--- ray/src/rt/ambcomp.c	2014/04/29 15:40:00	2.39
+++ ray/src/rt/ambcomp.c	2014/05/01 16:01:59	2.43
@@ -1,5 +1,5 @@
 #ifndef lint
-static const char	RCSid[] = "$Id: ambcomp.c,v 2.39 2014/04/29 15:40:00 greg Exp $";
+static const char	RCSid[] = "$Id: ambcomp.c,v 2.43 2014/05/01 16:01:59 greg Exp $";
 #endif
 /*
  * Routines to compute "ambient" values using Monte Carlo
@@ -32,8 +32,10 @@ typedef struct {
 	} sa[1];		/* sample array (extends struct) */
 }  AMBHEMI;		/* ambient sample hemisphere */
 
-#define ambsamp(h,i,j)	(h)->sa[(i)*(h)->ns + (j)]
+typedef struct s_ambsamp	AMBSAMP;
 
+#define ambsam(h,i,j)	(h)->sa[(i)*(h)->ns + (j)]
+
 typedef struct {
 	FVECT	r_i, r_i1, e_i, rcp, rI2_eJ2;
 	double	I1, I2;
@@ -59,8 +61,7 @@ inithemi(			/* initialize sampling hemisphere */
 	if (n < i)
 		n = i;
 					/* allocate sampling array */
-	hp = (AMBHEMI *)malloc(sizeof(AMBHEMI) +
-				sizeof(struct s_ambsamp)*(n*n - 1));
+	hp = (AMBHEMI *)malloc(sizeof(AMBHEMI) + sizeof(AMBSAMP)*(n*n - 1));
 	if (hp == NULL)
 		return(NULL);
 	hp->rp = r;
@@ -87,47 +88,66 @@ inithemi(			/* initialize sampling hemisphere */
 }
 
 
-static struct s_ambsamp *
-ambsample(				/* sample an ambient direction */
-	AMBHEMI	*hp,
-	int	i,
-	int	j
-)
+/* Sample ambient division and apply weighting coefficient */
+static int
+getambsamp(RAY *arp, AMBHEMI *hp, int i, int j, int n)
 {
-	struct s_ambsamp	*ap = &ambsamp(hp,i,j);
-	RAY			ar;
-	double			spt[2], zd;
-	int			ii;
+	int	hlist[3], ii;
+	double	spt[2], zd;
 					/* ambient coefficient for weight */
 	if (ambacc > FTINY)
-		setcolor(ar.rcoef, AVGREFL, AVGREFL, AVGREFL);
+		setcolor(arp->rcoef, AVGREFL, AVGREFL, AVGREFL);
 	else
-		copycolor(ar.rcoef, hp->acoef);
-	if (rayorigin(&ar, AMBIENT, hp->rp, ar.rcoef) < 0)
-		goto badsample;
+		copycolor(arp->rcoef, hp->acoef);
+	if (rayorigin(arp, AMBIENT, hp->rp, arp->rcoef) < 0)
+		return(0);
 	if (ambacc > FTINY) {
-		multcolor(ar.rcoef, hp->acoef);
-		scalecolor(ar.rcoef, 1./AVGREFL);
+		multcolor(arp->rcoef, hp->acoef);
+		scalecolor(arp->rcoef, 1./AVGREFL);
 	}
-					/* generate hemispherical sample */
-	SDsquare2disk(spt,	(i+.1+.8*frandom())/hp->ns,
-				(j+.1+.8*frandom())/hp->ns );
+	hlist[0] = hp->rp->rno;
+	hlist[1] = i;
+	hlist[2] = j;
+	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))
+			spt[0] = 0.1 + 0.8*frandom();
+		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);
 	zd = sqrt(1. - spt[0]*spt[0] - spt[1]*spt[1]);
 	for (ii = 3; ii--; )
-		ar.rdir[ii] =	spt[0]*hp->ux[ii] +
+		arp->rdir[ii] =	spt[0]*hp->ux[ii] +
 				spt[1]*hp->uy[ii] +
 				zd*hp->rp->ron[ii];
-	checknorm(ar.rdir);
+	checknorm(arp->rdir);
 	dimlist[ndims++] = i*hp->ns + j + 90171;
-	rayvalue(&ar);			/* evaluate ray */
-	ndims--;
+	rayvalue(arp);			/* evaluate ray */
+	ndims--;			/* apply coefficient */
+	multcolor(arp->rcol, arp->rcoef);
+	return(1);
+}
+
+
+static AMBSAMP *
+ambsample(				/* initial ambient division sample */
+	AMBHEMI	*hp,
+	int	i,
+	int	j
+)
+{
+	AMBSAMP	*ap = &ambsam(hp,i,j);
+	RAY	ar;
+					/* generate hemispherical sample */
+	if (!getambsamp(&ar, hp, i, j, 0))
+		goto badsample;
 					/* 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);
-	multcolor(ar.rcol, ar.rcoef);	/* apply coefficient */
 	copycolor(ap->v, ar.rcol);
 	return(ap);
 badsample:
@@ -137,6 +157,94 @@ badsample:
 }
 
