(row, col, fill, usecp, escore);
 		Edit e; e.reset();
 		unsorted_.clear();
@@ -939,6 +962,10 @@ class BtBranchTracer {
 	ELList<CpQuad> tri_;        // triangle to fill when doing mini-fills
 	EList<size_t>  ndep_;       // # triangles mini-filled at various depths
 
+	bool           resetPending_; // seenPaths_/sawcell_ reset owed by initRef
+	size_t         resetNdiag_;   // # diagonals of the last initRef's window
+	size_t         resetRflen_;   // largest rflen since the last reset
+
 #ifndef NDEBUG
 	ESet<size_t>  seen_;        // seedn branch ids; should never see same twice
 #endif

aligner_cache.cpp

@@ -88,12 +88,13 @@ bool AlignmentCache::addOnTheFlyImpl(
 		s->payload.len = botf - topf;
 		s->payload.topf = topf;
 		s->payload.topb = topb;
-		for(size_t j = 0; j < (botf-topf); j++) {
-			if(!salist_.add(pool(), OFF_MASK)) {
-				// Change the payload's len field
-				s->payload.len = (TIndexOffU)j;
-				return false; // Exhausted pool memory
-			}
+		// One OFF_MASK (offset not yet resolved) per element of the range
+		const size_t len = botf - topf;
+		const size_t j = salist_.addFillUpTo(pool(), len, OFF_MASK);
+		if(j < len) {
+			// Change the payload's len field
+			s->payload.len = (TIndexOffU)j;
+			return false; // Exhausted pool memory
 		}
 		assert(s->payload.repOk(*this));
 	}