--- ray/src/util/rmatrix.c	2023/11/27 22:04:45	2.61
+++ ray/src/util/rmatrix.c	2023/12/12 18:45:53	2.76
@@ -1,5 +1,5 @@
 #ifndef lint
-static const char RCSid[] = "$Id: rmatrix.c,v 2.61 2023/11/27 22:04:45 greg Exp $";
+static const char RCSid[] = "$Id: rmatrix.c,v 2.76 2023/12/12 18:45:53 greg Exp $";
 #endif
 /*
  * General matrix operations.
@@ -31,14 +31,16 @@ rmx_new(int nr, int nc, int n)
 		return(NULL);
 
 	dnew = (RMATRIX *)calloc(1, sizeof(RMATRIX));
-	if (dnew) {
-		dnew->dtype = DTdouble;
-		dnew->nrows = nr;
-		dnew->ncols = nc;
-		dnew->ncomp = n;
-		setcolor(dnew->cexp, 1.f, 1.f, 1.f);
-		memcpy(dnew->wlpart, WLPART, sizeof(dnew->wlpart));
-	}
+	if (!dnew)
+		return(NULL);
+
+	dnew->dtype = DTdouble;
+	dnew->nrows = nr;
+	dnew->ncols = nc;
+	dnew->ncomp = n;
+	setcolor(dnew->cexp, 1.f, 1.f, 1.f);
+	memcpy(dnew->wlpart, WLPART, sizeof(dnew->wlpart));
+
 	return(dnew);
 }
 
@@ -68,19 +70,33 @@ rmx_alloc(int nr, int nc, int n)
 	return(dnew);
 }
 
-/* Free a RMATRIX array */
+/* Clear state by freeing info and matrix data */
 void
-rmx_free(RMATRIX *rm)
+rmx_reset(RMATRIX *rm)
 {
 	if (!rm) return;
-	if (rm->info)
+	if (rm->info) {
 		free(rm->info);
+		rm->info = NULL;
+	}
+	if (rm->mtx) {
 #ifdef MAP_FILE
-	if (rm->mapped)
-		munmap(rm->mapped, mapped_size(rm));
-	else
+		if (rm->mapped) {
+			munmap(rm->mapped, mapped_size(rm));
+			rm->mapped = NULL;
+		} else
 #endif
-		free(rm->mtx);
+			free(rm->mtx);
+		rm->mtx = NULL;
+	}
+}
+
+/* Free an RMATRIX struct and data */
+void
+rmx_free(RMATRIX *rm)
+{
+	if (!rm) return;
+	rmx_reset(rm);
 	free(rm);
 }
 
@@ -172,226 +188,227 @@ get_dminfo(char *s, void *p)
 }
 
 static int
-rmx_load_ascii(RMATRIX *rm, FILE *fp)
+rmx_load_ascii(double *drp, const RMATRIX *rm, FILE *fp)
 {
-	int	i, j, k;
+	int	j, k;
 
-	if (!rmx_prepare(rm))
-		return(0);
-	for (i = 0; i < rm->nrows; i++)
-	    for (j = 0; j < rm->ncols; j++) {
-	    	double	*dp = rmx_lval(rm,i,j);
-	        for (k = 0; k < rm->ncomp; k++)
-		    if (fscanf(fp, "%lf", &dp[k]) != 1)
-			return(0);
-	    }
+	for (j = 0; j < rm->ncols; j++)
+	        for (k = rm->ncomp; k-- > 0; )
+			if (fscanf(fp, "%lf", drp++) != 1)
+				return(0);
 	return(1);
 }
 
 static int
-rmx_load_float(RMATRIX *rm, FILE *fp)
+rmx_load_float(double *drp, const RMATRIX *rm, FILE *fp)
 {
-	int	i, j, k;
+	int	j, k;
 	float	val[100];
 
 	if (rm->ncomp > 100) {
 		fputs("Unsupported # components in rmx_load_float()\n", stderr);
 		exit(1);
 	}
-	if (!rmx_prepare(rm))
-		return(0);
-	for (i = 0; i < rm->nrows; i++)
-	    for (j = 0; j < rm->ncols; j++) {
-	    	double	*dp = rmx_lval(rm,i,j);
+	for (j = 0; j < rm->ncols; j++) {
 		if (getbinary(val, sizeof(val[0]), rm->ncomp, fp) != rm->ncomp)
-		    return(0);
+			return(0);
 		if (rm->swapin)
-		    swap32((char *)val, rm->ncomp);
-	        for (k = rm->ncomp; k--; )
-		     dp[k] = val[k];
-	    }
+			swap32((char *)val, rm->ncomp);
+	        for (k = 0; k < rm->ncomp; k++)
+			*drp++ = val[k];
+	}
 	return(1);
 }
 