 
+/* Estimate errors based on ambient division differences */
+static float *
+getambdiffs(AMBHEMI *hp)
+{
+	float	*earr = calloc(hp->ns*hp->ns, sizeof(float));
+	float	*ep;
+	AMBSAMP	*ap;
+	double	b, d2;
+	int	i, j;
+
+	if (earr == NULL)		/* out of memory? */
+		return(NULL);
+					/* compute squared neighbor diffs */
+	for (ap = hp->sa, ep = earr, i = 0; i < hp->ns; i++)
+	    for (j = 0; j < hp->ns; j++, ap++, ep++) {
+		b = bright(ap[0].v);
+		if (i) {		/* from above */
+			d2 = b - bright(ap[-hp->ns].v);
+			d2 *= d2;
+			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;
+		}
+	    }
+					/* 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;
+	for (i = 1; i < hp->ns-1; i++) {
+		earr[i*hp->ns] *= 4./3.;
+		earr[i*hp->ns + hp->ns-1] *= 4./3.;
+	}
+	for (j = 1; j < hp->ns-1; j++) {
+		earr[j] *= 4./3.;
+		earr[(hp->ns-1)*hp->ns + j] *= 4./3.;
+	}
+	return(earr);
+}
+
+
+/* Perform super-sampling on hemisphere (introduces bias) */
+static void
+ambsupersamp(double acol[3], AMBHEMI *hp, int cnt)
+{
+	float	*earr = getambdiffs(hp);
+	double	e2sum = 0;
+	AMBSAMP	*ap;
+	RAY	ar;
+	COLOR	asum;
+	float	*ep;
+	int	i, j, n;
+
+	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;
+	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.);
+		for (n = 1; n <= nss; n++) {
+			if (!getambsamp(&ar, hp, i, j, n)) {
+				nss = n-1;
+				break;
+			}
+			addcolor(asum, ar.rcol);
+		}
+		if (nss) {		/* update returned ambient value */
+			const double	ssf = 1./(nss + 1);
+			for (n = 3; n--; )
+				acol[n] += ssf*colval(asum,n) +
+						(ssf - 1.)*colval(ap->v,n);
+		}
+		e2sum -= *ep++;		/* update remainders */
+		cnt -= nss;
+	}
+	free(earr);
+}
+
+
 /* Compute vectors and coefficients for Hessian/gradient calcs */
 static void
 comp_fftri(FFTRI *ftp, FVECT ap0, FVECT ap1, FVECT rop)
@@ -278,8 +386,7 @@ add2gradient(FVECT grad, FVECT egrad1, FVECT egrad2, F
 
