--- ray/src/rt/ambcomp.c	2014/05/01 16:06:11	2.44
+++ ray/src/rt/ambcomp.c	2014/05/09 22:53:11	2.57
@@ -1,5 +1,5 @@
 #ifndef lint
-static const char	RCSid[] = "$Id: ambcomp.c,v 2.44 2014/05/01 16:06:11 greg Exp $";
+static const char	RCSid[] = "$Id: ambcomp.c,v 2.57 2014/05/09 22:53:11 greg Exp $";
 #endif
 /*
  * Routines to compute "ambient" values using Monte Carlo
@@ -8,6 +8,10 @@ static const char	RCSid[] = "$Id: ambcomp.c,v 2.44 201
  *	for Irradiance Caching" by Schwarzhaupt, Wann Jensen, & Jarosz
  *	from ACM SIGGRAPH Asia 2012 conference proceedings.
  *
+ *  Added book-keeping optimization to avoid calculations that would
+ *	cancel due to traversal both directions on edges that are adjacent
+ *	to same-valued triangles.  This cuts about half of Hessian math.
+ *
  *  Declarations of external symbols in ambient.h
  */
 
@@ -23,6 +27,7 @@ extern void		SDsquare2disk(double ds[2], double seedx,
 
 typedef struct {
 	COLOR	v;		/* hemisphere sample value */
+	float	d;		/* reciprocal distance (1/rt) */
 	FVECT	p;		/* intersection point */
 } AMBSAMP;		/* sample value */
 
@@ -34,7 +39,8 @@ typedef struct {
 	AMBSAMP	sa[1];		/* sample array (extends struct) */
 }  AMBHEMI;		/* ambient sample hemisphere */
 
-#define ambsam(h,i,j)	(h)->sa[(i)*(h)->ns + (j)]
+#define AI(h,i,j)	((i)*(h)->ns + (j))
+#define ambsam(h,i,j)	(h)->sa[AI(h,i,j)]
 
 typedef struct {
 	FVECT	r_i, r_i1, e_i, rcp, rI2_eJ2;
@@ -106,23 +112,23 @@ getambsamp(RAY *arp, AMBHEMI *hp, int i, int j, int n)
 		scalecolor(arp->rcoef, 1./AVGREFL);
 	}
 	hlist[0] = hp->rp->rno;
-	hlist[1] = i;
-	hlist[2] = j;
+	hlist[1] = j;
+	hlist[2] = i;
 	multisamp(spt, 2, urand(ilhash(hlist,3)+n));
 	if (!n) {			/* avoid border samples for n==0 */
-		if ((spt[0] < 0.1) | (spt[0] > 0.9))
+		if ((spt[0] < 0.1) | (spt[0] >= 0.9))
 			spt[0] = 0.1 + 0.8*frandom();
-		if ((spt[1] < 0.1) | (spt[1] > 0.9))
+		if ((spt[1] < 0.1) | (spt[1] >= 0.9))
 			spt[1] = 0.1 + 0.8*frandom();
 	}
-	SDsquare2disk(spt, (i+spt[0])/hp->ns, (j+spt[1])/hp->ns);
+	SDsquare2disk(spt, (j+spt[1])/hp->ns, (i+spt[0])/hp->ns);
 	zd = sqrt(1. - spt[0]*spt[0] - spt[1]*spt[1]);
 	for (ii = 3; ii--; )
 		arp->rdir[ii] =	spt[0]*hp->ux[ii] +
 				spt[1]*hp->uy[ii] +
 				zd*hp->rp->ron[ii];
 	checknorm(arp->rdir);
-	dimlist[ndims++] = i*hp->ns + j + 90171;
+	dimlist[ndims++] = AI(hp,i,j) + 90171;
 	rayvalue(arp);			/* evaluate ray */
 	ndims--;			/* apply coefficient */
 	multcolor(arp->rcol, arp->rcoef);
@@ -140,20 +146,16 @@ ambsample(				/* initial ambient division sample */
 	AMBSAMP	*ap = &ambsam(hp,i,j);
 	RAY	ar;
 					/* generate hemispherical sample */
-	if (!getambsamp(&ar, hp, i, j, 0))
-		goto badsample;
-					/* limit vertex distance */
+	if (!getambsamp(&ar, hp, i, j, 0) || ar.rt <= FTINY) {
+		memset(ap, 0, sizeof(AMBSAMP));
+		return(NULL);
+	}
+	ap->d = 1.0/ar.rt;		/* limit vertex distance */
 	if (ar.rt > 10.0*thescene.cusize)
 		ar.rt = 10.0*thescene.cusize;
-	else if (ar.rt <= FTINY)	/* should never happen! */
-		goto badsample;
 	VSUM(ap->p, ar.rorg, ar.rdir, ar.rt);
 	copycolor(ap->v, ar.rcol);
 	return(ap);
-badsample:
-	setcolor(ap->v, 0., 0., 0.);
-	VCOPY(ap->p, hp->rp->rop);
-	return(NULL);
 }
 
