--- ray/src/util/rmatrix.c	2024/05/19 15:32:24	2.79
+++ ray/src/util/rmatrix.c	2025/04/16 22:47:16	2.93
@@ -1,5 +1,5 @@
 #ifndef lint
-static const char RCSid[] = "$Id: rmatrix.c,v 2.79 2024/05/19 15:32:24 greg Exp $";
+static const char RCSid[] = "$Id: rmatrix.c,v 2.93 2025/04/16 22:47:16 greg Exp $";
 #endif
 /*
  * General matrix operations.
@@ -31,7 +31,7 @@ rmx_new(int nr, int nc, int n)
 	if (!dnew)
 		return(NULL);
 
-	dnew->dtype = DTdouble;
+	dnew->dtype = DTrmx_native;
 	dnew->nrows = nr;
 	dnew->ncols = nc;
 	dnew->ncomp = n;
@@ -46,12 +46,12 @@ int
 rmx_prepare(RMATRIX *rm)
 {
 	if (!rm) return(0);
-	if (rm->mtx)
+	if (rm->mtx)			/* assume it's right size */
 		return(1);
 	if ((rm->nrows <= 0) | (rm->ncols <= 0) | (rm->ncomp <= 0))
 		return(0);
-	rm->mtx = (double *)malloc(rmx_array_size(rm));
-	rm->pflags |= RMF_OURMEM;
+	rm->mtx = (rmx_dtype *)malloc(rmx_array_size(rm));
+	rm->pflags |= RMF_FREEMEM;
 	return(rm->mtx != NULL);
 }
 
@@ -61,9 +61,9 @@ rmx_alloc(int nr, int nc, int n)
 {
 	RMATRIX	*dnew = rmx_new(nr, nc, n);
 
-	if (dnew && !rmx_prepare(dnew)) {
+	if (!rmx_prepare(dnew)) {
 		rmx_free(dnew);
-		dnew = NULL;
+		return(NULL);
 	}
 	return(dnew);
 }
@@ -77,15 +77,15 @@ rmx_reset(RMATRIX *rm)
 		free(rm->info);
 		rm->info = NULL;
 	}
-	if (rm->mtx && rm->pflags & RMF_OURMEM) {
 #ifdef MAP_FILE
-		if (rm->mapped) {
-			munmap(rm->mapped, rmx_mapped_size(rm));
-			rm->mapped = NULL;
-		} else
+	if (rm->mapped) {
+		munmap(rm->mapped, rmx_mapped_size(rm));
+		rm->mapped = NULL;
+	} else
 #endif
-			free(rm->mtx);
-		rm->pflags &= ~RMF_OURMEM;
+	if (rm->pflags & RMF_FREEMEM) {
+		free(rm->mtx);
+		rm->pflags &= ~RMF_FREEMEM;
 	}
 	rm->mtx = NULL;
 }
@@ -116,13 +116,12 @@ rmx_newtype(int dtyp1, int dtyp2)
 int
 rmx_addinfo(RMATRIX *rm, const char *info)
 {
-	int	oldlen = 0;
+	size_t	oldlen = 0;
 
 	if (!rm || !info || !*info)
 		return(0);
 	if (!rm->info) {
 		rm->info = (char *)malloc(strlen(info)+1);
-		if (rm->info) rm->info[0] = '\0';
 	} else {
 		oldlen = strlen(rm->info);
 		rm->info = (char *)realloc(rm->info,
@@ -141,19 +140,19 @@ get_dminfo(char *s, void *p)
 	char	fmt[MAXFMTLEN];
 	int	i;
 
