--- ray/src/rt/rtrace.c	2012/06/09 07:16:47	2.66
+++ ray/src/rt/rtrace.c	2020/04/03 17:06:16	2.89
@@ -1,5 +1,5 @@
 #ifndef lint
-static const char	RCSid[] = "$Id: rtrace.c,v 2.66 2012/06/09 07:16:47 greg Exp $";
+static const char	RCSid[] = "$Id: rtrace.c,v 2.89 2020/04/03 17:06:16 greg Exp $";
 #endif
 /*
  *  rtrace.c - program and variables for individual ray tracing.
@@ -28,6 +28,7 @@ static const char	RCSid[] = "$Id: rtrace.c,v 2.66 2012
 #include  "ambient.h"
 #include  "source.h"
 #include  "otypes.h"
+#include  "otspecial.h"
 #include  "resolu.h"
 #include  "random.h"
 
@@ -53,12 +54,19 @@ OBJECT	traset[MAXTSET+1]={0};		/* trace include/exclud
 
 static RAY  thisray;			/* for our convenience */
 
+static FILE  *inpfp = NULL;		/* input stream pointer */
+
+static FVECT	*inp_queue = NULL;	/* ray input queue if flushing */
+static int	inp_qpos = 0;		/* next ray to return */
+static int	inp_qend = 0;		/* number of rays in this work group */
+
 typedef void putf_t(RREAL *v, int n);
-static putf_t puta, putd, putf;
+static putf_t puta, putd, putf, putrgbe;
 
 typedef void oputf_t(RAY *r);
 static oputf_t  oputo, oputd, oputv, oputV, oputl, oputL, oputc, oputp,
-		oputn, oputN, oputs, oputw, oputW, oputm, oputM, oputtilde;
+		oputr, oputR, oputx, oputX, oputn, oputN, oputs,
+		oputw, oputW, oputm, oputM, oputtilde;
 
 static void setoutput(char *vs);
 extern void tranotify(OBJECT obj);
@@ -68,10 +76,12 @@ static void rayirrad(RAY *r);
 static void rtcompute(FVECT org, FVECT dir, double dmax);
 static int printvals(RAY *r);
 static int getvec(FVECT vec, int fmt, FILE *fp);
+static int iszerovec(const FVECT vec);
+static double nextray(FVECT org, FVECT dir);
 static void tabin(RAY *r);
 static void ourtrace(RAY *r);
 
-static oputf_t *ray_out[16], *every_out[16];
+static oputf_t *ray_out[32], *every_out[32];
 static putf_t *putreal;
 