 
@@ -161,7 +163,7 @@ badsample:
 static float *
 getambdiffs(AMBHEMI *hp)
 {
-	float	*earr = calloc(hp->ns*hp->ns, sizeof(float));
+	float	*earr = (float *)calloc(hp->ns*hp->ns, sizeof(float));
 	float	*ep;
 	AMBSAMP	*ap;
 	double	b, d2;
@@ -179,25 +181,31 @@ getambdiffs(AMBHEMI *hp)
 			ep[0] += d2;
 			ep[-hp->ns] += d2;
 		}
-		if (j) {		/* from behind */
-			d2 = b - bright(ap[-1].v);
-			d2 *= d2;
-			ep[0] += d2;
-			ep[-1] += d2;
-		}
+		if (!j) continue;
+					/* from behind */
+		d2 = b - bright(ap[-1].v);
+		d2 *= d2;
+		ep[0] += d2;
+		ep[-1] += d2;
+		if (!i) continue;
+					/* diagonal */
+		d2 = b - bright(ap[-hp->ns-1].v);
+		d2 *= d2;
+		ep[0] += d2;
+		ep[-hp->ns-1] += d2;
 	    }
 					/* correct for number of neighbors */
-	earr[0] *= 2.f;
-	earr[hp->ns-1] *= 2.f;
-	earr[(hp->ns-1)*hp->ns] *= 2.f;
-	earr[(hp->ns-1)*hp->ns + hp->ns-1] *= 2.f;
+	earr[0] *= 8./3.;
+	earr[hp->ns-1] *= 8./3.;
+	earr[(hp->ns-1)*hp->ns] *= 8./3.;
+	earr[(hp->ns-1)*hp->ns + hp->ns-1] *= 8./3.;
 	for (i = 1; i < hp->ns-1; i++) {
-		earr[i*hp->ns] *= 4./3.;
-		earr[i*hp->ns + hp->ns-1] *= 4./3.;
+		earr[i*hp->ns] *= 8./5.;
+		earr[i*hp->ns + hp->ns-1] *= 8./5.;
 	}
 	for (j = 1; j < hp->ns-1; j++) {
-		earr[j] *= 4./3.;
-		earr[(hp->ns-1)*hp->ns + j] *= 4./3.;
+		earr[j] *= 8./5.;
+		earr[(hp->ns-1)*hp->ns + j] *= 8./5.;
 	}
 	return(earr);
 }
@@ -208,23 +216,25 @@ static void
 ambsupersamp(double acol[3], AMBHEMI *hp, int cnt)
 {
 	float	*earr = getambdiffs(hp);
-	double	e2sum = 0;
+	double	e2rem = 0;
 	AMBSAMP	*ap;
 	RAY	ar;
-	COLOR	asum;
+	double	asum[3];
 	float	*ep;
-	int	i, j, n;
+	int	i, j, n, nss;
 