-	if (headidval(NULL, s))
+	if (isheadid(s))
 		return(0);
 	if (isncomp(s)) {
 		ip->ncomp = ncompval(s);
-		return(0);
+		return(ip->ncomp - 1);
 	}
 	if (!strncmp(s, "NROWS=", 6)) {
 		ip->nrows = atoi(s+6);
-		return(0);
+		return(ip->nrows - 1);
 	}
 	if (!strncmp(s, "NCOLS=", 6)) {
 		ip->ncols = atoi(s+6);
-		return(0);
+		return(ip->ncols - 1);
 	}
 	if ((i = isbigendian(s)) >= 0) {
 		if (nativebigendian() != i)
@@ -165,18 +164,17 @@ get_dminfo(char *s, void *p)
 	if (isexpos(s)) {
 		float	f = exposval(s);
 		scalecolor(ip->cexp, f);
-		return(0);
+		return(f > .0 ? 0 : -1);
 	}
 	if (iscolcor(s)) {
 		COLOR	ctmp;
-		colcorval(ctmp, s);
+		if (!colcorval(ctmp, s)) return(-1);
 		multcolor(ip->cexp, ctmp);
 		return(0);
 	}
-	if (iswlsplit(s)) {
-		wlsplitval(ip->wlpart, s);
-		return(0);
-	}
+	if (iswlsplit(s))
+		return(wlsplitval(ip->wlpart, s) - 1);
+
 	if (!formatval(fmt, s)) {
 		rmx_addinfo(ip, s);
 		return(0);
@@ -186,28 +184,34 @@ get_dminfo(char *s, void *p)
 			ip->dtype = i;
 			return(0);
 		}
-	return(-1);
+	return(-1);		/* bad format */
 }
 
 static int
-rmx_load_ascii(double *drp, const RMATRIX *rm, FILE *fp)
+rmx_load_ascii(rmx_dtype *drp, const RMATRIX *rm, FILE *fp)
 {
 	int	j, k;
 
 	for (j = 0; j < rm->ncols; j++)
 	        for (k = rm->ncomp; k-- > 0; )
-			if (fscanf(fp, "%lf", drp++) != 1)
+			if (fscanf(fp, rmx_scanfmt, drp++) != 1)
 				return(0);
 	return(1);
 }
 
 static int
-rmx_load_float(double *drp, const RMATRIX *rm, FILE *fp)
+rmx_load_float(rmx_dtype *drp, const RMATRIX *rm, FILE *fp)
 {
+#if DTrmx_native==DTfloat
+	if (getbinary(drp, sizeof(*drp)*rm->ncomp, rm->ncols, fp) != rm->ncols)
+		return(0);
+	if (rm->pflags & RMF_SWAPIN)
+		swap32((char *)drp, rm->ncols*rm->ncomp);
+#else
 	int	j, k;
-	float	val[100];
+	float	val[MAXCOMP];
 
-	if (rm->ncomp > 100) {
+	if (rm->ncomp > MAXCOMP) {
 		fputs("Unsupported # components in rmx_load_float()\n", stderr);
 		exit(1);
 	}
@@ -219,21 +223,40 @@ rmx_load_float(double *drp, const RMATRIX *rm, FILE *f
 	        for (k = 0; k < rm->ncomp; k++)
 			*drp++ = val[k];
 	}
+#endif
 	return(1);
 }
 
 static int
-rmx_load_double(double *drp, const RMATRIX *rm, FILE *fp)
+rmx_load_double(rmx_dtype *drp, const RMATRIX *rm, FILE *fp)
 {
+#if DTrmx_native==DTdouble
 	if (getbinary(drp, sizeof(*drp)*rm->ncomp, rm->ncols, fp) != rm->ncols)
 		return(0);
 	if (rm->pflags & RMF_SWAPIN)
 		swap64((char *)drp, rm->ncols*rm->ncomp);
+#else
+	int	j, k;
+	double	val[MAXCOMP];
+
+	if (rm->ncomp > MAXCOMP) {
+		fputs("Unsupported # components in rmx_load_double()\n", stderr);
+		exit(1);
+	}
+	for (j = 0; j < rm->ncols; j++) {
+		if (getbinary(val, sizeof(val[0]), rm->ncomp, fp) != rm->ncomp)
+			return(0);
+		if (rm->pflags & RMF_SWAPIN)
+			swap64((char *)val, rm->ncomp);
+	        for (k = 0; k < rm->ncomp; k++)
+			*drp++ = (float)val[k];
+	}
+#endif
 	return(1);
 }
 