 static int
-rmx_load_double(RMATRIX *rm, FILE *fp)
+rmx_load_double(double *drp, const RMATRIX *rm, FILE *fp)
 {
-	int	i;
-#ifdef MAP_FILE
-	long	pos;		/* map memory for file > 1MB if possible */
-	if (!rm->swapin && array_size(rm) >= 1L<<20 &&
-			(pos = ftell(fp)) >= 0 && !(pos % sizeof(double))) {
-		rm->mapped = mmap(NULL, array_size(rm)+pos, PROT_READ|PROT_WRITE,
-					MAP_PRIVATE, fileno(fp), 0);
-		if (rm->mapped != MAP_FAILED) {
-			rm->mtx = (double *)rm->mapped + pos/sizeof(double);
-			return(1);
-		}		/* else fall back on reading into memory */
-		rm->mapped = NULL;
-	}
-#endif
-	if (!rmx_prepare(rm))
+	if (getbinary(drp, sizeof(*drp)*rm->ncomp, rm->ncols, fp) != rm->ncols)
 		return(0);
-	for (i = 0; i < rm->nrows; i++) {
-		if (getbinary(rmx_lval(rm,i,0), sizeof(double)*rm->ncomp,
-					rm->ncols, fp) != rm->ncols)
-			return(0);
-		if (rm->swapin)
-			swap64((char *)rmx_lval(rm,i,0), rm->ncols*rm->ncomp);
-	}
+	if (rm->swapin)
+		swap64((char *)drp, rm->ncols*rm->ncomp);
 	return(1);
 }
 
 static int
-rmx_load_rgbe(RMATRIX *rm, FILE *fp)
+rmx_load_rgbe(double *drp, const RMATRIX *rm, FILE *fp)
 {
-	COLOR	*scan = (COLOR *)malloc(sizeof(COLOR)*rm->ncols);
-	int	i, j;
+	COLR	*scan;
+	COLOR	col;
+	int	j;
 
+	if (rm->ncomp != 3)
+		return(0);
+	scan = (COLR *)tempbuffer(sizeof(COLR)*rm->ncols);
 	if (!scan)
 		return(0);
-	if (!rmx_prepare(rm)) {
-		free(scan);
+	if (freadcolrs(scan, rm->ncols, fp) < 0)
 		return(0);
+	for (j = 0; j < rm->ncols; j++) {
+		colr_color(col, scan[j]);
+		*drp++ = colval(col,RED);
+		*drp++ = colval(col,GRN);
+		*drp++ = colval(col,BLU);
 	}
-	for (i = 0; i < rm->nrows; i++) {
-	    double	*dp = rmx_lval(rm,i,0);
-	    if (freadscan(scan, rm->ncols, fp) < 0) {
-		free(scan);
-		return(0);
-	    }
-	    for (j = 0; j < rm->ncols; j++, dp += 3) {
-	        dp[0] = colval(scan[j],RED);
-	        dp[1] = colval(scan[j],GRN);
-	        dp[2] = colval(scan[j],BLU);
-	    }
-	}
-	free(scan);
 	return(1);
 }
 
 static int
-rmx_load_spec(RMATRIX *rm, FILE *fp)
+rmx_load_spec(double *drp, const RMATRIX *rm, FILE *fp)
 {
 	uby8	*scan;
 	SCOLOR	scol;
-	int	i, j, k;
+	int	j, k;
 
-	if (rm->ncomp < 3)
+	if ((rm->ncomp < 3) | (rm->ncomp > MAXCSAMP))
 		return(0);
-	scan = (uby8 *)malloc((rm->ncomp+1)*rm->ncols);
+	scan = (uby8 *)tempbuffer((rm->ncomp+1)*rm->ncols);
 	if (!scan)
 		return(0);
-	if (!rmx_prepare(rm)) {
-		free(scan);
+	if (freadscolrs(scan, rm->ncomp, rm->ncols, fp) < 0)
 		return(0);
+	for (j = 0; j < rm->ncols; j++) {
+		scolr2scolor(scol, scan+j*(rm->ncomp+1), rm->ncomp);
+		for (k = 0; k < rm->ncomp; k++)
+			*drp++ = scol[k];
 	}
-	for (i = 0; i < rm->nrows; i++) {
-	    double	*dp = rmx_lval(rm,i,0);
-	    if (freadscolrs(scan, rm->ncomp, rm->ncols, fp) < 0) {
-		free(scan);
+	return(1);
+}
+
+/* Read matrix header from input stream (cannot be XML) */
+int
+rmx_load_header(RMATRIX *rm, FILE *fp)
+{
+	if (!rm | !fp)
 		return(0);
-	    }
-	    for (j = 0; j < rm->ncols; j++) {
-	    	scolr2scolor(scol, scan+j*(rm->ncomp+1), rm->ncomp);
-	    	for (k = 0; k < rm->ncomp; k++)
-	    		*dp++ = scol[k];
-	    }
+	rmx_reset(rm);				/* clear state */
+	if (rm->nrows | rm->ncols | !rm->dtype) {
+		rm->nrows = rm->ncols = 0;
+		rm->ncomp = 3;
+		setcolor(rm->cexp, 1.f, 1.f, 1.f);
+		memcpy(rm->wlpart, WLPART, sizeof(rm->wlpart));
+		rm->swapin = 0;
 	}
-	free(scan);
+	rm->dtype = DTascii;			/* assumed w/o FORMAT */
+	if (getheader(fp, get_dminfo, rm) < 0) {
+		fputs("Unrecognized matrix format\n", stderr);
+		return(0);
+	}
+	if ((rm->dtype == DTrgbe) | (rm->dtype == DTxyze) &&
+			rm->ncomp != 3)
+		return(0);
+	if (rm->ncols <= 0 &&			/* resolution string? */
+			!fscnresolu(&rm->ncols, &rm->nrows, fp))
+		return(0);
+	if (rm->dtype == DTascii)		/* set file type (WINDOWS) */
+		SET_FILE_TEXT(fp);
+	else
+		SET_FILE_BINARY(fp);
 	return(1);
 }
 