 	if (earr == NULL)		/* just skip calc. if no memory */
 		return;
-					/* add up estimated variances */
-	for (ep = earr + hp->ns*hp->ns; ep-- > earr; )
-		e2sum += *ep;
+					/* accumulate estimated variances */
+	for (ep = earr + hp->ns*hp->ns; ep > earr; )
+		e2rem += *--ep;
 	ep = earr;			/* perform super-sampling */
 	for (ap = hp->sa, i = 0; i < hp->ns; i++)
 	    for (j = 0; j < hp->ns; j++, ap++) {
-		int	nss = *ep/e2sum*cnt + frandom();
-		setcolor(asum, 0., 0., 0.);
+		if (e2rem <= FTINY)
+			goto done;	/* nothing left to do */
+		nss = *ep/e2rem*cnt + frandom();
+		asum[0] = asum[1] = asum[2] = 0.0;
 		for (n = 1; n <= nss; n++) {
 			if (!getambsamp(&ar, hp, i, j, n)) {
 				nss = n-1;
@@ -233,28 +243,44 @@ ambsupersamp(double acol[3], AMBHEMI *hp, int cnt)
 			addcolor(asum, ar.rcol);
 		}
 		if (nss) {		/* update returned ambient value */
-			const double	ssf = 1./(nss + 1);
+			const double	ssf = 1./(nss + 1.);
 			for (n = 3; n--; )
-				acol[n] += ssf*colval(asum,n) +
+				acol[n] += ssf*asum[n] +
 						(ssf - 1.)*colval(ap->v,n);
 		}
-		e2sum -= *ep++;		/* update remainders */
+		e2rem -= *ep++;		/* update remainders */
 		cnt -= nss;
 	}
+done:
 	free(earr);
 }
 
 
+/* Return brightness of farthest ambient sample */
+static double
+back_ambval(AMBHEMI *hp, const int n1, const int n2, const int n3)
+{
+	if (hp->sa[n1].d <= hp->sa[n2].d) {
+		if (hp->sa[n1].d <= hp->sa[n3].d)
+			return(colval(hp->sa[n1].v,CIEY));
+		return(colval(hp->sa[n3].v,CIEY));
+	}
+	if (hp->sa[n2].d <= hp->sa[n3].d)
+		return(colval(hp->sa[n2].v,CIEY));
+	return(colval(hp->sa[n3].v,CIEY));
+}
+
+
 /* Compute vectors and coefficients for Hessian/gradient calcs */
 static void
-comp_fftri(FFTRI *ftp, FVECT ap0, FVECT ap1, FVECT rop)
+comp_fftri(FFTRI *ftp, AMBHEMI *hp, const int n0, const int n1)
 {
 	double	rdot_cp, dot_e, dot_er, rdot_r, rdot_r1, J2;
-	int	i;
+	int	ii;
 
-	VSUB(ftp->r_i, ap0, rop);
-	VSUB(ftp->r_i1, ap1, rop);
-	VSUB(ftp->e_i, ap1, ap0);
+	VSUB(ftp->r_i, hp->sa[n0].p, hp->rp->rop);
+	VSUB(ftp->r_i1, hp->sa[n1].p, hp->rp->rop);
+	VSUB(ftp->e_i, hp->sa[n1].p, hp->sa[n0].p);
 	VCROSS(ftp->rcp, ftp->r_i, ftp->r_i1);
 	rdot_cp = 1.0/DOT(ftp->rcp,ftp->rcp);
 	dot_e = DOT(ftp->e_i,ftp->e_i);
@@ -266,8 +292,8 @@ comp_fftri(FFTRI *ftp, FVECT ap0, FVECT ap1, FVECT rop
 	ftp->I2 = ( DOT(ftp->e_i, ftp->r_i1)*rdot_r1 - dot_er*rdot_r +
 			dot_e*ftp->I1 )*0.5*rdot_cp;
 	J2 =  ( 0.5*(rdot_r - rdot_r1) - dot_er*ftp->I2 ) / dot_e;
-	for (i = 3; i--; )
-		ftp->rI2_eJ2[i] = ftp->I2*ftp->r_i[i] + J2*ftp->e_i[i];
+	for (ii = 3; ii--; )
+		ftp->rI2_eJ2[ii] = ftp->I2*ftp->r_i[ii] + J2*ftp->e_i[ii];
 }
 
 
@@ -316,7 +342,7 @@ comp_hessian(FVECT hess[3], FFTRI *ftp, FVECT nrm)
 		hess[i][j] = m1[i][j] + d1*( I3*m2[i][j] + K3*m3[i][j] +
 						2.0*J3*m4[i][j] );
 		hess[i][j] += d2*(i==j);
-		hess[i][j] *= 1.0/PI;
+		hess[i][j] *= -1.0/PI;
 	    }
 }
 