 static int
-rmx_load_rgbe(double *drp, const RMATRIX *rm, FILE *fp)
+rmx_load_rgbe(rmx_dtype *drp, const RMATRIX *rm, FILE *fp)
 {
 	COLR	*scan;
 	COLOR	col;
@@ -256,15 +279,15 @@ rmx_load_rgbe(double *drp, const RMATRIX *rm, FILE *fp
 }
 
 static int
-rmx_load_spec(double *drp, const RMATRIX *rm, FILE *fp)
+rmx_load_spec(rmx_dtype *drp, const RMATRIX *rm, FILE *fp)
 {
-	uby8	*scan;
-	SCOLOR	scol;
+	COLRV	*scan;
+	COLORV	scol[MAXCOMP];
 	int	j, k;
 
-	if ((rm->ncomp < 3) | (rm->ncomp > MAXCSAMP))
+	if ((rm->ncomp < 3) | (rm->ncomp > MAXCOMP))
 		return(0);
-	scan = (uby8 *)tempbuffer((rm->ncomp+1)*rm->ncols);
+	scan = (COLRV *)tempbuffer((rm->ncomp+1)*rm->ncols);
 	if (!scan)
 		return(0);
 	if (freadscolrs(scan, rm->ncomp, rm->ncols, fp) < 0)
@@ -293,7 +316,7 @@ rmx_load_header(RMATRIX *rm, FILE *fp)
 	}
 	rm->dtype = DTascii;			/* assumed w/o FORMAT */
 	if (getheader(fp, get_dminfo, rm) < 0) {
-		fputs("Unrecognized matrix format\n", stderr);
+		fputs("Bad matrix header\n", stderr);
 		return(0);
 	}
 	if ((rm->dtype == DTrgbe) | (rm->dtype == DTxyze) &&
@@ -309,9 +332,9 @@ rmx_load_header(RMATRIX *rm, FILE *fp)
 	return(1);
 }
 
-/* Load next row as double (cannot be XML) */
+/* Load next row as rmx_dtype (cannot be XML) */
 int
-rmx_load_row(double *drp, const RMATRIX *rm, FILE *fp)
+rmx_load_row(rmx_dtype *drp, const RMATRIX *rm, FILE *fp)
 {
 	switch (rm->dtype) {
 	case DTascii:
@@ -338,14 +361,16 @@ 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->pflags & RMF_SWAPIN) &&
-			rmx_array_size(rm) >= 1L<<20 &&
-			(pos = ftell(fp)) >= 0 && !(pos % sizeof(double))) {
+	if ((rm->dtype == DTrmx_native) & !(rm->pflags & RMF_SWAPIN) &
+			(rmx_array_size(rm) >= 1L<<20) &&
+			(pos = ftell(fp)) >= 0 && !(pos % sizeof(rmx_dtype))) {
 		rm->mapped = mmap(NULL, rmx_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);
-			rm->pflags |= RMF_OURMEM;
+			if (rm->pflags & RMF_FREEMEM)
+				free(rm->mtx);
+			rm->mtx = (rmx_dtype *)rm->mapped + pos/sizeof(rmx_dtype);
+			rm->pflags &= ~RMF_FREEMEM;
 			return(1);
 		}		/* else fall back on reading into memory */
 		rm->mapped = NULL;
@@ -378,13 +403,9 @@ rmx_load(const char *inspec, RMPref rmp)
 		fp = stdin;
 	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 */
+	else {					/* check suffix */
+		const char	*sp = strrchr(inspec, '.');
+		if (sp > inspec && !strcasecmp(sp+1, "XML")) {	/* BSDF? */
 			CMATRIX	*cm = rmp==RMPnone ? (CMATRIX *)NULL :
 					rmp==RMPtrans ? cm_loadBTDF(inspec) :
 					cm_loadBRDF(inspec, rmp==RMPreflB) ;
@@ -432,53 +453,68 @@ rmx_load(const char *inspec, RMPref rmp)
 						/* undo exposure? */
 	if ((dnew->cexp[0] != 1.f) |
 			(dnew->cexp[1] != 1.f) | (dnew->cexp[2] != 1.f)) {
-		double	cmlt[MAXCSAMP];
+		double	cmlt[MAXCOMP];
 		int	i;
-		cmlt[0] = 1./dnew->cexp[0];
-		cmlt[1] = 1./dnew->cexp[1];
-		cmlt[2] = 1./dnew->cexp[2];
-		if (dnew->ncomp > MAXCSAMP) {
+		if (dnew->ncomp > MAXCOMP) {
 			fprintf(stderr, "Excess spectral components in: %s\n",
 					inspec);
 			rmx_free(dnew);
 			return(NULL);
 		}
+		cmlt[0] = 1./dnew->cexp[0];
+		cmlt[1] = 1./dnew->cexp[1];
+		cmlt[2] = 1./dnew->cexp[2];
 		for (i = dnew->ncomp; i-- > 3; )
-			cmlt[i] = cmlt[1];
+			cmlt[i] = cmlt[1];	/* XXX hack! */
 		rmx_scale(dnew, cmlt);
 		setcolor(dnew->cexp, 1.f, 1.f, 1.f);
 	}
 	return(dnew);
 }
 