 /* Return brightness of furthest ambient sample */
 static COLORV
-back_ambval(struct s_ambsamp *ap1, struct s_ambsamp *ap2,
-		struct s_ambsamp *ap3, FVECT orig)
+back_ambval(AMBSAMP *ap1, AMBSAMP *ap2, AMBSAMP *ap3, FVECT orig)
 {
 	COLORV	vback;
 	FVECT	vec;
@@ -385,8 +492,8 @@ ambHessian(				/* anisotropic radii & pos. gradient */
 	}
 					/* compute first row of edges */
 	for (j = 0; j < hp->ns-1; j++) {
-		comp_fftri(&fftr, ambsamp(hp,0,j).p,
-				ambsamp(hp,0,j+1).p, hp->rp->rop);
+		comp_fftri(&fftr, ambsam(hp,0,j).p,
+				ambsam(hp,0,j+1).p, hp->rp->rop);
 		if (hessrow != NULL)
 			comp_hessian(hessrow[j], &fftr, hp->rp->ron);
 		if (gradrow != NULL)
@@ -396,8 +503,8 @@ 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, ambsamp(hp,i,0).p,
-			ambsamp(hp,i+1,0).p, hp->rp->rop);
+	    comp_fftri(&fftr, ambsam(hp,i,0).p,
+			ambsam(hp,i+1,0).p, hp->rp->rop);
 	    if (hessrow != NULL)
 		comp_hessian(hesscol, &fftr, hp->rp->ron);
 	    if (gradrow != NULL)
@@ -406,11 +513,11 @@ ambHessian(				/* anisotropic radii & pos. gradient */
 		FVECT	hessdia[3];	/* compute triangle contributions */
 		FVECT	graddia;
 		COLORV	backg;
-		backg = back_ambval(&ambsamp(hp,i,j), &ambsamp(hp,i,j+1),
-					&ambsamp(hp,i+1,j), hp->rp->rop);
+		backg = back_ambval(&ambsam(hp,i,j), &ambsam(hp,i,j+1),
+					&ambsam(hp,i+1,j), hp->rp->rop);
 					/* diagonal (inner) edge */
-		comp_fftri(&fftr, ambsamp(hp,i,j+1).p,
-				ambsamp(hp,i+1,j).p, hp->rp->rop);
+		comp_fftri(&fftr, ambsam(hp,i,j+1).p,
+				ambsam(hp,i+1,j).p, hp->rp->rop);
 		if (hessrow != NULL) {
 		    comp_hessian(hessdia, &fftr, hp->rp->ron);
 		    rev_hessian(hesscol);
@@ -422,16 +529,16 @@ ambHessian(				/* anisotropic radii & pos. gradient */
 		    add2gradient(gradient, gradrow[j], graddia, gradcol, backg);
 		}
 					/* initialize edge in next row */
-		comp_fftri(&fftr, ambsamp(hp,i+1,j+1).p,
-				ambsamp(hp,i+1,j).p, hp->rp->rop);
+		comp_fftri(&fftr, ambsam(hp,i+1,j+1).p,
+				ambsam(hp,i+1,j).p, hp->rp->rop);
 		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(&ambsamp(hp,i,j+1), &ambsamp(hp,i+1,j+1),
-					&ambsamp(hp,i+1,j), hp->rp->rop);
-		comp_fftri(&fftr, ambsamp(hp,i,j+1).p, ambsamp(hp,i+1,j+1).p,
+		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);
 		if (hessrow != NULL) {
 		    comp_hessian(hesscol, &fftr, hp->rp->ron);
@@ -466,11 +573,11 @@ ambHessian(				/* anisotropic radii & pos. gradient */
 static void
 ambdirgrad(AMBHEMI *hp, FVECT uv[2], float dg[2])
 {
-	struct s_ambsamp	*ap;
-	double			dgsum[2];
-	int			n;
-	FVECT			vd;
-	double			gfact;
+	AMBSAMP	*ap;
+	double	dgsum[2];
+	int	n;
+	FVECT	vd;
+	double	gfact;
 
 	dgsum[0] = dgsum[1] = 0.0;	/* sum values times -tan(theta) */
 	for (ap = hp->sa, n = hp->ns*hp->ns; n--; ap++) {
@@ -478,9 +585,9 @@ ambdirgrad(AMBHEMI *hp, FVECT uv[2], float dg[2])
 		VSUB(vd, ap->p, hp->rp->rop);
 					/* brightness over cosine factor */
 		gfact = colval(ap->v,CIEY) / DOT(hp->rp->ron, vd);
-					/* -sine = -proj_radius/vd_length */
-		dgsum[0] += DOT(uv[1], vd) * gfact;
-		dgsum[1] -= DOT(uv[0], vd) * gfact;
+					/* sine = proj_radius/vd_length */
+		dgsum[0] -= DOT(uv[1], vd) * gfact;
+		dgsum[1] += DOT(uv[0], vd) * gfact;
 	}
 	dg[0] = dgsum[0] / (hp->ns*hp->ns);
 	dg[1] = dgsum[1] / (hp->ns*hp->ns);
@@ -498,12 +605,12 @@ doambient(				/* compute ambient component */
 	float	dg[2]			/* returned (optional) */
 )
 {
-	AMBHEMI			*hp = inithemi(rcol, r, wt);
-	int			cnt = 0;
-	FVECT			my_uv[2];
-	double			d, K, acol[3];
-	struct s_ambsamp	*ap;
-	int			i, j;
+	AMBHEMI	*hp = inithemi(rcol, r, wt);
+	int	cnt = 0;
+	FVECT	my_uv[2];
+	double	d, K, acol[3];
+	AMBSAMP	*ap;
+	int	i, j;
 					/* check/initialize */
 	if (hp == NULL)
 		return(0);
@@ -528,9 +635,16 @@ doambient(				/* compute ambient component */
 		free(hp);
 		return(0);		/* no valid samples */
 	}
+	if (cnt < hp->ns*hp->ns) {	/* incomplete sampling? */
+		copycolor(rcol, acol);
+		free(hp);
+		return(-1);		/* return value w/o Hessian */
+	}
+	cnt = ambssamp*wt + 0.5;	/* perform super-sampling? */
+	if (cnt > 0)
+		ambsupersamp(acol, hp, cnt);
 	copycolor(rcol, acol);		/* final indirect irradiance/PI */
-	if (cnt < hp->ns*hp->ns ||	/* incomplete sampling? */
-			(ra == NULL) & (pg == NULL) & (dg == NULL)) {
+	if ((ra == NULL) & (pg == NULL) & (dg == NULL)) {
 		free(hp);
 		return(-1);		/* no radius or gradient calc. */
 	}