@@ -338,7 +364,7 @@ rev_hessian(FVECT hess[3])
 /* Add to radiometric Hessian from the given triangle */
 static void
 add2hessian(FVECT hess[3], FVECT ehess1[3],
-		FVECT ehess2[3], FVECT ehess3[3], COLORV v)
+		FVECT ehess2[3], FVECT ehess3[3], double v)
 {
 	int	i, j;
 
@@ -359,7 +385,7 @@ comp_gradient(FVECT grad, FFTRI *ftp, FVECT nrm)
 	f1 = 2.0*DOT(nrm, ftp->rcp);
 	VCROSS(ncp, nrm, ftp->e_i);
 	for (i = 3; i--; )
-		grad[i] = (-0.5/PI)*( ftp->I1*ncp[i] + f1*ftp->rI2_eJ2[i] );
+		grad[i] = (0.5/PI)*( ftp->I1*ncp[i] + f1*ftp->rI2_eJ2[i] );
 }
 
 
@@ -375,7 +401,7 @@ rev_gradient(FVECT grad)
 
 /* Add to displacement gradient from the given triangle */
 static void
-add2gradient(FVECT grad, FVECT egrad1, FVECT egrad2, FVECT egrad3, COLORV v)
+add2gradient(FVECT grad, FVECT egrad1, FVECT egrad2, FVECT egrad3, double v)
 {
 	int	i;
 