 
@@ -115,33 +125,42 @@ rtrace(				/* trace rays from file */
 {
 	unsigned long  vcount = (hresolu > 1) ? (unsigned long)hresolu*vresolu
 					      : (unsigned long)vresolu;
-	long  nextflush = (vresolu > 0) & (hresolu > 1) ? 0 : hresolu;
+	long  nextflush = (!vresolu | (hresolu <= 1)) * hresolu;
+	int  something2flush = 0;
 	FILE  *fp;
 	double	d;
 	FVECT  orig, direc;
 					/* set up input */
 	if (fname == NULL)
-		fp = stdin;
-	else if ((fp = fopen(fname, "r")) == NULL) {
+		inpfp = stdin;
+	else if ((inpfp = fopen(fname, "r")) == NULL) {
 		sprintf(errmsg, "cannot open input file \"%s\"", fname);
 		error(SYSTEM, errmsg);
 	}
+#ifdef getc_unlocked
+	flockfile(inpfp);		/* avoid lock/unlock overhead */
+	flockfile(stdout);
+#endif
 	if (inform != 'a')
-		SET_FILE_BINARY(fp);
+		SET_FILE_BINARY(inpfp);
 					/* set up output */
 	setoutput(outvals);
 	if (imm_irrad)
 		castonly = 0;
 	else if (castonly)
 		nproc = 1;		/* don't bother multiprocessing */
+	if ((nextflush > 0) & (nproc > nextflush)) {
+		error(WARNING, "reducing number of processes to match flush interval");
+		nproc = nextflush;
+	}
 	switch (outform) {
 	case 'a': putreal = puta; break;
 	case 'f': putreal = putf; break;
 	case 'd': putreal = putd; break;
 	case 'c': 
-		if (strcmp(outvals, "v"))
-			error(USER, "color format with value output only");
-		break;
+		if (outvals[0] && (outvals[1] || !strchr("vrx", outvals[0])))
+			error(USER, "color format only with -ov, -or, -ox");
+		putreal = putrgbe; break;
 	default:
 		error(CONSISTENCY, "botched output format");
 	}
@@ -152,49 +171,50 @@ rtrace(				/* trace rays from file */
 	if (hresolu > 0) {
 		if (vresolu > 0)
 			fprtresolu(hresolu, vresolu, stdout);
-		fflush(stdout);
+		else
+			fflush(stdout);
 	}
-					/* process file */
-	while (getvec(orig, inform, fp) == 0 &&
-			getvec(direc, inform, fp) == 0) {
-
-		d = normalize(direc);
-		if (d == 0.0) {				/* zero ==> flush */
-			if (--nextflush <= 0 || !vcount) {
-				if (nproc > 1 && ray_fifo_flush() < 0)
-					error(USER, "lost children");
+					/* process input rays */
+	while ((d = nextray(orig, direc)) >= 0.0) {
+		if (d == 0.0) {				/* flush request? */
+			if (something2flush) {
+				if (ray_pnprocs > 1 && ray_fifo_flush() < 0)
+					error(USER, "child(ren) died");
 				bogusray();
 				fflush(stdout);
-				nextflush = (vresolu > 0) & (hresolu > 1) ? 0 :
-								hresolu;
+				nextflush = (!vresolu | (hresolu <= 1)) * hresolu;
+				something2flush = 0;
 			} else
 				bogusray();
 		} else {				/* compute and print */
 			rtcompute(orig, direc, lim_dist ? d : 0.0);
-							/* flush if time */
-			if (!--nextflush) {
-				if (nproc > 1 && ray_fifo_flush() < 0)
-					error(USER, "lost children");
+			if (!--nextflush) {		/* flush if time */
+				if (ray_pnprocs > 1 && ray_fifo_flush() < 0)
+					error(USER, "child(ren) died");
 				fflush(stdout);
 				nextflush = hresolu;
-			}
+			} else
+				something2flush = 1;
 		}
 		if (ferror(stdout))
 			error(SYSTEM, "write error");
 		if (vcount && !--vcount)		/* check for end */
 			break;
 	}
-	if (nproc > 1) {				/* clean up children */
+	if (ray_pnprocs > 1) {				/* clean up children */
 		if (ray_fifo_flush() < 0)
 			error(USER, "unable to complete processing");
 		ray_pclose(0);
 	}
+	if (vcount)
+		error(WARNING, "unexpected EOF on input");
 	if (fflush(stdout) < 0)
 		error(SYSTEM, "write error");
-	if (vcount)
-		error(USER, "unexpected EOF on input");
-	if (fname != NULL)
-		fclose(fp);
+	if (fname != NULL) {
+		fclose(inpfp);
+		inpfp = NULL;
+	}
+	nextray(NULL, NULL);
 }
 
 
@@ -235,6 +255,22 @@ setoutput(				/* set up output tables */
 		case 'd':				/* direction */
 			*table++ = oputd;
 			break;
+		case 'r':				/* reflected contrib. */
+			*table++ = oputr;
+			castonly = 0;
+			break;
+		case 'R':				/* reflected distance */
+			*table++ = oputR;
+			castonly = 0;
+			break;
+		case 'x':				/* xmit contrib. */
+			*table++ = oputx;
+			castonly = 0;
+			break;
+		case 'X':				/* xmit distance */
+			*table++ = oputX;
+			castonly = 0;
+			break;
 		case 'v':				/* value */
 			*table++ = oputv;
 			castonly = 0;
@@ -273,7 +309,8 @@ setoutput(				/* set up output tables */
 			break;
 		case 'W':				/* coefficient */
 			*table++ = oputW;
-			if (ambounce > 0 && (ambacc > FTINY || ambssamp > 0))
+			castonly = 0;
+			if (ambounce > 0 && (ambacc > FTINY) | (ambssamp > 0))
 				error(WARNING,
 					"-otW accuracy depends on -aa 0 -as 0");
 			break;
@@ -288,15 +325,21 @@ setoutput(				/* set up output tables */
 			break;
 		}
 	*table = NULL;
+							/* compatibility */
+	for (table = ray_out; *table != NULL; table++) {
+		if ((*table == oputV) | (*table == oputW))
+			error(WARNING, "-oVW options require trace mode");
+		if ((do_irrad | imm_irrad) &&
+				(*table == oputr) | (*table == oputR) |
+				(*table == oputx) | (*table == oputX))
+			error(WARNING, "-orRxX options incompatible with -I+ and -i+");
+	}
 }
 