+#if DTrmx_native==DTdouble
 static int
-rmx_write_ascii(const double *dp, int nc, int len, FILE *fp)
+rmx_write_float(const rmx_dtype *dp, int len, FILE *fp)
 {
-	while (len-- > 0) {
-		int	k = nc;
-		while (k-- > 0)
-			fprintf(fp, " %.7e", *dp++);
-		fputc('\t', fp);
+	float	val;
+
+	while (len--) {
+		val = (float)*dp++;
+		if (putbinary(&val, sizeof(val), 1, fp) != 1)
+			return(0);
 	}
-	return(fputc('\n', fp) != EOF);
+	return(1);
 }
-
+#else
 static int
-rmx_write_float(const double *dp, int len, FILE *fp)
+rmx_write_double(const rmx_dtype *dp, int len, FILE *fp)
 {
-	float	val;
+	double	val;
 
 	while (len--) {
 		val = *dp++;
-		if (putbinary(&val, sizeof(float), 1, fp) != 1)
+		if (putbinary(&val, sizeof(val), 1, fp) != 1)
 			return(0);
 	}
 	return(1);
 }
+#endif
 
 static int
-rmx_write_rgbe(const double *dp, int nc, int len, FILE *fp)
+rmx_write_ascii(const rmx_dtype *dp, int nc, int len, FILE *fp)
 {
+	while (len-- > 0) {
+		int	k = nc;
+		while (k-- > 0)
+			fprintf(fp, " %.7e", *dp++);
+		fputc('\t', fp);
+	}
+	return(fputc('\n', fp) != EOF);
+}
+
+static int
+rmx_write_rgbe(const rmx_dtype *dp, int nc, int len, FILE *fp)
+{
 	COLR	*scan;
 	int	j;
 
@@ -496,14 +532,14 @@ rmx_write_rgbe(const double *dp, int nc, int len, FILE
 }
 
 static int
-rmx_write_spec(const double *dp, int nc, int len, FILE *fp)
+rmx_write_spec(const rmx_dtype *dp, int nc, int len, FILE *fp)
 {
-	uby8	*scan;
-	SCOLOR	scol;
+	COLRV	*scan;
+	COLORV	scol[MAXCOMP];
 	int	j, k;
 
-	if (nc < 3) return(0);
-	scan = (uby8 *)tempbuffer((nc+1)*len);
+	if ((nc < 3) | (nc > MAXCOMP)) return(0);
+	scan = (COLRV *)tempbuffer((nc+1)*len);
 	if (!scan) return(0);
 	for (j = 0; j < len; j++, dp += nc) {
 	    	for (k = nc; k--; )
@@ -538,9 +574,13 @@ rmx_write_header(const RMATRIX *rm, int dtype, FILE *f
 		return(0);
 	if (rm->info)
 		fputs(rm->info, fp);
-	if (dtype == DTfromHeader)
+	if (dtype == DTfromHeader) {
 		dtype = rm->dtype;
-	else if (dtype == DTrgbe && (rm->dtype == DTxyze ||
+#if DTrmx_native==DTfloat
+		if (dtype == DTdouble)		/* but stored as float? */
+			dtype = DTfloat;
+#endif
+	} else if (dtype == DTrgbe && (rm->dtype == DTxyze ||
 					findCIEprims(rm->info)))
 		dtype = DTxyze;
 	else if ((dtype == DTxyze) & (rm->dtype == DTrgbe))
@@ -584,15 +624,20 @@ rmx_write_header(const RMATRIX *rm, int dtype, FILE *f
 