+/* Load next row as double (cannot be XML) */
+int
+rmx_load_row(double *drp, const RMATRIX *rm, FILE *fp)
+{
+	switch (rm->dtype) {
+	case DTascii:
+		return(rmx_load_ascii(drp, rm, fp));
+	case DTfloat:
+		return(rmx_load_float(drp, rm, fp));
+	case DTdouble:
+		return(rmx_load_double(drp, rm, fp));
+	case DTrgbe:
+	case DTxyze:
+		return(rmx_load_rgbe(drp, rm, fp));
+	case DTspec:
+		return(rmx_load_spec(drp, rm, fp));
+	default:
+		fputs("Unsupported data type in rmx_load_row()\n", stderr);
+	}
+	return(0);
+}
+
+/* Allocate & load post-header data from stream given type set in rm->dtype */
+int
+rmx_load_data(RMATRIX *rm, FILE *fp)
+{
+	int	i;
+#ifdef MAP_FILE
+	long	pos;		/* map memory for file > 1MB if possible */
+	if ((rm->dtype == DTdouble) & !rm->swapin && array_size(rm) >= 1L<<20 &&
+			(pos = ftell(fp)) >= 0 && !(pos % sizeof(double))) {
+		rm->mapped = mmap(NULL, array_size(rm)+pos, PROT_READ|PROT_WRITE,
+					MAP_PRIVATE, fileno(fp), 0);
+		if (rm->mapped != MAP_FAILED) {
+			rm->mtx = (double *)rm->mapped + pos/sizeof(double);
+			return(1);
+		}		/* else fall back on reading into memory */
+		rm->mapped = NULL;
+	}
+#endif
+	if (!rmx_prepare(rm)) {	/* need in-core matrix array */
+		fprintf(stderr, "Cannot allocate %g MByte matrix array\n",
+				(1./(1L<<20))*(double)array_size(rm));
+		return(0);
+	}
+	for (i = 0; i < rm->nrows; i++)
+		if (!rmx_load_row(rmx_lval(rm,i,0), rm, fp))
+			return(0);
+	return(1);
+}
+
 /* Load matrix from supported file type */
 RMATRIX *
 rmx_load(const char *inspec, RMPref rmp)
 {
 	FILE		*fp;
 	RMATRIX		*dnew;
+	int		ok;
 