 
 static void
 bogusray(void)			/* print out empty record */
 {
-	thisray.rorg[0] = thisray.rorg[1] = thisray.rorg[2] =
-	thisray.rdir[0] = thisray.rdir[1] = thisray.rdir[2] = 0.0;
-	thisray.rmax = 0.0;
 	rayorigin(&thisray, PRIMARY, NULL, NULL);
 	printvals(&thisray);
 }
@@ -323,8 +366,8 @@ rayirrad(			/* compute irradiance rather than radiance
 )
 {
 	void	(*old_revf)(RAY *) = r->revf;
-
-	r->rot = 1e-5;			/* pretend we hit surface */
+					/* pretend we hit surface */
+	r->rxt = r->rot = 1e-5;
 	VSUM(r->rop, r->rorg, r->rdir, r->rot);
 	r->ron[0] = -r->rdir[0];
 	r->ron[1] = -r->rdir[1];
@@ -389,7 +432,7 @@ printvals(			/* print requested ray values */
 
 
 static int
-getvec(		/* get a vector from fp */
+getvec(			/* get a vector from fp */
 	FVECT  vec,
 	int  fmt,
 	FILE  *fp
@@ -410,12 +453,12 @@ getvec(		/* get a vector from fp */
 		}
 		break;
 	case 'f':					/* binary float */
-		if (fread((char *)vf, sizeof(float), 3, fp) != 3)
+		if (getbinary(vf, sizeof(float), 3, fp) != 3)
 			return(-1);
 		VCOPY(vec, vf);
 		break;
 	case 'd':					/* binary double */
-		if (fread((char *)vd, sizeof(double), 3, fp) != 3)
+		if (getbinary(vd, sizeof(double), 3, fp) != 3)
 			return(-1);
 		VCOPY(vec, vd);
 		break;
@@ -426,6 +469,64 @@ getvec(		/* get a vector from fp */
 }
 
 
+static int
+iszerovec(const FVECT vec)
+{
+	return (vec[0] == 0.0) & (vec[1] == 0.0) & (vec[2] == 0.0);
+}
+
+
+static double
+nextray(		/* return next ray in work group (-1.0 if EOF) */
+	FVECT org,
+	FVECT dir
+)
+{
+	const size_t	qlength = !vresolu * hresolu;
+
+	if ((org == NULL) | (dir == NULL)) {
+		if (inp_queue != NULL)	/* asking to free queue */
+			free(inp_queue);
+		inp_queue = NULL;
+		inp_qpos = inp_qend = 0;
+		return(-1.);
+	}
+	if (!inp_qend) {		/* initialize FIFO queue */
+		int	rsiz = 6*20;	/* conservative ascii ray size */
+		if (inform == 'f') rsiz = 6*sizeof(float);
+		else if (inform == 'd') rsiz = 6*sizeof(double);
+		if ((inpfp == stdin) & (qlength*rsiz > 512))	/* pipe limit */
+			inp_queue = (FVECT *)malloc(sizeof(FVECT)*2*qlength);
+		inp_qend = -(inp_queue == NULL);	/* flag for no queue */
+	}
+	if (inp_qend < 0) {		/* not queuing? */
+		if (getvec(org, inform, inpfp) < 0 ||
+				getvec(dir, inform, inpfp) < 0)
+			return(-1.);
+		return normalize(dir);
+	}
+	if (inp_qpos >= inp_qend) {	/* need to refill input queue? */
+		for (inp_qend = 0; inp_qend < qlength; inp_qend++) {
+			if (getvec(inp_queue[2*inp_qend], inform, inpfp) < 0
+					|| getvec(inp_queue[2*inp_qend+1],
+							inform, inpfp) < 0)
+				break;		/* hit EOF */
+			if (iszerovec(inp_queue[2*inp_qend+1])) {
+				++inp_qend;	/* flush request */
+				break;
+			}
+		}
+		inp_qpos = 0;
+	}
+	if (inp_qpos >= inp_qend)	/* unexpected EOF? */
+		return(-1.);
+	VCOPY(org, inp_queue[2*inp_qpos]);
+	VCOPY(dir, inp_queue[2*inp_qpos+1]);
+	++inp_qpos;
+	return normalize(dir);
+}
+
+
 void
 tranotify(			/* record new modifier */
 	OBJECT	obj
@@ -508,20 +609,59 @@ oputd(				/* print direction */
 
 
 static void
+oputr(				/* print mirrored contribution */
+	RAY  *r
+)
+{
+	RREAL	cval[3];
+
+	cval[0] = colval(r->mcol,RED);
+	cval[1] = colval(r->mcol,GRN);
+	cval[2] = colval(r->mcol,BLU);
+	(*putreal)(cval, 3);
+}
+
+
+
+static void
+oputR(				/* print mirrored distance */
+	RAY  *r
+)
+{
+	(*putreal)(&r->rmt, 1);
+}
+
+
+static void
+oputx(				/* print unmirrored contribution */
+	RAY  *r
+)
+{
+	RREAL	cval[3];
+
+	cval[0] = colval(r->rcol,RED) - colval(r->mcol,RED);
+	cval[1] = colval(r->rcol,GRN) - colval(r->mcol,GRN);
+	cval[2] = colval(r->rcol,BLU) - colval(r->mcol,BLU);
+	(*putreal)(cval, 3);
+}
+
+
+static void
+oputX(				/* print unmirrored distance */
+	RAY  *r
+)
+{
+	(*putreal)(&r->rxt, 1);
+}
+
+
+static void
 oputv(				/* print value */
 	RAY  *r
 )
 {
 	RREAL	cval[3];
 