@@ -384,33 +410,8 @@ add2gradient(FVECT grad, FVECT egrad1, FVECT egrad2, F
 }
 
 
-/* Return brightness of furthest ambient sample */
-static COLORV
-back_ambval(AMBSAMP *ap1, AMBSAMP *ap2, AMBSAMP *ap3, FVECT orig)
-{
-	COLORV	vback;
-	FVECT	vec;
-	double	d2, d2best;
-
-	VSUB(vec, ap1->p, orig);
-	d2best = DOT(vec,vec);
-	vback = colval(ap1->v,CIEY);
-	VSUB(vec, ap2->p, orig);
-	d2 = DOT(vec,vec);
-	if (d2 > d2best) {
-		d2best = d2;
-		vback = colval(ap2->v,CIEY);
-	}
-	VSUB(vec, ap3->p, orig);
-	d2 = DOT(vec,vec);
-	if (d2 > d2best)
-		return(colval(ap3->v,CIEY));
-	return(vback);
-}
-
-
 /* Compute anisotropic radii and eigenvector directions */
-static int
+static void
 eigenvectors(FVECT uv[2], float ra[2], FVECT hessian[3])
 {
 	double	hess2[2][2];
@@ -432,9 +433,10 @@ eigenvectors(FVECT uv[2], float ra[2], FVECT hessian[3
 	if (i == 1)			/* double-root (circle) */
 		evalue[1] = evalue[0];
 	if (!i || ((evalue[0] = fabs(evalue[0])) <= FTINY*FTINY) |
-			((evalue[1] = fabs(evalue[1])) <= FTINY*FTINY) )
-		error(INTERNAL, "bad eigenvalue calculation");
-
+			((evalue[1] = fabs(evalue[1])) <= FTINY*FTINY) ) {
+		ra[0] = ra[1] = maxarad;
+		return;
+	}
 	if (evalue[0] > evalue[1]) {
 		ra[0] = sqrt(sqrt(4.0/evalue[0]));
 		ra[1] = sqrt(sqrt(4.0/evalue[1]));
@@ -492,8 +494,7 @@ ambHessian(				/* anisotropic radii & pos. gradient */
 	}
 					/* compute first row of edges */
 	for (j = 0; j < hp->ns-1; j++) {
-		comp_fftri(&fftr, ambsam(hp,0,j).p,
-				ambsam(hp,0,j+1).p, hp->rp->rop);
+		comp_fftri(&fftr, hp, AI(hp,0,j), AI(hp,0,j+1));
 		if (hessrow != NULL)
 			comp_hessian(hessrow[j], &fftr, hp->rp->ron);
 		if (gradrow != NULL)
@@ -503,8 +504,7 @@ ambHessian(				/* anisotropic radii & pos. gradient */
 	for (i = 0; i < hp->ns-1; i++) {
 	    FVECT	hesscol[3];	/* compute first vertical edge */
 	    FVECT	gradcol;
-	    comp_fftri(&fftr, ambsam(hp,i,0).p,
-			ambsam(hp,i+1,0).p, hp->rp->rop);
+	    comp_fftri(&fftr, hp, AI(hp,i,0), AI(hp,i+1,0));
 	    if (hessrow != NULL)
 		comp_hessian(hesscol, &fftr, hp->rp->ron);
 	    if (gradrow != NULL)
@@ -512,12 +512,11 @@ ambHessian(				/* anisotropic radii & pos. gradient */
 	    for (j = 0; j < hp->ns-1; j++) {
 		FVECT	hessdia[3];	/* compute triangle contributions */
 		FVECT	graddia;
-		COLORV	backg;
-		backg = back_ambval(&ambsam(hp,i,j), &ambsam(hp,i,j+1),
-					&ambsam(hp,i+1,j), hp->rp->rop);
+		double	backg;
+		backg = back_ambval(hp, AI(hp,i,j),
+					AI(hp,i,j+1), AI(hp,i+1,j));
 					/* diagonal (inner) edge */
-		comp_fftri(&fftr, ambsam(hp,i,j+1).p,
-				ambsam(hp,i+1,j).p, hp->rp->rop);
+		comp_fftri(&fftr, hp, AI(hp,i,j+1), AI(hp,i+1,j));
 		if (hessrow != NULL) {
 		    comp_hessian(hessdia, &fftr, hp->rp->ron);
 		    rev_hessian(hesscol);
@@ -529,17 +528,15 @@ ambHessian(				/* anisotropic radii & pos. gradient */
 		    add2gradient(gradient, gradrow[j], graddia, gradcol, backg);
 		}
 					/* initialize edge in next row */
-		comp_fftri(&fftr, ambsam(hp,i+1,j+1).p,
-				ambsam(hp,i+1,j).p, hp->rp->rop);
+		comp_fftri(&fftr, hp, AI(hp,i+1,j+1), AI(hp,i+1,j));
 		if (hessrow != NULL)
 		    comp_hessian(hessrow[j], &fftr, hp->rp->ron);
 		if (gradrow != NULL)
 		    comp_gradient(gradrow[j], &fftr, hp->rp->ron);
 					/* new column edge & paired triangle */
-		backg = back_ambval(&ambsam(hp,i,j+1), &ambsam(hp,i+1,j+1),
-					&ambsam(hp,i+1,j), hp->rp->rop);
-		comp_fftri(&fftr, ambsam(hp,i,j+1).p, ambsam(hp,i+1,j+1).p,
-				hp->rp->rop);
+		backg = back_ambval(hp, AI(hp,i+1,j+1),
+					AI(hp,i+1,j), AI(hp,i,j+1));
+		comp_fftri(&fftr, hp, AI(hp,i,j+1), AI(hp,i+1,j+1));
 		if (hessrow != NULL) {
 		    comp_hessian(hesscol, &fftr, hp->rp->ron);
 		    rev_hessian(hessdia);
@@ -594,6 +591,51 @@ ambdirgrad(AMBHEMI *hp, FVECT uv[2], float dg[2])
 }
 