 	if (!inspec)
 		inspec = stdin_name;
 	else if (!*inspec)
 		return(NULL);
-	if (inspec == stdin_name) {		/* reading from stdin? */
+	if (inspec == stdin_name)		/* reading from stdin? */
 		fp = stdin;
-	} else if (inspec[0] == '!') {
-		if (!(fp = popen(inspec+1, "r")))
-			return(NULL);
-	} else {
+	else if (inspec[0] == '!')
+		fp = popen(inspec+1, "r");
+	else {
 		const char	*sp = inspec;	/* check suffix */
 		while (*sp)
 			++sp;
 		while (sp > inspec && sp[-1] != '.')
 			--sp;
 		if (!strcasecmp(sp, "XML")) {	/* assume it's a BSDF */
-			CMATRIX	*cm = rmp==RMPtrans ? cm_loadBTDF(inspec) :
+			CMATRIX	*cm = rmp==RMPnone ? (CMATRIX *)NULL :
+					rmp==RMPtrans ? cm_loadBTDF(inspec) :
 					cm_loadBRDF(inspec, rmp==RMPreflB) ;
 			if (!cm)
 				return(NULL);
 			dnew = rmx_from_cmatrix(cm);
 			cm_free(cm);
 			dnew->dtype = DTascii;
-			return(dnew);
-		}
-						/* else open it ourselves */
-		if (!(fp = fopen(inspec, "r")))
-			return(NULL);
+			return(dnew);		/* return here */
+		}				/* else open it ourselves */
+		fp = fopen(inspec, "r");
 	}
-	SET_FILE_BINARY(fp);
+	if (!fp)
+		return(NULL);
 #ifdef getc_unlocked
 	flockfile(fp);
 #endif
-	if (!(dnew = rmx_new(0,0,3))) {
-		fclose(fp);
+	SET_FILE_BINARY(fp);			/* load header info */
+	if (!rmx_load_header(dnew = rmx_new(0,0,3), fp)) {
+		fprintf(stderr, "Bad header in: %s\n", inspec);
+		if (inspec[0] == '!') pclose(fp);
+		else fclose(fp);
+		rmx_free(dnew);
 		return(NULL);
 	}
-	dnew->dtype = DTascii;			/* assumed w/o FORMAT */
-	if (getheader(fp, get_dminfo, dnew) < 0) {
-		fclose(fp);
-		return(NULL);
-	}
-	if ((dnew->nrows <= 0) | (dnew->ncols <= 0)) {
-		if (!fscnresolu(&dnew->ncols, &dnew->nrows, fp)) {
-			fclose(fp);
-			return(NULL);
-		}
-		if ((dnew->dtype == DTrgbe) | (dnew->dtype == DTxyze) &&
-				dnew->ncomp != 3) {
-			fclose(fp);
-			return(NULL);
-		}
-	}
-	switch (dnew->dtype) {
-	case DTascii:
-		SET_FILE_TEXT(fp);
-		if (!rmx_load_ascii(dnew, fp))
-			goto loaderr;
-		dnew->dtype = DTascii;		/* should leave double? */
-		break;
-	case DTfloat:
-		if (!rmx_load_float(dnew, fp))
-			goto loaderr;
-		dnew->dtype = DTfloat;
-		break;
-	case DTdouble:
-		if (!rmx_load_double(dnew, fp))
-			goto loaderr;
-		dnew->dtype = DTdouble;
-		break;
-	case DTrgbe:
-	case DTxyze:
-		if (!rmx_load_rgbe(dnew, fp))
-			goto loaderr;
-		break;
-	case DTspec:
-		if (!rmx_load_spec(dnew, fp))
-			goto loaderr;
-		break;
-	default:
-		goto loaderr;
-	}
-	if (fp != stdin) {
+	ok = rmx_load_data(dnew, fp);		/* allocate & load data */
+
+	if (fp != stdin) {			/* close input stream */
 		if (inspec[0] == '!')
 			pclose(fp);
 		else
@@ -401,6 +418,11 @@ rmx_load(const char *inspec, RMPref rmp)
 	else
 		funlockfile(fp);
 #endif
+	if (!ok) {				/* load failure? */
+		fprintf(stderr, "Error loading data from: %s\n", inspec);
+		rmx_free(dnew);
+		return(NULL);
+	}
 						/* undo exposure? */
 	if ((dnew->cexp[0] != 1.f) |
 			(dnew->cexp[1] != 1.f) | (dnew->cexp[2] != 1.f)) {
@@ -409,125 +431,80 @@ rmx_load(const char *inspec, RMPref rmp)
 		cmlt[0] = 1./dnew->cexp[0];
 		cmlt[1] = 1./dnew->cexp[1];
 		cmlt[2] = 1./dnew->cexp[2];
-		if (dnew->ncomp > MAXCSAMP)
-			goto loaderr;
+		if (dnew->ncomp > MAXCSAMP) {
+			fprintf(stderr, "Excess spectral components in: %s\n",
+					inspec);
+			rmx_free(dnew);
+			return(NULL);
+		}
 		for (i = dnew->ncomp; i-- > 3; )
 			cmlt[i] = cmlt[1];
 		rmx_scale(dnew, cmlt);
 		setcolor(dnew->cexp, 1.f, 1.f, 1.f);
 	}
 	return(dnew);
-loaderr:					/* should report error? */
-	if (inspec[0] == '!')
-		pclose(fp);
-	else
-		fclose(fp);
-	rmx_free(dnew);
-	return(NULL);
 }
 
 static int
-rmx_write_ascii(const RMATRIX *rm, FILE *fp)
+rmx_write_ascii(const double *dp, int nc, int len, FILE *fp)
 {
-	const char	*fmt = (rm->dtype == DTfloat) ? " %.7e" :
-			(rm->dtype == DTrgbe) | (rm->dtype == DTxyze) |
-			(rm->dtype == DTspec) ? " %.3e" : " %.15e" ;
-	int	i, j, k;
-
-	for (i = 0; i < rm->nrows; i++) {
-	    for (j = 0; j < rm->ncols; j++) {
-		const double	*dp = rmx_lval(rm,i,j);
-	        for (k = 0; k < rm->ncomp; k++)
-		    fprintf(fp, fmt, dp[k]);
+	while (len-- > 0) {
+		int	k = nc;
+		while (k-- > 0)
+			fprintf(fp, " %.7e", *dp++);
 		fputc('\t', fp);
-	    }
-	    fputc('\n', fp);
 	}
-	return(1);
+	return(fputc('\n', fp) != EOF);
 }
 
 static int
-rmx_write_float(const RMATRIX *rm, FILE *fp)
+rmx_write_float(const double *dp, int len, FILE *fp)
 {
-	int	i, j, k;
-	float	val[100];
+	float	val;
 