 /* Write out matrix data (usually by row) */
 int
-rmx_write_data(const double *dp, int nc, int len, int dtype, FILE *fp)
+rmx_write_data(const rmx_dtype *dp, int nc, int len, int dtype, FILE *fp)
 {
 	switch (dtype) {
-	case DTascii:
-		return(rmx_write_ascii(dp, nc, len, fp));
+#if DTrmx_native==DTdouble
 	case DTfloat:
 		return(rmx_write_float(dp, nc*len, fp));
+#else
 	case DTdouble:
+		return(rmx_write_double(dp, nc*len, fp));
+#endif
+	case DTrmx_native:
 		return(putbinary(dp, sizeof(*dp)*nc, len, fp) == len);
+	case DTascii:
+		return(rmx_write_ascii(dp, nc, len, fp));
 	case DTrgbe:
 	case DTxyze:
 		return(rmx_write_rgbe(dp, nc, len, fp));
@@ -615,7 +660,7 @@ rmx_write(const RMATRIX *rm, int dtype, FILE *fp)
 #ifdef getc_unlocked
 	flockfile(fp);
 #endif
-	if (dtype == DTdouble)			/* write all at once? */
+	if (dtype == DTrmx_native)		/* write all at once? */
 		ok = rmx_write_data(rm->mtx, rm->ncomp,
 				rm->nrows*rm->ncols, dtype, fp);
 	else					/* else row by row */
@@ -644,7 +689,7 @@ rmx_identity(const int dim, const int n)
 		return(NULL);
 	memset(rid->mtx, 0, rmx_array_size(rid));
 	for (i = dim; i--; ) {
-	    double	*dp = rmx_lval(rid,i,i);
+	    rmx_dtype	*dp = rmx_lval(rid,i,i);
 	    for (k = n; k--; )
 		dp[k] = 1.;
 	}
@@ -680,68 +725,110 @@ rmx_copy(const RMATRIX *rm)
 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))
+	if (!rdst | !rsrc)
 		return(0);
-
 	if (dometa) {		/* transfer everything? */
 		rmx_reset(rdst);
 		*rdst = *rsrc;
 		rsrc->info = NULL; rsrc->mapped = NULL; rsrc->mtx = NULL;
 		return(1);
 	}
-	if (rdst->pflags & RMF_OURMEM) {
-#ifdef MAP_FILE			/* just matrix data -- leave metadata */
-		if (rdst->mapped)
-			munmap(rdst->mapped, rmx_mapped_size(rdst));
-		else
+				/* just matrix data -- leave metadata */
+	if ((rdst->nrows != rsrc->nrows) |
+			(rdst->ncols != rsrc->ncols) |
+			(rdst->ncomp != rsrc->ncomp))
+		return(0);
+#ifdef MAP_FILE
+	if (rdst->mapped)
+		munmap(rdst->mapped, rmx_mapped_size(rdst));
+	else
 #endif
-		if (rdst->mtx)
-			free(rdst->mtx);
+	if (rdst->pflags & RMF_FREEMEM) {
+		free(rdst->mtx);
+		rdst->pflags &= ~RMF_FREEMEM;
 	}
 	rdst->mapped = rsrc->mapped;
 	rdst->mtx = rsrc->mtx;
-	if (rsrc->pflags & RMF_OURMEM)
-		rdst->pflags |= RMF_OURMEM;
-	else
-		rdst->pflags &= ~RMF_OURMEM;
+	rdst->pflags |= rsrc->pflags & RMF_FREEMEM;
 	rsrc->mapped = NULL; rsrc->mtx = NULL;
 	return(1);
 }
 
