--- ray/src/util/rcomb.c	2024/06/03 18:55:51	2.19
+++ ray/src/util/rcomb.c	2024/08/21 01:09:25	2.25
@@ -1,12 +1,12 @@
 #ifndef lint
-static const char RCSid[] = "$Id: rcomb.c,v 2.19 2024/06/03 18:55:51 greg Exp $";
+static const char RCSid[] = "$Id: rcomb.c,v 2.25 2024/08/21 01:09:25 greg Exp $";
 #endif
 /*
  * General component matrix combiner, operating on a row at a time.
  *
  * Multi-processing mode under Unix creates children that each work
  * on one input row at a time, fed by the original process.  Final conversion
- * and output to stdout is sorted by first child while its siblings send it
+ * and output to stdout is sorted by last child while its siblings send it
  * their record calculations.
  */
 
@@ -62,9 +62,9 @@ SUBPROC		*cproc = NULL;			/* child process array */
 int		nchildren = 0;			/* # of child processes */
 int		inchild = -1;			/* our child ID (-1: parent) */
 
-static int	checksymbolic(ROPMAT *rop);
+extern int	checksymbolic(ROPMAT *rop);
 
-static int
+int
 split_input(ROPMAT *rop)
 {
 	if (rop->rmp == &rop->imx && !(rop->rmp = rmx_copy(&rop->imx))) {
@@ -76,7 +76,7 @@ split_input(ROPMAT *rop)
 }
 
 /* Check/set transform based on a reference input file */
-static int
+int
 checkreffile(ROPMAT *rop)
 {
 	static const char	*curRF = NULL;
@@ -139,7 +139,7 @@ checkreffile(ROPMAT *rop)
 }
 
 /* Compute conversion row from spectrum to one channel of RGB */
-static void
+void
 rgbrow(ROPMAT *rop, int r, int p)
 {
 	const int	nc = rop->imx.ncomp;
@@ -156,7 +156,7 @@ rgbrow(ROPMAT *rop, int r, int p)
 }
 
 /* Compute conversion row from spectrum to one channel of XYZ */
-static void
+void
 xyzrow(ROPMAT *rop, int r, int p)
 {
 	const int	nc = rop->imx.ncomp;
@@ -173,8 +173,8 @@ xyzrow(ROPMAT *rop, int r, int p)
 }
 
 /* Use the spectral sensitivity function to compute matrix coefficients */
-static void
-sensrow(ROPMAT *rop, int r, double (*sf)(SCOLOR sc, int ncs, const float wlpt[4]))
+void
+sensrow(ROPMAT *rop, int r, double (*sf)(const SCOLOR sc, int ncs, const float wlpt[4]))
 {
 	const int	nc = rop->imx.ncomp;
 	int		i;
@@ -188,7 +188,7 @@ sensrow(ROPMAT *rop, int r, double (*sf)(SCOLOR sc, in
 }
 
 /* Check/set symbolic transform */
-static int
+int
 checksymbolic(ROPMAT *rop)
 {
 	const int	nc = rop->imx.ncomp;
@@ -302,7 +302,7 @@ checksymbolic(ROPMAT *rop)
 	return(1);
 }
 
-static int
+int
 get_component_xfm(ROPMAT *rop)
 {
 	int	i, j;
@@ -385,7 +385,7 @@ get_component_xfm(ROPMAT *rop)
 	return(1);
 }
 
-static int
+int
 apply_op(RMATRIX *dst, const RMATRIX *src, const RUNARYOP *ro)
 {
 	if (ro->clen > 0) {
@@ -404,7 +404,7 @@ apply_op(RMATRIX *dst, const RMATRIX *src, const RUNAR
 	return(1);
 }
 
-static int
+int
 open_input(ROPMAT *rop)
 {
 	int	outtype;
@@ -426,7 +426,7 @@ open_input(ROPMAT *rop)
 }
 
 /* Return nominal wavelength associated with input component (return nm) */
-static double
+double
 l_wavelength(char *nam)
 {
 	double	comp = argument(1);
@@ -450,7 +450,7 @@ l_wavelength(char *nam)
 }
 