-	if (rm->ncomp > 100) {
-		fputs("Unsupported # components in rmx_write_float()\n", stderr);
-		exit(1);
-	}
-	for (i = 0; i < rm->nrows; i++)
-	    for (j = 0; j < rm->ncols; j++) {
-	    	const double	*dp = rmx_lval(rm,i,j);
-	        for (k = rm->ncomp; k--; )
-		    val[k] = (float)dp[k];
-		if (putbinary(val, sizeof(float), rm->ncomp, fp) != rm->ncomp)
+	while (len--) {
+		val = *dp++;
+		if (putbinary(&val, sizeof(float), 1, fp) != 1)
 			return(0);
-	    }
+	}
 	return(1);
 }
 
 static int
-rmx_write_double(const RMATRIX *rm, FILE *fp)
+rmx_write_rgbe(const double *dp, int nc, int len, FILE *fp)
 {
-	int	i;
+	COLR	*scan;
+	int	j;
 
-	for (i = 0; i < rm->nrows; i++)
-		if (putbinary(rmx_lval(rm,i,0), sizeof(double)*rm->ncomp,
-					rm->ncols, fp) != rm->ncols)
-			return(0);
-	return(1);
-}
+	if ((nc != 1) & (nc != 3)) return(0);
+	scan = (COLR *)tempbuffer(sizeof(COLR)*len);
+	if (!scan) return(0);
 
-static int
-rmx_write_rgbe(const RMATRIX *rm, FILE *fp)
-{
-	COLR	*scan = (COLR *)malloc(sizeof(COLR)*rm->ncols);
-	int	i, j;
-
-	if (!scan)
-		return(0);
-	for (i = 0; i < rm->nrows; i++) {
-	    for (j = rm->ncols; j--; ) {
-	    	const double	*dp = rmx_lval(rm,i,j);
-	    	if (rm->ncomp == 1)
+	for (j = 0; j < len; j++, dp += nc)
+	    	if (nc == 1)
 	    		setcolr(scan[j], dp[0], dp[0], dp[0]);
 	    	else
 	        	setcolr(scan[j], dp[0], dp[1], dp[2]);
-	    }
-	    if (fwritecolrs(scan, rm->ncols, fp) < 0) {
-		free(scan);
-		return(0);
-	    }
-	}
-	free(scan);
-	return(1);
+
+	return(fwritecolrs(scan, len, fp) >= 0);
 }
 
 static int
-rmx_write_spec(const RMATRIX *rm, FILE *fp)
+rmx_write_spec(const double *dp, int nc, int len, FILE *fp)
 {
-	uby8	*scan = (uby8 *)malloc((rm->ncomp+1)*rm->ncols);
-	int	ok = 1;
+	uby8	*scan;
 	SCOLOR	scol;
-	int	i, j, k;
+	int	j, k;
 
-	if (!scan)
-		return(0);
-	for (i = 0; i < rm->nrows; i++) {
-	    for (j = rm->ncols; j--; ) {
-	    	const double	*dp = rmx_lval(rm,i,j);
-	    	for (k = rm->ncomp; k--; )
+	if (nc < 3) return(0);
+	scan = (uby8 *)tempbuffer((nc+1)*len);
+	if (!scan) return(0);
+	for (j = len; j--; dp += nc) {
+	    	for (k = nc; k--; )
 	    		scol[k] = dp[k];
-		scolor2scolr(scan+j*(rm->ncomp+1), scol, rm->ncomp);
-	    }
-	    if (fwritescolrs(scan, rm->ncomp, rm->ncols, fp) < 0) {
-		ok = 0;
-		break;
-	    }
+		scolor2scolr(scan+j*(nc+1), scol, nc);
 	}
-	free(scan);
-	return(ok);
+	return(fwritescolrs(scan, nc, len, fp) >= 0);
 }
 
 /* Check if CIE XYZ primaries were specified */
@@ -547,18 +524,12 @@ findCIEprims(const char *info)
 			(prims[BLU][CIEX] < .01) & (prims[BLU][CIEY] < .01));
 }
 