-/* Allocate and assign transposed matrix */
-RMATRIX *
-rmx_transpose(const RMATRIX *rm)
+/* Transpose the given matrix */
+int
+rmx_transpose(RMATRIX *rm)
 {
-	RMATRIX	*dnew;
-	int	i, j;
+#define	bmbyte(r,c)	bmap[((r)*rm->ncols+(c))>>3]
+#define	bmbit(r,c)	(1 << ((r)*rm->ncols+(c) & 7))
+#define	bmop(r,c, op)	(bmbyte(r,c) op bmbit(r,c))
+#define	bmtest(r,c)	bmop(r,c,&)
+#define	bmset(r,c)	bmop(r,c,|=)
+#define bmtestandset(r,c)	(!bmtest(r,c) && bmset(r,c))
+	uby8		*bmap;
+	rmx_dtype	val[MAXCOMP];
+	RMATRIX		dold;
+	int		i, j;
 
-	if (!rm || !rm->mtx)
+	if (!rm || !rm->mtx | (rm->ncomp > MAXCOMP))
 		return(0);
-	if ((rm->nrows == 1) | (rm->ncols == 1)) {
-		dnew = rmx_copy(rm);
-		if (!dnew)
-			return(NULL);
-		dnew->nrows = rm->ncols;
-		dnew->ncols = rm->nrows;
-		return(dnew);
+	if (rm->info)
+		rmx_addinfo(rm, "Transposed rows and columns\n");
+	if ((rm->nrows == 1) | (rm->ncols == 1)) { /* vector? */
+		j = rm->ncols;
+		rm->ncols = rm->nrows;
+		rm->nrows = j;
+		return(1);
 	}
-	dnew = rmx_alloc(rm->ncols, rm->nrows, rm->ncomp);
-	if (!dnew)
-		return(NULL);
-	if (rm->info) {
-		rmx_addinfo(dnew, rm->info);
-		rmx_addinfo(dnew, "Transposed rows and columns\n");
+	if (rm->nrows == rm->ncols) {	/* square matrix case */
+	     for (i = rm->nrows; i--; )
+	     	for (j = rm->ncols; j--; ) {
+		    if (i == j) continue;
+		    memcpy(val, rmx_val(rm,i,j),
+				sizeof(rmx_dtype)*rm->ncomp);
+		    memcpy(rmx_lval(rm,i,j), rmx_val(rm,j,i),
+				sizeof(rmx_dtype)*rm->ncomp);
+		    memcpy(rmx_val(rm,j,i), val,
+				sizeof(rmx_dtype)*rm->ncomp);
+		}
+		return(1);
 	}
-	dnew->dtype = rm->dtype;
-	copycolor(dnew->cexp, rm->cexp);
-	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_val(rm,j,i),
-	    			sizeof(double)*dnew->ncomp);
-	return(dnew);
+					/* clear completion bitmap */
+	bmap = (uby8 *)calloc(((size_t)rm->nrows*rm->ncols+7)>>3, 1);
+	if (!bmap)
+		return(0);
+	dold = *rm;
+	rm->ncols = dold.nrows; rm->nrows = dold.ncols;
+	for (i = rm->nrows; i--; )
+	    for (j = rm->ncols; j--; ) {
+	    	int	i0, j0;
+	    	int	i1 = i;
+	    	size_t	j1 = j;
+		if (!bmtestandset(i,j)) continue;
+		memcpy(val, rmx_val(rm,i,j),
+			sizeof(rmx_dtype)*rm->ncomp);
+		for ( ; ; ) {		/* value transpose loop */
+		    const rmx_dtype	*src;
+		    i0 = i1; j0 = (int)j1;
+		    src = rmx_val(&dold, j0, i0);
+		    j1 = (src - dold.mtx)/dold.ncomp;
+		    i1 = j1 / rm->ncols;
+		    j1 -= i1*rm->ncols;
+		    if (!bmtestandset(i1,j1)) break;
+		    memcpy(rmx_lval(rm,i0,j0), src,
+				sizeof(rmx_dtype)*rm->ncomp);
+		}
+					/* close the loop */
+		memcpy(rmx_lval(rm,i0,j0), val,
+			sizeof(rmx_dtype)*rm->ncomp);
+	    }
+	free(bmap);			/* all done! */
+	return(1);
+#undef	bmbyte
+#undef	bmbit
+#undef	bmop
+#undef	bmtest
+#undef	bmset
+#undef	bmtestandset
 }
 
 /* Multiply (concatenate) two matrices and allocate the result */
@@ -767,8 +854,9 @@ rmx_multiply(const RMATRIX *m1, const RMATRIX *m2)
 	        for (k = mres->ncomp; k--; ) {
 		    double	d = 0;
 		    for (h = m1->ncols; h--; )
-			d += rmx_val(m1,i,h)[k] * rmx_val(m2,h,j)[k];
-		    rmx_lval(mres,i,j)[k] = d;
+			d += (double)rmx_val(m1,i,h)[k] *
+					rmx_val(m2,h,j)[k];
+		    rmx_lval(mres,i,j)[k] = (rmx_dtype)d;
 		}
 	return(mres);
 }
@@ -793,7 +881,7 @@ rmx_elemult(RMATRIX *m1, const RMATRIX *m2, int divide
 	for (i = m1->nrows; i--; )
 	    for (j = m1->ncols; j--; )
 		if (divide) {
-		    double	d;
+		    rmx_dtype	d;
 		    if (m2->ncomp == 1) {
 			d = rmx_val(m2,i,j)[0];
 			if (d == 0) {
@@ -816,7 +904,7 @@ rmx_elemult(RMATRIX *m1, const RMATRIX *m2, int divide
 			}
 		} else {
 		    if (m2->ncomp == 1) {
-			const double	d = rmx_val(m2,i,j)[0];
+			const rmx_dtype	d = rmx_val(m2,i,j)[0];
 		        for (k = m1->ncomp; k--; )
 			    rmx_lval(m1,i,j)[k] *= d;
 		    } else
@@ -857,10 +945,10 @@ 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_val(madd,i,j);
-	    	double		*ds = rmx_lval(msum,i,j);
+	    	const rmx_dtype	*da = rmx_val(madd,i,j);
+	    	rmx_dtype	*ds = rmx_lval(msum,i,j);
 		for (k = msum->ncomp; k--; )
-		     ds[k] += sf[k] * da[k];
+		     ds[k] += (rmx_dtype)sf[k] * da[k];
 	    }
 	if (mysf)
 		free(mysf);
@@ -877,9 +965,9 @@ rmx_scale(RMATRIX *rm, const double sf[])
 		return(0);
 	for (i = rm->nrows; i--; )
 	    for (j = rm->ncols; j--; ) {
-	    	double	*dp = rmx_lval(rm,i,j);
+	    	rmx_dtype	*dp = rmx_lval(rm,i,j);
 		for (k = rm->ncomp; k--; )
-		    dp[k] *= sf[k];
+		    dp[k] *= (rmx_dtype)sf[k];
 	    }
 	if (rm->info)
 		rmx_addinfo(rm, "Applied scalar\n");
@@ -909,12 +997,12 @@ 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_val(msrc,i,j);
+		const rmx_dtype	*ds = rmx_val(msrc,i,j);
 	        for (kd = dnew->ncomp; kd--; ) {
 		    double	d = 0;
 		    for (ks = msrc->ncomp; ks--; )
 		        d += cmat[kd*msrc->ncomp + ks] * ds[ks];
-		    rmx_lval(dnew,i,j)[kd] = d;
+		    rmx_lval(dnew,i,j)[kd] = (rmx_dtype)d;
 		}
 	    }
 	return(dnew);
@@ -924,7 +1012,6 @@ rmx_transform(const RMATRIX *msrc, int n, const double
 RMATRIX *
 rmx_from_cmatrix(const CMATRIX *cm)
 {
-	int	i, j;
 	RMATRIX	*dnew;
 