 /* Return ith input with optional channel selector */
-static double
+double
 l_chanin(char *nam)
 {
 	double	inp = argument(1);
@@ -475,7 +475,7 @@ l_chanin(char *nam)
 	return(mop[mi].rmp->mtx[cur_col*in_ncomp + chan]);
 }
 
-static int
+int
 initialize(RMATRIX *imp)
 {
 	int	i;
@@ -488,7 +488,7 @@ initialize(RMATRIX *imp)
 		restype = mop[i].rmp->dtype;
 		if (!imp->dtype || (restype = rmx_newtype(restype, imp->dtype)) > 0)
 			imp->dtype = restype;
-		else
+		else if (!nowarn)
 			fprintf(stderr, "%s: warning - data type mismatch\n",
 					mop[i].inspec);
 		if (!i) {
@@ -522,7 +522,7 @@ initialize(RMATRIX *imp)
 	return(1);
 }
 
-static void
+void
 output_headinfo(FILE *fp)
 {
 	int	i;
@@ -545,30 +545,15 @@ output_headinfo(FILE *fp)
 	}
 }
 
-static void
-reset_inputs(void)
-{
-	int	i = nmats;
-
-	fpurge(stdin);		/* discard previous matrix input */
-	while (i--) {
-		if (mop[i].infp != stdin)
-			fclose(mop[i].infp);	/* ! pclose() */
-		mop[i].infp = stdin;
-		mop[i].imx.dtype = DTdouble;
-		mop[i].imx.pflags &= ~RMF_SWAPIN;
-	}
-}
-
-static int
+int
 spawned_children(int np)
 {
-	size_t	recsize = 0;
 	int	i, rv;
 
 #if defined(_WIN32) || defined(_WIN64)
 	if (np > 1) {
-		fputs("Warning: only one process under Windows\n", stderr);
+		if (!nowarn)
+			fputs("Warning: only one process under Windows\n", stderr);
 		np = 1;
 	} else
 #endif
@@ -579,27 +564,27 @@ spawned_children(int np)
 		mop[i].imx.nrows = 1;
 		if (!rmx_prepare(&mop[i].imx))
 			goto memerror;
-		recsize += rmx_array_size(&mop[i].imx);
 		if (mop[i].rmp != &mop[i].imx) {
 			mop[i].rmp->nrows = 1;
 			if (!rmx_prepare(mop[i].rmp))
 				goto memerror;
 		}
 	}
-				/* prep output row buffer */
-	if (mcat || mop[nmats].preop.clen > 0) {
+				/* prep output row buffer(s) */
+	if (mop[nmats].preop.clen > 0) {
 		if (!split_input(&mop[nmats]))	/* need separate buffer */
-			return(0);
-		if (mop[nmats].preop.clen > 0)
-			mop[nmats].rmp->ncomp = mop[nmats].preop.clen /
-						mop[nmats].imx.ncomp;
-		mop[nmats].rmp->nrows = 1;
-		if (!mcat | !mcat_last && !rmx_prepare(mop[nmats].rmp))
 			goto memerror;
+		mop[nmats].rmp->ncomp = mop[nmats].preop.clen /
+					mop[nmats].imx.ncomp;
 	}
 	mop[nmats].imx.nrows = 1;
 	if (!rmx_prepare(&mop[nmats].imx))
 		goto memerror;
+	if (mop[nmats].rmp != &mop[nmats].imx) {
+		mop[nmats].rmp->nrows = 1;
+		if (!rmx_prepare(mop[nmats].rmp))
+			goto memerror;
+	}
 	if (np <= 1) {		/* single process return */
 #ifdef getc_unlocked
 		for (i = 0; i < nmats; i++)
@@ -609,74 +594,86 @@ spawned_children(int np)
 		return(0);
 	}
 	fflush(stdout);		/* flush header & spawn children */
-	nchildren = np+1;
+	nchildren = np + 1;	/* extra child to sequence output */
 	cproc = (SUBPROC *)malloc(sizeof(SUBPROC)*nchildren);
 	if (!cproc)
 		goto memerror;
-				/* first child becomes stdout filter */
-	cproc[0].flags = PF_FILT_OUT;
-	cproc[0].w = fileno(stdout);
-	cproc[0].r = 0;
-	cproc[0].pid = -1;
-	rv = open_process(&cproc[i], NULL);
+	for (i = nchildren; i--; ) cproc[i] = sp_inactive;
+	cproc[nchildren-1].flags |= PF_FILT_OUT;
+				/* start each child from parent */
+	for (i = 0; i < nchildren; i++)
+		if ((rv = open_process(&cproc[i], NULL)) <= 0)
+			break;	/* child breaks here */
 	if (rv < 0) {
-		perror("fork");
+		perror("fork");	/* WTH? */
+		close_processes(cproc, i);
 		exit(1);
 	}
-	if (rv == 0) {		/* loop if first child */
-		inchild = 0;
-		reset_inputs();
-		output_loop();
-		_exit(0);
+	if (i != nchildren-1) {	/* last child is sole reader */
+		int	j = i;
+		while (j-- > 0) {
+			close(cproc[j].r);
+			cproc[j].r = -1;
+		}
 	}
-				/* start other children */
-	for (i = 1; (i < nchildren) & (rv > 0); i++) {
-		cproc[i].flags = 0;
-		cproc[i].w = 1;
-		cproc[i].r = 0;
-		cproc[i].pid = -1;
-		rv = open_process(&cproc[i], NULL);
+	if (rv > 0)
+		return(1);	/* parent return value */
+
+	inchild = i;		/* else set our child index */
+	while (i-- > 0)		/* only parent writes siblings */
+		close(cproc[i].w);
+
+	i = nmats;		/* close matrix streams (carefully) */
+	while (i-- > 0) {
+		if (mop[i].infp != stdin) {
+			close(fileno(mop[i].infp));	/* avoid lseek() */
+			fclose(mop[i].infp);		/* ! pclose() */
+		}
+		mop[i].infp = NULL;
 	}
-	if (rv < 0) {
-		perror("fork");
-		close_processes(cproc, i);
-		exit(1);
+	fpurge(stdin);		/* discard previously buffered input */
+
+	if (inchild == nchildren-1)
+		return(-1);	/* output process return value */
+
+	i = nmats;		/* get matrix rows from parent */
+	while (i-- > 0) {
+		mop[i].infp = stdin;
+		mop[i].imx.dtype = DTrmx_native;
+		mop[i].imx.pflags &= ~RMF_SWAPIN;
 	}
-	if (rv > 0)		/* parent return? */
-		return(1);
-	inchild = i;		/* our child index */
-	while (--i > 0)		/* don't share siblings' pipes */
-		close(cproc[i].w);
-	reset_inputs();
 #ifdef getc_unlocked
 	flockfile(stdin);
 #endif
-	return(0);		/* child return */
+	mop[nmats].rmp->dtype = DTrmx_native;
+	return(0);		/* worker child return value */
 memerror:
 	fputs("Out of memory in spawned_children()\n", stderr);
 	exit(1);
 }
 
-static int
+int
 parent_loop(void)
 {
 	int	i;
 
-	rmx_reset(&mop[nmats].imx);		/* not doing output side */
-	if (mop[nmats].rmp != &mop[nmats].imx)
-		rmx_reset(mop[nmats].rmp);
+	rmx_reset(&mop[nmats].imx);		/* not touching output side */
+	if (mop[nmats].rmp != &mop[nmats].imx) {
+		rmx_free(mop[nmats].rmp);
+		mop[nmats].rmp = &mop[nmats].imx;
+	}
 #ifdef getc_unlocked
-	for (i = 0; i < nmats; i++)
+	for (i = 0; i < nmats; i++)		/* we handle matrix inputs */
 		flockfile(mop[i].infp);
 #endif
 						/* load & send rows to kids */
 	for (cur_row = 0; (in_nrows <= 0) | (cur_row < in_nrows); cur_row++) {
-	    int	wfd = cproc[1 + cur_row%(nchildren-1)].w;
+	    int		wfd = cproc[cur_row % (nchildren-1)].w;
 	    for (i = 0; i < nmats; i++)
 		if (!rmx_load_row(mop[i].imx.mtx, &mop[i].imx, mop[i].infp)) {
 			if (cur_row > in_nrows)	/* unknown #input rows? */
 				break;
-			fprintf(stderr, "%s: read error at row %d\n",
+			fprintf(stderr, "%s: parent_loop() load error at row %d\n",
 					mop[i].inspec, cur_row);
 			return(0);
 		}
@@ -684,39 +681,41 @@ parent_loop(void)
 		break;
 	    for (i = 0; i < nmats; i++)
 	    	if (writebuf(wfd, mop[i].imx.mtx, rmx_array_size(&mop[i].imx))
-	    				!= rmx_array_size(&mop[i].imx))
+	    				!= rmx_array_size(&mop[i].imx)) {
+			fprintf(stderr, "%s: parent_loop() write error at row %d\n",
+					mop[i].inspec, cur_row);
 			return(0);
+		}
 	}
-	i = close_processes(cproc, nchildren);
-	free(cproc); cproc = NULL;
+	i = close_processes(cproc, nchildren);	/* collect family */
+	free(cproc); cproc = NULL; nchildren = 0;
 	if (i < 0) {
-		fputs("Warning: missing child in parent_loop()\n", stderr);
+		if (!nowarn)
+			fputs("Warning: lost child in parent_loop()\n", stderr);
 		return(1);
 	}
 	if (i > 0) {
 		fprintf(stderr, "Child exited with status %d\n", i);
 		return(0);
 	}
-	return(1);
-memerror:
-	fputs("Out of memory in parent_loop()\n", stderr);
-	exit(1);
+	return(1);				/* return success! */
 }
 
-static int
+int
 combine_input(void)
 {
-	const int	row0 = (inchild > 0)*(inchild-1);
-	const int	rstep = nchildren + !nchildren;
+	const int	row0 = (inchild >= 0)*inchild;
+	const int	rstep = nchildren ? nchildren-1 : 1;
 	ROPMAT		*res = &mop[nmats];
 	int		set_r, set_c;
 	RMATRIX		*tmp = NULL;
 	int		co_set;
 	int		i;
 
-	if (mcat && mcat_last &&
-			!(tmp = rmx_alloc(1, res->imx.ncols, res->rmp->ncomp)))
-		goto memerror;
+	if (mcat_last && !(tmp = rmx_alloc(1, res->imx.ncols, res->rmp->ncomp))) {
+		fputs("Out of buffer space in combine_input()\n", stderr);
+		return(0);
+	}
 					/* figure out what the user set */
 	co_set = fundefined("co");
 	if (!co_set)
@@ -738,7 +737,7 @@ combine_input(void)
 		if (!rmx_load_row(mop[i].imx.mtx, &mop[i].imx, mop[i].infp)) {
 			if (cur_row > in_nrows)	/* unknown #input rows? */
 				break;
-			fprintf(stderr, "%s: read error at row %d\n",
+			fprintf(stderr, "%s: combine_input() load error at row %d\n",
 					mop[i].inspec, cur_row);
 			return(0);
 		}
@@ -783,26 +782,60 @@ combine_input(void)
 			return(0);
 	    }
 	    rmx_free(mres); mres = NULL;
-	    if (inchild > 0) {		/* children share stdout */
-	    	i = getc(stdin);	/* signals it's our turn */
-	    	if (i != EOF) ungetc(i, stdin);
-	    }
 	    if (!rmx_write_data(res->rmp->mtx, res->rmp->ncomp,
-	    			res->rmp->ncols, res->rmp->dtype, stdout))
+	    			res->rmp->ncols, res->rmp->dtype, stdout) ||
+	    			 (inchild >= 0 && fflush(stdout) == EOF)) {
+		fprintf(stderr, "Conversion/write error at row %d in combine_input()\n",
+				cur_row);
 	    	return(0);
-	    if (inchild > 0 && fflush(stdout) == EOF)
-	    	return(0);
+	    }
 	}
-	return(inchild > 0 || fflush(stdout) != EOF);
-memerror:
-	fputs("Out of buffer space in combine_input()\n", stderr);
-	return(0);
+	return(inchild >= 0 || fflush(stdout) != EOF);
 multerror:
 	fputs("Unexpected matrix multiply error in combine_input()\n", stderr);
 	return(0);
 }
 
-static int
+int
+output_loop(void)
+{
+	const size_t	row_size = rmx_array_size(mop[nmats].rmp);
+	int		cur_child = 0;
+	int		i = nmats;
+
+	while (i-- > 0) {				/* free input buffers */
+		rmx_reset(&mop[i].imx);
+		if (mop[i].rmp != &mop[i].imx) {
+			rmx_free(mop[i].rmp);
+			mop[i].rmp = &mop[i].imx;
+		}
+	}
+	if (mop[nmats].rmp != &mop[nmats].imx)		/* output is split? */
+		rmx_reset(&mop[nmats].imx);
+#ifdef getc_unlocked
+	flockfile(stdout);				/* we own this, now */
+#endif
+	for ( ; ; ) {					/* loop until no more */
+		ssize_t		rv;
+		rv = readbuf(cproc[cur_child].r, mop[nmats].rmp->mtx, row_size);
+		if (!rv)				/* out of rows? */
+			break;
+		if (rv != row_size) {
+			fputs("Read error in output_loop()\n", stderr);
+			return(0);
+		}					/* do final conversion */
+		if (!rmx_write_data(mop[nmats].rmp->mtx, mop[nmats].rmp->ncomp,
+	    			mop[nmats].rmp->ncols, mop[nmats].rmp->dtype, stdout)) {
+			fputs("Conversion/write error in output_loop()\n", stderr);
+			return(0);
+		}
+		cur_child++;
+		cur_child *= (cur_child < inchild);	/* loop over workers */
+	}
+	return(fflush(stdout) != EOF);
+}
+
+int
 get_factors(double da[], int n, char *av[])
 {
 	int	ac;
@@ -812,7 +845,7 @@ get_factors(double da[], int n, char *av[])
 	return(ac);
 }
 
-static void
+void
 resize_inparr(int n2alloc)
 {
 	int	i;
@@ -914,17 +947,17 @@ main(int argc, char *argv[])
 				}
 				break;
 			case 'C':
+				mcat_last = 0;
 				if (!n || isflt(argv[i+1]))
 					goto userr;
 				defCsym = mop[nmats].preop.csym = argv[++i];
 				mop[nmats].preop.clen = 0;
-				mcat_last = 0;
 				break;
 			case 'c':
+				mcat_last = 0;
 				if (n && !isflt(argv[i+1])) {
 					mop[nmats].preop.csym = argv[++i];
 					mop[nmats].preop.clen = 0;
-					mcat_last = 0;
 					break;
 				}
 				if (n > MAXCOMP*MAXCOMP) n = MAXCOMP*MAXCOMP;
@@ -937,16 +970,15 @@ main(int argc, char *argv[])
 					goto userr;
 				}
 				mop[nmats].preop.csym = NULL;
-				mcat_last = 0;
 				break;
 			case 'm':
+				mcat_last = 1;
 				if (!n) goto userr;
 				if (argv[++i][0] == '-' && !argv[i][1]) {
 					if (stdin_used++) goto stdin_error;
 					mcat_spec = stdin_name;
 				} else
 					mcat_spec = argv[i];
-				mcat_last = 1;
 				break;
 			default:
 				fprintf(stderr, "%s: unknown option '%s'\n",
@@ -1015,9 +1047,12 @@ main(int argc, char *argv[])
 		return(1);
 	}
 	doptimize(1);			/* optimize definitions */
-	if (spawned_children(nproc))	/* running in parent process? */
+	i = spawned_children(nproc);	/* create multiple processes if requested */
+	if (i > 0)			/* running in parent process? */
 		return(parent_loop() ? 0 : 1);
-					/* process & write rows */
+	if (i < 0)			/* running in output process? */
+		return(output_loop() ? 0 : 1);
+					/* else we are a worker process */
 	return(combine_input() ? 0 : 1);
 stdin_error:
 	fprintf(stderr, "%s: %s used for more than one input\n",