-/* Write matrix to file type indicated by dtype */
+/* Finish writing header data with resolution and format, returning type used */
 int
-rmx_write(const RMATRIX *rm, int dtype, FILE *fp)
+rmx_write_header(const RMATRIX *rm, int dtype, FILE *fp)
 {
-	int	ok = 1;
-
-	if (!rm | !fp || !rm->mtx)
+	if (!rm | !fp || rm->ncols <= 0)
 		return(0);
-#ifdef getc_unlocked
-	flockfile(fp);
-#endif
-						/* complete header */
 	if (rm->info)
 		fputs(rm->info, fp);
 	if (dtype == DTfromHeader)
@@ -568,54 +539,91 @@ rmx_write(const RMATRIX *rm, int dtype, FILE *fp)
 		dtype = DTxyze;
 	else if ((dtype == DTxyze) & (rm->dtype == DTrgbe))
 		dtype = DTrgbe;
+	if ((dtype < DTspec) & (rm->ncomp > 3))
+		dtype = DTspec;
+	else if ((dtype == DTspec) & (rm->ncomp <= 3))
+		return(0);
+
+	if (dtype == DTascii)			/* set file type (WINDOWS) */
+		SET_FILE_TEXT(fp);
+	else
+		SET_FILE_BINARY(fp);
 						/* write exposure? */
 	if (rm->ncomp == 3 && (rm->cexp[RED] != rm->cexp[GRN]) |
 			(rm->cexp[GRN] != rm->cexp[BLU]))
 		fputcolcor(rm->cexp, fp);
 	else if (rm->cexp[GRN] != 1.f)
 		fputexpos(rm->cexp[GRN], fp);
-	if ((dtype != DTrgbe) & (dtype != DTxyze)) {
-		if (dtype == DTspec) {
-			if (rm->ncomp < 3)
-				return(0);	/* bad # components */
-			fputwlsplit(rm->wlpart, fp);
-		} else {
+						/* matrix size? */
+	if ((dtype > DTspec) | (rm->nrows <= 0)) {
+		if (rm->nrows > 0)
 			fprintf(fp, "NROWS=%d\n", rm->nrows);
-			fprintf(fp, "NCOLS=%d\n", rm->ncols);
-		}
+		fprintf(fp, "NCOLS=%d\n", rm->ncols);
+	}
+	if (dtype >= DTspec) {			/* # components & split? */
 		fputncomp(rm->ncomp, fp);
+		if (rm->ncomp > 3 &&
+				memcmp(rm->wlpart, WLPART, sizeof(WLPART)))
+			fputwlsplit(rm->wlpart, fp);
 	} else if ((rm->ncomp != 3) & (rm->ncomp != 1))
 		return(0);			/* wrong # components */
 	if ((dtype == DTfloat) | (dtype == DTdouble))
 		fputendian(fp);			/* important to record */
 	fputformat(cm_fmt_id[dtype], fp);
-	fputc('\n', fp);
-	switch (dtype) {			/* write data */
+	fputc('\n', fp);			/* end of header */
+	if ((dtype <= DTspec) & (rm->nrows > 0))
+		fprtresolu(rm->ncols, rm->nrows, fp);
+	return(dtype);
+}
+
+/* Write out matrix data (usually by row) */
+int
+rmx_write_data(const double *dp, int nc, int len, int dtype, FILE *fp)
+{
+	switch (dtype) {
 	case DTascii:
-		ok = rmx_write_ascii(rm, fp);
-		break;
+		return(rmx_write_ascii(dp, nc, len, fp));
 	case DTfloat:
-		ok = rmx_write_float(rm, fp);
-		break;
+		return(rmx_write_float(dp, nc*len, fp));
 	case DTdouble:
-		ok = rmx_write_double(rm, fp);
-		break;
+		return(putbinary(dp, sizeof(*dp)*nc, len, fp) == len);
 	case DTrgbe:
 	case DTxyze:
-		fprtresolu(rm->ncols, rm->nrows, fp);
-		ok = rmx_write_rgbe(rm, fp);
-		break;
+		return(rmx_write_rgbe(dp, nc, len, fp));
 	case DTspec:
-		fprtresolu(rm->ncols, rm->nrows, fp);
-		ok = rmx_write_spec(rm, fp);
-		break;
-	default:
-		return(0);
+		return(rmx_write_spec(dp, nc, len, fp));
 	}
-	ok &= (fflush(fp) == 0);
+	return(0);
+}
+
+/* Write matrix using file format indicated by dtype */
+int
+rmx_write(const RMATRIX *rm, int dtype, FILE *fp)
+{
+	int	ok = 0;
+	int	i;
+						/* complete header */
+	dtype = rmx_write_header(rm, dtype, fp);
+	if (dtype <= 0)
+		return(0);
 #ifdef getc_unlocked
+	flockfile(fp);
+#endif
+	if (dtype == DTdouble)			/* write all at once? */
+		ok = rmx_write_data(rm->mtx, rm->ncomp,
+				rm->nrows*rm->ncols, dtype, fp);
+	else					/* else row by row */
+		for (i = 0; i < rm->nrows; i++) {
+			ok = rmx_write_data(rmx_val(rm,i,0), rm->ncomp,
+					rm->ncols, dtype, fp);
+			if (!ok) break;
+		}
+
+	if (ok) ok = (fflush(fp) == 0);
+#ifdef getc_unlocked
 	funlockfile(fp);
 #endif
+	if (!ok) fputs("Error writing matrix\n", stderr);
 	return(ok);
 }
 