 	if (!cm)
@@ -932,15 +1019,23 @@ rmx_from_cmatrix(const CMATRIX *cm)
 	dnew = rmx_alloc(cm->nrows, cm->ncols, 3);
 	if (!dnew)
 		return(NULL);
-	dnew->dtype = DTfloat;
-	for (i = dnew->nrows; i--; )
-	    for (j = dnew->ncols; j--; ) {
-		const COLORV	*cv = cm_lval(cm,i,j);
-		double		*dp = rmx_lval(dnew,i,j);
-		dp[0] = cv[0];
-		dp[1] = cv[1];
-		dp[2] = cv[2];
-	    }
+
+	dnew->dtype = sizeof(COLORV)==sizeof(float) ?
+			DTfloat : DTdouble;
+
+	if (sizeof(COLORV) == sizeof(rmx_dtype)) {
+		memcpy(dnew->mtx, cm->cmem, rmx_array_size(dnew));
+	} else {
+		int	i, j;
+		for (i = dnew->nrows; i--; )
+		    for (j = dnew->ncols; j--; ) {
+			const COLORV	*cv = cm_lval(cm,i,j);
+			rmx_dtype	*dp = rmx_lval(dnew,i,j);
+			dp[0] = cv[0];
+			dp[1] = cv[1];
+			dp[2] = cv[2];
+		    }
+	}
 	return(dnew);
 }
 
@@ -948,33 +1043,41 @@ rmx_from_cmatrix(const CMATRIX *cm)
 CMATRIX *
 cm_from_rmatrix(const RMATRIX *rm)
 {
-	int	i, j;
 	CMATRIX	*cnew;
 
-	if (!rm || !rm->mtx | (rm->ncomp == 2))
+	if (!rm || !rm->mtx | (rm->ncomp == 2) | (rm->ncomp > MAXCOMP))
 		return(NULL);
 	cnew = cm_alloc(rm->nrows, rm->ncols);
 	if (!cnew)
 		return(NULL);
-	for (i = cnew->nrows; i--; )
-	    for (j = cnew->ncols; j--; ) {
-		const double	*dp = rmx_val(rm,i,j);
-		COLORV		*cv = cm_lval(cnew,i,j);
-		switch (rm->ncomp) {
-		case 3:
-	    	    setcolor(cv, dp[0], dp[1], dp[2]);
-	    	    break;
-		case 1:
-		    setcolor(cv, dp[0], dp[0], dp[0]);
-		    break;
-		default: {
-			SCOLOR	scol;
-			int	k;
-			for (k = rm->ncomp; k--; )
-				scol[k] = dp[k];
-			scolor2color(cv, scol, rm->ncomp, rm->wlpart);
-		    } break;
-		}
-	    }
+	if ((sizeof(COLORV) == sizeof(rmx_dtype)) & (rm->ncomp == 3)) {
+		memcpy(cnew->cmem, rm->mtx, rmx_array_size(rm));
+	} else {
+		int	i, j;
+		for (i = cnew->nrows; i--; )
+		    for (j = cnew->ncols; j--; ) {
+			const rmx_dtype	*dp = rmx_val(rm,i,j);
+			COLORV		*cv = cm_lval(cnew,i,j);
+			switch (rm->ncomp) {
+			case 3:
+			    setcolor(cv, dp[0], dp[1], dp[2]);
+			    break;
+			case 1:
+			    setcolor(cv, dp[0], dp[0], dp[0]);
+			    break;
+			default:
+			    if (sizeof(COLORV) == sizeof(rmx_dtype)) {
+				scolor2color(cv, (const COLORV *)dp,
+						rm->ncomp, rm->wlpart);
+			    } else {
+				COLORV	scol[MAXCOMP];
+				int	k = rm->ncomp;
+				while (k--) scol[k] = dp[k];
+				scolor2color(cv, scol, rm->ncomp, rm->wlpart);
+			    }
+			    break;
+			}
+		    }
+	}
 	return(cnew);
 }