 
+/* Compute potential light leak direction flags for cache value */
+static uint32
+ambcorral(AMBHEMI *hp, FVECT uv[2], const double r0, const double r1)
+{
+	const double	max_d = 1.0/(minarad*ambacc + 0.001);
+	const double	ang_res = 0.5*PI/(hp->ns-1);
+	const double	ang_step = ang_res/((int)(16/PI*ang_res) + (1+FTINY));
+	double		avg_d = 0;
+	uint32		flgs = 0;
+	int		i, j;
+					/* don't bother for a few samples */
+	if (hp->ns < 12)
+		return(0);
+					/* check distances overhead */
+	for (i = hp->ns*3/4; i-- > hp->ns>>2; )
+	    for (j = hp->ns*3/4; j-- > hp->ns>>2; )
+		avg_d += ambsam(hp,i,j).d;
+	avg_d *= 4.0/(hp->ns*hp->ns);
+	if (avg_d*r0 >= 1.0)		/* ceiling too low for corral? */
+		return(0);
+	if (avg_d >= max_d)		/* insurance */
+		return(0);
+					/* else circle around perimeter */
+	for (i = 0; i < hp->ns; i++)
+	    for (j = 0; j < hp->ns; j += !i|(i==hp->ns-1) ? 1 : hp->ns-1) {
+		AMBSAMP	*ap = &ambsam(hp,i,j);
+		FVECT	vec;
+		double	u, v;
+		double	ang, a1;
+		int	abp;
+		if ((ap->d <= FTINY) | (ap->d >= max_d))
+			continue;	/* too far or too near */
+		VSUB(vec, ap->p, hp->rp->rop);
+		u = DOT(vec, uv[0]) * ap->d;
+		v = DOT(vec, uv[1]) * ap->d;
+		if ((r0*r0*u*u + r1*r1*v*v) * ap->d*ap->d <= 1.0)
+			continue;	/* occluder outside ellipse */
+		ang = atan2a(v, u);	/* else set direction flags */
+		for (a1 = ang-.5*ang_res; a1 <= ang+.5*ang_res; a1 += ang_step)
+			flgs |= 1L<<(int)(16/PI*(a1 + 2.*PI*(a1 < 0)));
+	    }
+	return(flgs);
+}
+
+
 int
 doambient(				/* compute ambient component */
 	COLOR	rcol,			/* input/output color */
@@ -602,11 +644,12 @@ doambient(				/* compute ambient component */
 	FVECT	uv[2],			/* returned (optional) */
 	float	ra[2],			/* returned (optional) */
 	float	pg[2],			/* returned (optional) */
-	float	dg[2]			/* returned (optional) */
+	float	dg[2],			/* returned (optional) */
+	uint32	*crlp			/* returned (optional) */
 )
 {
 	AMBHEMI	*hp = inithemi(rcol, r, wt);
-	int	cnt = 0;
+	int	cnt;
 	FVECT	my_uv[2];
 	double	d, K, acol[3];
 	AMBSAMP	*ap;
@@ -622,8 +665,11 @@ doambient(				/* compute ambient component */
 		pg[0] = pg[1] = 0.0;
 	if (dg != NULL)
 		dg[0] = dg[1] = 0.0;
+	if (crlp != NULL)
+		*crlp = 0;
 					/* sample the hemisphere */
 	acol[0] = acol[1] = acol[2] = 0.0;
+	cnt = 0;
 	for (i = hp->ns; i--; )
 		for (j = hp->ns; j--; )
 			if ((ap = ambsample(hp, i, j)) != NULL) {
@@ -641,21 +687,21 @@ doambient(				/* compute ambient component */
 		return(-1);		/* return value w/o Hessian */
 	}
 	cnt = ambssamp*wt + 0.5;	/* perform super-sampling? */
-	if (cnt > 0)
+	if (cnt > 8)
 		ambsupersamp(acol, hp, cnt);
 	copycolor(rcol, acol);		/* final indirect irradiance/PI */
 	if ((ra == NULL) & (pg == NULL) & (dg == NULL)) {
 		free(hp);
 		return(-1);		/* no radius or gradient calc. */
 	}
-	if (bright(acol) > FTINY) {	/* normalize Y values */
-		d = 0.99*cnt/bright(acol);
+	if ((d = bright(acol)) > FTINY) {	/* normalize Y values */
+		d = 0.99*(hp->ns*hp->ns)/d;
 		K = 0.01;
-	} else {			/* geometric Hessian fall-back */
-		d = 0.0;
+	} else {			/* or fall back on geometric Hessian */
 		K = 1.0;
 		pg = NULL;
 		dg = NULL;
+		crlp = NULL;
 	}
 	ap = hp->sa;			/* relative Y channel from here on... */
 	for (i = hp->ns*hp->ns; i--; ap++)
@@ -691,6 +737,9 @@ doambient(				/* compute ambient component */
 			if (ra[0] > maxarad)
 				ra[0] = maxarad;
 		}
+					/* flag encroached directions */
+		if ((wt >= 0.89*AVGREFL) & (crlp != NULL))
+			*crlp = ambcorral(hp, uv, ra[0]*ambacc, ra[1]*ambacc);
 		if (pg != NULL) {	/* cap gradient if necessary */
 			d = pg[0]*pg[0]*ra[0]*ra[0] + pg[1]*pg[1]*ra[1]*ra[1];
 			if (d > 1.0) {