@@ -637,7 +645,7 @@ rmx_identity(const int dim, const int n)
 	return(rid);
 }
 
-/* Duplicate the given matrix */
+/* Duplicate the given matrix (may be unallocated) */
 RMATRIX *
 rmx_copy(const RMATRIX *rm)
 {
@@ -645,17 +653,51 @@ rmx_copy(const RMATRIX *rm)
 
 	if (!rm)
 		return(NULL);
-	dnew = rmx_alloc(rm->nrows, rm->ncols, rm->ncomp);
+	dnew = rmx_new(rm->nrows, rm->ncols, rm->ncomp);
 	if (!dnew)
 		return(NULL);
+	if (rm->mtx) {
+		if (!rmx_prepare(dnew)) {
+			rmx_free(dnew);
+			return(NULL);
+		}
+		memcpy(dnew->mtx, rm->mtx, array_size(dnew));
+	}
 	rmx_addinfo(dnew, rm->info);
 	dnew->dtype = rm->dtype;
 	copycolor(dnew->cexp, rm->cexp);
 	memcpy(dnew->wlpart, rm->wlpart, sizeof(dnew->wlpart));
-	memcpy(dnew->mtx, rm->mtx, array_size(dnew));
 	return(dnew);
 }
 
+/* Replace data in first matrix with data from second */
+int
+rmx_transfer_data(RMATRIX *rdst, RMATRIX *rsrc, int dometa)
+{
+	if (!rdst | !rsrc || (rdst->nrows != rsrc->nrows) |
+			(rdst->ncols != rsrc->ncols) |
+			(rdst->ncomp != rsrc->ncomp))
+		return(0);
+
+	if (dometa) {		/* transfer everything? */
+		rmx_reset(rdst);
+		*rdst = *rsrc;
+		rsrc->info = NULL; rsrc->mapped = NULL; rsrc->mtx = NULL;
+		return(1);
+	}
+#ifdef MAP_FILE			/* just matrix data -- leave metadata */
+	if (rdst->mapped)
+		munmap(rdst->mapped, mapped_size(rdst));
+	else
+#endif
+	if (rdst->mtx)
+		free(rdst->mtx);
+	rdst->mapped = rsrc->mapped;
+	rdst->mtx = rsrc->mtx;
+	rsrc->mapped = NULL; rsrc->mtx = NULL;
+	return(1);
+}
+
 /* Allocate and assign transposed matrix */
 RMATRIX *
 rmx_transpose(const RMATRIX *rm)
@@ -663,7 +705,7 @@ rmx_transpose(const RMATRIX *rm)
 	RMATRIX	*dnew;
 	int	i, j;
 
-	if (!rm)
+	if (!rm || !rm->mtx)
 		return(0);
 	if ((rm->nrows == 1) | (rm->ncols == 1)) {
 		dnew = rmx_copy(rm);
@@ -685,7 +727,7 @@ rmx_transpose(const RMATRIX *rm)
 	memcpy(dnew->wlpart, rm->wlpart, sizeof(dnew->wlpart));
 	for (j = dnew->ncols; j--; )
 	    for (i = dnew->nrows; i--; )
-	    	memcpy(rmx_lval(dnew,i,j), rmx_lval(rm,j,i),
+	    	memcpy(rmx_lval(dnew,i,j), rmx_val(rm,j,i),
 	    			sizeof(double)*dnew->ncomp);
 	return(dnew);
 }
@@ -697,7 +739,8 @@ rmx_multiply(const RMATRIX *m1, const RMATRIX *m2)
 	RMATRIX	*mres;
 	int	i, j, k, h;
 
-	if (!m1 | !m2 || (m1->ncomp != m2->ncomp) | (m1->ncols != m2->nrows))
+	if (!m1 | !m2 || !m1->mtx | !m2->mtx |
+			(m1->ncomp != m2->ncomp) | (m1->ncols != m2->nrows))
 		return(NULL);
 	mres = rmx_alloc(m1->nrows, m2->ncols, m1->ncomp);
 	if (!mres)
@@ -712,7 +755,7 @@ rmx_multiply(const RMATRIX *m1, const RMATRIX *m2)
 	        for (k = mres->ncomp; k--; ) {
 		    double	d = 0;
 		    for (h = m1->ncols; h--; )
-			d += rmx_lval(m1,i,h)[k] * rmx_lval(m2,h,j)[k];
+			d += rmx_val(m1,i,h)[k] * rmx_val(m2,h,j)[k];
 		    rmx_lval(mres,i,j)[k] = d;
 		}
 	return(mres);
@@ -725,7 +768,8 @@ rmx_elemult(RMATRIX *m1, const RMATRIX *m2, int divide
 	int	zeroDivides = 0;
 	int	i, j, k;
 