-	if (outform == 'c') {
-		COLR  cout;
-		setcolr(cout,	colval(r->rcol,RED),
-				colval(r->rcol,GRN),
-				colval(r->rcol,BLU));
-		fwrite(cout, sizeof(cout), 1, stdout);
-		return;
-	}
 	cval[0] = colval(r->rcol,RED);
 	cval[1] = colval(r->rcol,GRN);
 	cval[2] = colval(r->rcol,BLU);
@@ -547,7 +687,9 @@ oputl(				/* print effective distance */
 	RAY  *r
 )
 {
-	(*putreal)(&r->rt, 1);
+	RREAL	d = raydistance(r);
+
+	(*putreal)(&d, 1);
 }
 
 
@@ -577,7 +719,7 @@ oputp(				/* print point */
 	RAY  *r
 )
 {
-	if (r->rot < FHUGE)
+	if (r->rot < FHUGE*.99)
 		(*putreal)(r->rop, 3);
 	else
 		(*putreal)(vdummy, 3);
@@ -589,9 +731,16 @@ oputN(				/* print unperturbed normal */
 	RAY  *r
 )
 {
-	if (r->rot < FHUGE)
-		(*putreal)(r->ron, 3);
-	else
+	if (r->rot < FHUGE*.99) {
+		if (r->rflips & 1) {	/* undo any flippin' flips */
+			FVECT	unrm;
+			unrm[0] = -r->ron[0];
+			unrm[1] = -r->ron[1];
+			unrm[2] = -r->ron[2];
+			(*putreal)(unrm, 3);
+		} else
+			(*putreal)(r->ron, 3);
+	} else
 		(*putreal)(vdummy, 3);
 }
 
@@ -603,7 +752,7 @@ oputn(				/* print perturbed normal */
 {
 	FVECT  pnorm;
 
-	if (r->rot >= FHUGE) {
+	if (r->rot >= FHUGE*.99) {
 		(*putreal)(vdummy, 3);
 		return;
 	}
@@ -642,8 +791,12 @@ oputW(				/* print coefficient */
 )
 {
 	RREAL	contr[3];
+				/* shadow ray not on source? */
+	if (r->rsrc >= 0 && source[r->rsrc].so != r->ro)
+		setcolor(contr, 0.0, 0.0, 0.0);
+	else
+		raycontrib(contr, r, PRIMARY);
 
-	raycontrib(contr, r, PRIMARY);
 	(*putreal)(contr, 3);
 }
 
@@ -706,17 +859,27 @@ puta(				/* print ascii value(s) */
 
 
 static void
-putd(RREAL *v, int n)		/* print binary double(s) */
+putd(RREAL *v, int n)		/* output binary double(s) */
 {
-	if (sizeof(RREAL) != sizeof(double))
+#ifdef	SMLFLT
+	double	da[3];
+	int	i;
+
+	if (n > 3)
 		error(INTERNAL, "code error in putd()");
-	fwrite(v, sizeof(RREAL), n, stdout);
+	for (i = n; i--; )
+		da[i] = v[i];
+	putbinary(da, sizeof(double), n, stdout);
+#else
+	putbinary(v, sizeof(RREAL), n, stdout);
+#endif
 }
 
 
 static void
-putf(RREAL *v, int n)		/* print binary float(s) */
+putf(RREAL *v, int n)		/* output binary float(s) */
 {
+#ifndef	SMLFLT
 	float	fa[3];
 	int	i;
 
@@ -724,5 +887,20 @@ putf(RREAL *v, int n)		/* print binary float(s) */
 		error(INTERNAL, "code error in putf()");
 	for (i = n; i--; )
 		fa[i] = v[i];
-	fwrite(fa, sizeof(float), n, stdout);
+	putbinary(fa, sizeof(float), n, stdout);
+#else
+	putbinary(v, sizeof(RREAL), n, stdout);
+#endif
+}
+
+
+static void
+putrgbe(RREAL *v, int n)	/* output RGBE color */
+{
+	COLR  cout;
+
+	if (n != 3)
+		error(INTERNAL, "putrgbe() not called with 3 components");
+	setcolr(cout, v[0], v[1], v[2]);
+	putbinary(cout, sizeof(cout), 1, stdout);
 }