-	if (!m1 | !m2 || (m1->ncols != m2->ncols) | (m1->nrows != m2->nrows))
+	if (!m1 | !m2 || !m1->mtx | !m2->mtx |
+			 (m1->ncols != m2->ncols) | (m1->nrows != m2->nrows))
 		return(0);
 	if ((m2->ncomp > 1) & (m2->ncomp != m1->ncomp))
 		return(0);
@@ -739,7 +783,7 @@ rmx_elemult(RMATRIX *m1, const RMATRIX *m2, int divide
 		if (divide) {
 		    double	d;
 		    if (m2->ncomp == 1) {
-			d = rmx_lval(m2,i,j)[0];
+			d = rmx_val(m2,i,j)[0];
 			if (d == 0) {
 			    ++zeroDivides;
 			    for (k = m1->ncomp; k--; )
@@ -751,7 +795,7 @@ rmx_elemult(RMATRIX *m1, const RMATRIX *m2, int divide
 			}
 		    } else
 		        for (k = m1->ncomp; k--; ) {
-			    d = rmx_lval(m2,i,j)[k];
+			    d = rmx_val(m2,i,j)[k];
 			    if (d == 0) {
 			        ++zeroDivides;
 				rmx_lval(m1,i,j)[k] = 0;
@@ -760,12 +804,12 @@ rmx_elemult(RMATRIX *m1, const RMATRIX *m2, int divide
 			}
 		} else {
 		    if (m2->ncomp == 1) {
-			const double	d = rmx_lval(m2,i,j)[0];
+			const double	d = rmx_val(m2,i,j)[0];
 		        for (k = m1->ncomp; k--; )
 			    rmx_lval(m1,i,j)[k] *= d;
 		    } else
 		        for (k = m1->ncomp; k--; )
-			    rmx_lval(m1,i,j)[k] *= rmx_lval(m2,i,j)[k];
+			    rmx_lval(m1,i,j)[k] *= rmx_val(m2,i,j)[k];
 		}
 	if (zeroDivides) {
 		rmx_addinfo(m1, "WARNING: zero divide(s) corrupted results\n");
@@ -781,7 +825,7 @@ rmx_sum(RMATRIX *msum, const RMATRIX *madd, const doub
 	double	*mysf = NULL;
 	int	i, j, k;
 
-	if (!msum | !madd ||
+	if (!msum | !madd || !msum->mtx | !madd->mtx |
 			(msum->nrows != madd->nrows) |
 			(msum->ncols != madd->ncols) |
 			(msum->ncomp != madd->ncomp))
@@ -801,7 +845,7 @@ rmx_sum(RMATRIX *msum, const RMATRIX *madd, const doub
 		rmx_addinfo(msum, rmx_mismatch_warn);
 	for (i = msum->nrows; i--; )
 	    for (j = msum->ncols; j--; ) {
-	    	const double	*da = rmx_lval(madd,i,j);
+	    	const double	*da = rmx_val(madd,i,j);
 	    	double		*ds = rmx_lval(msum,i,j);
 		for (k = msum->ncomp; k--; )
 		     ds[k] += sf[k] * da[k];
@@ -817,7 +861,7 @@ rmx_scale(RMATRIX *rm, const double sf[])
 {
 	int	i, j, k;
 
-	if (!rm | !sf)
+	if (!rm | !sf || !rm->mtx)
 		return(0);
 	for (i = rm->nrows; i--; )
 	    for (j = rm->ncols; j--; ) {
@@ -838,7 +882,7 @@ rmx_transform(const RMATRIX *msrc, int n, const double
 	int	i, j, ks, kd;
 	RMATRIX	*dnew;
 
-	if (!msrc | (n <= 0) | !cmat)
+	if (!msrc | (n <= 0) | !cmat || !msrc->mtx)
 		return(NULL);
 	dnew = rmx_alloc(msrc->nrows, msrc->ncols, n);
 	if (!dnew)
@@ -853,7 +897,7 @@ rmx_transform(const RMATRIX *msrc, int n, const double
 	dnew->dtype = msrc->dtype;
 	for (i = dnew->nrows; i--; )
 	    for (j = dnew->ncols; j--; ) {
-		const double	*ds = rmx_lval(msrc,i,j);
+		const double	*ds = rmx_val(msrc,i,j);
 	        for (kd = dnew->ncomp; kd--; ) {
 		    double	d = 0;
 		    for (ks = msrc->ncomp; ks--; )
@@ -902,7 +946,7 @@ cm_from_rmatrix(const RMATRIX *rm)
 		return(NULL);
 	for (i = cnew->nrows; i--; )
 	    for (j = cnew->ncols; j--; ) {
-		const double	*dp = rmx_lval(rm,i,j);
+		const double	*dp = rmx_val(rm,i,j);
 		COLORV		*cv = cm_lval(cnew,i,j);
 		switch (rm->ncomp) {
 		case 3: