A small optimization: handle most N-channel <-> M-channel conversions

inside the mixer and the demultiplexer. This way, aucat will not
trigger the heavy conversion code when only channel conversions are
required. Cuts ~50% of the CPU usage on envy(4) devices, can improve
surround 4.0, 5.1 and 7.1 capable devices. No functionnal change.

8 files changed, 129 insertions, 53 deletions

diff --git a/usr.bin/aucat/abuf.c b/usr.bin/aucat/abuf.c
index 61a18a17293..c15b138ad31 100644
--- a/usr.bin/aucat/abuf.c
+++ b/usr.bin/aucat/abuf.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: abuf.c,v 1.7 2008/10/26 08:49:43 ratchov Exp $	*/
+/*	$OpenBSD: abuf.c,v 1.8 2008/11/03 22:25:13 ratchov Exp $	*/
 /*
  * Copyright (c) 2008 Alexandre Ratchov <alex@caoua.org>
  *
@@ -38,6 +38,7 @@
 #include <stdarg.h>
 
 #include "conf.h"
+#include "aparams.h"
 #include "aproc.h"
 #include "abuf.h"
 
@@ -63,11 +64,12 @@ abuf_dprn(int n, struct abuf *buf, char *fmt, ...)
 #endif
 
 struct abuf *
-abuf_new(unsigned nfr, unsigned bpf)
+abuf_new(unsigned nfr, struct aparams *par)
 {
 	struct abuf *buf;
-	unsigned len;
+	unsigned len, bpf;
 
+	bpf = aparams_bpf(par);
 	len = nfr * bpf;
 	buf = malloc(sizeof(struct abuf) + len);
 	if (buf == NULL) {
@@ -75,6 +77,8 @@ abuf_new(unsigned nfr, unsigned bpf)
 		abort();
 	}
 	buf->bpf = bpf;
+	buf->cmin = par->cmin;
+	buf->cmax = par->cmax;
 	buf->inuse = 0;
 
 	/*
diff --git a/usr.bin/aucat/abuf.h b/usr.bin/aucat/abuf.h
index aefe8698fb8..ae1ab690c68 100644
--- a/usr.bin/aucat/abuf.h
+++ b/usr.bin/aucat/abuf.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: abuf.h,v 1.9 2008/10/26 08:49:43 ratchov Exp $	*/
+/*	$OpenBSD: abuf.h,v 1.10 2008/11/03 22:25:13 ratchov Exp $	*/
 /*
  * Copyright (c) 2008 Alexandre Ratchov <alex@caoua.org>
  *
@@ -21,6 +21,7 @@
 
 struct abuf;
 struct aproc;
+struct aparam;
 
 struct abuf {
 	/*
@@ -32,9 +33,9 @@ struct abuf {
 	 * intput->output path.
 	 */
 	int mixvol;		/* input gain */
-	unsigned mixdone;	/* input of mixer */
-	unsigned mixtodo;	/* output of mixer */
-	unsigned subdone;	/* output if sub */
+	unsigned mixdone;	/* input already processed */
+	unsigned mixtodo;	/* output not yet complete */
+	unsigned subdone;	/* output already processed */
 #define XRUN_IGNORE	0	/* on xrun silently insert/discard samples */
 #define XRUN_SYNC	1	/* catchup to sync to the mix/sub */
 #define XRUN_ERROR	2	/* xruns are errors, eof/hup buffer */
@@ -46,6 +47,7 @@ struct abuf {
 	 * fifo parameters
 	 */
 	unsigned bpf;		/* bytes per frame */
+	unsigned cmin, cmax;	/* channel range of this buf */
 	unsigned start;		/* offset where data starts */
 	unsigned used;		/* valid data */
 	unsigned len;		/* size of the ring */
@@ -92,7 +94,7 @@ struct abuf {
  */
 #define ABUF_EMPTY(b) ((b)->used == 0)
 
-struct abuf *abuf_new(unsigned, unsigned);
+struct abuf *abuf_new(unsigned, struct aparams *);
 void abuf_del(struct abuf *);
 unsigned char *abuf_rgetblk(struct abuf *, unsigned *, unsigned);
 unsigned char *abuf_wgetblk(struct abuf *, unsigned *, unsigned);
diff --git a/usr.bin/aucat/aparams.c b/usr.bin/aucat/aparams.c
index 40cc9cd7fb3..2ee1b4d763e 100644
--- a/usr.bin/aucat/aparams.c
+++ b/usr.bin/aucat/aparams.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: aparams.c,v 1.2 2008/10/26 08:49:43 ratchov Exp $	*/
+/*	$OpenBSD: aparams.c,v 1.3 2008/11/03 22:25:13 ratchov Exp $	*/
 /*
  * Copyright (c) 2008 Alexandre Ratchov <alex@caoua.org>
  *
@@ -183,17 +183,14 @@ aparams_print2(struct aparams *par1, struct aparams *par2)
 }
 
 /*
- * Return true if both parameters are the same.
+ * Return true if both encodings are the same.
  */
 int
-aparams_eq(struct aparams *par1, struct aparams *par2)
+aparams_eqenc(struct aparams *par1, struct aparams *par2)
 {
 	if (par1->bps != par2->bps ||
 	    par1->bits != par2->bits ||
-	    par1->sig != par2->sig ||
-	    par1->cmin != par2->cmin ||
-	    par1->cmax != par2->cmax ||
-	    par1->rate != par2->rate)
+	    par1->sig != par2->sig)
 		return 0;
 	if ((par1->bits != 8 * par1->bps) && par1->msb != par2->msb)
 		return 0;
@@ -203,6 +200,40 @@ aparams_eq(struct aparams *par1, struct aparams *par2)
 }
 
 /*
+ * Return true if both parameters are the same.
+ */
+int
+aparams_eq(struct aparams *par1, struct aparams *par2)
+{
+	if (!aparams_eqenc(par1, par2) ||
+	    par1->cmin != par2->cmin ||
+	    par1->cmax != par2->cmax ||
+	    par1->rate != par2->rate)
+		return 0;
+	return 1;
+}
+
+/*
+ * Retrurn true if first channel range includes second range
+ */
+int
+aparams_subset(struct aparams *subset, struct aparams *set)
+{
+	return subset->cmin >= set->cmin && subset->cmax <= set->cmax;
+}
+
+/*
+ * Return true if rates are the same
+ */
+int
+aparams_eqrate(struct aparams *p1, struct aparams *p2)
+{
+	/* XXX: allow 1/9 halftone of difference */
+	return p1->rate == p2->rate;
+}
+
+
+/*
  * Return the number of bytes per frame with the given parameters.
  */
 unsigned
diff --git a/usr.bin/aucat/aparams.h b/usr.bin/aucat/aparams.h
index 8d394502229..1e226f44899 100644
--- a/usr.bin/aucat/aparams.h
+++ b/usr.bin/aucat/aparams.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: aparams.h,v 1.2 2008/10/26 08:49:43 ratchov Exp $	*/
+/*	$OpenBSD: aparams.h,v 1.3 2008/11/03 22:25:13 ratchov Exp $	*/
 /*
  * Copyright (c) 2008 Alexandre Ratchov <alex@caoua.org>
  *
@@ -72,7 +72,10 @@ struct aparams {
 void aparams_init(struct aparams *, unsigned, unsigned, unsigned);
 void aparams_print(struct aparams *);
 void aparams_print2(struct aparams *, struct aparams *);
+int aparams_eqrate(struct aparams *, struct aparams *);
+int aparams_eqenc(struct aparams *, struct aparams *);
 int aparams_eq(struct aparams *, struct aparams *);
+int aparams_subset(struct aparams *, struct aparams *);
 unsigned aparams_bpf(struct aparams *);
 int aparams_strtoenc(struct aparams *, char *);
 int aparams_enctostr(struct aparams *, char *);
diff --git a/usr.bin/aucat/aproc.c b/usr.bin/aucat/aproc.c
index 527a874042d..bf619514662 100644
--- a/usr.bin/aucat/aproc.c
+++ b/usr.bin/aucat/aproc.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: aproc.c,v 1.12 2008/10/26 08:49:43 ratchov Exp $	*/
+/*	$OpenBSD: aproc.c,v 1.13 2008/11/03 22:25:13 ratchov Exp $	*/
 /*
  * Copyright (c) 2008 Alexandre Ratchov <alex@caoua.org>
  *
@@ -348,30 +348,38 @@ void
 mix_badd(struct abuf *ibuf, struct abuf *obuf)
 {
 	short *idata, *odata;
-	unsigned i, scount, icount, ocount;
 	int vol = ibuf->mixvol;
+	unsigned i, j, icnt, onext, ostart;
+	unsigned scount, icount, ocount;
 
 	DPRINTFN(4, "mix_badd: todo = %u, done = %u\n",
 	    obuf->mixtodo, ibuf->mixdone);
 
 	idata = (short *)abuf_rgetblk(ibuf, &icount, 0);
-	icount -= icount % ibuf->bpf;
+	icount /= ibuf->bpf;
 	if (icount == 0)
 		return;
 
 	odata = (short *)abuf_wgetblk(obuf, &ocount, ibuf->mixdone);
-	ocount -= ocount % obuf->bpf;
+	ocount /= obuf->bpf;
 	if (ocount == 0)
 		return;
 
+	ostart = ibuf->cmin - obuf->cmin; 
+	onext = obuf->cmax - ibuf->cmax + ostart;
+	icnt = ibuf->cmax - ibuf->cmin + 1;
+	odata += ostart;
 	scount = (icount < ocount) ? icount : ocount;
-	for (i = scount / sizeof(short); i > 0; i--) {
-		*odata += (*idata * vol) >> ADATA_SHIFT;
-		idata++;
-		odata++;
-	}	
-	abuf_rdiscard(ibuf, scount);
-	ibuf->mixdone += scount;
+	for (i = scount; i > 0; i--) {
+		for (j = icnt; j > 0; j--) {
+			*odata += (*idata * vol) >> ADATA_SHIFT;
+			idata++;
+			odata++;
+		}
+		odata += onext;
+	}
+	abuf_rdiscard(ibuf, scount * ibuf->bpf);
+	ibuf->mixdone += scount * obuf->bpf;
 
 	DPRINTFN(4, "mix_badd: added %u, done = %u, todo = %u\n",
 	    scount, ibuf->mixdone, obuf->mixtodo);
@@ -509,6 +517,12 @@ mix_hup(struct aproc *p, struct abuf *obuf)
 void
 mix_newin(struct aproc *p, struct abuf *ibuf)
 {
+	struct abuf *obuf = LIST_FIRST(&p->obuflist);
+
+	if (!obuf || ibuf->cmin < obuf->cmin || ibuf->cmax > obuf->cmax) {
+		fprintf(stderr, "mix_newin: channel ranges mismatch\n");
+		abort();
+	}
 	ibuf->mixdone = 0;
 	ibuf->mixvol = ADATA_UNIT;
 	ibuf->xrun = XRUN_IGNORE;
@@ -589,22 +603,34 @@ mix_setmaster(struct aproc *p)
 void
 sub_bcopy(struct abuf *ibuf, struct abuf *obuf)
 {
-	unsigned char *idata, *odata;
+	short *idata, *odata;
+	unsigned i, j, ocnt, inext, istart;
 	unsigned icount, ocount, scount;
 
-	idata = abuf_rgetblk(ibuf, &icount, obuf->subdone);
-	icount -= icount % ibuf->bpf;
+	idata = (short *)abuf_rgetblk(ibuf, &icount, obuf->subdone);
+	icount /= ibuf->bpf;
 	if (icount == 0)
 		return;
-	odata = abuf_wgetblk(obuf, &ocount, 0);
-	ocount -= icount % obuf->bpf;
+	odata = (short *)abuf_wgetblk(obuf, &ocount, 0);
+	ocount /= obuf->bpf;
 	if (ocount == 0)
 		return;
+	istart = obuf->cmin - ibuf->cmin;
+	inext = ibuf->cmax - obuf->cmax + istart;
+	ocnt = obuf->cmax - obuf->cmin + 1;
 	scount = (icount < ocount) ? icount : ocount;
-	memcpy(odata, idata, scount);
-	abuf_wcommit(obuf, scount);	
-	obuf->subdone += scount;
-	DPRINTFN(4, "sub_bcopy: %u bytes\n", scount);
+	idata += istart;
+	for (i = scount; i > 0; i--) {
+		for (j = ocnt; j > 0; j--) {
+			*odata = *idata;
+			odata++;
+			idata++;
+		}
+		idata += inext;
+	}
+	abuf_wcommit(obuf, scount * obuf->bpf);	
+	obuf->subdone += scount * ibuf->bpf;
+	DPRINTFN(4, "sub_bcopy: %u frames\n", scount);
 }
 
 int
@@ -713,6 +739,12 @@ sub_hup(struct aproc *p, struct abuf *obuf)
 void
 sub_newout(struct aproc *p, struct abuf *obuf)
 {
+	struct abuf *ibuf = LIST_FIRST(&p->ibuflist);
+
+	if (!ibuf || obuf->cmin < ibuf->cmin || obuf->cmax > ibuf->cmax) {
+		fprintf(stderr, "sub_newout: channel ranges mismatch\n");
+		abort();
+	}
 	obuf->subdone = 0;
 	obuf->xrun = XRUN_IGNORE;
 }
diff --git a/usr.bin/aucat/aucat.c b/usr.bin/aucat/aucat.c
index 22757cc9d75..5e41f46c65a 100644
--- a/usr.bin/aucat/aucat.c
+++ b/usr.bin/aucat/aucat.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: aucat.c,v 1.30 2008/10/26 12:38:38 ratchov Exp $	*/
+/*	$OpenBSD: aucat.c,v 1.31 2008/11/03 22:25:13 ratchov Exp $	*/
 /*
  * Copyright (c) 2008 Alexandre Ratchov <alex@caoua.org>
  *
@@ -239,7 +239,7 @@ newinput(struct farg *fa)
 	 */
 	f = wav_new_in(&wav_ops, fd, fa->name, &fa->par, fa->hdr);
 	nfr = dev_bufsz * fa->par.rate / dev_rate;
-	buf = abuf_new(nfr, aparams_bpf(&fa->par));
+	buf = abuf_new(nfr, &fa->par);
 	proc = rpipe_new((struct file *)f);
 	aproc_setout(proc, buf);
 	abuf_fill(buf); /* XXX: move this in dev_attach() ? */
@@ -275,7 +275,7 @@ newoutput(struct farg *fa)
 	f = wav_new_out(&wav_ops, fd, fa->name, &fa->par, fa->hdr);
 	nfr = dev_bufsz * fa->par.rate / dev_rate;
 	proc = wpipe_new((struct file *)f);
-	buf = abuf_new(nfr, aparams_bpf(&fa->par));
+	buf = abuf_new(nfr, &fa->par);
 	aproc_setin(proc, buf);
 	dev_attach(fa->name, NULL, NULL, 0, buf, &fa->par, fa->xrun);
 }
diff --git a/usr.bin/aucat/dev.c b/usr.bin/aucat/dev.c
index 0f418a63fc8..70391760653 100644
--- a/usr.bin/aucat/dev.c
+++ b/usr.bin/aucat/dev.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: dev.c,v 1.3 2008/10/26 08:49:43 ratchov Exp $	*/
+/*	$OpenBSD: dev.c,v 1.4 2008/11/03 22:25:13 ratchov Exp $	*/
 /*
  * Copyright (c) 2008 Alexandre Ratchov <alex@caoua.org>
  *
@@ -138,14 +138,14 @@ dev_init(char *devpath,
 		 * create the read end
 		 */
 		dev_rec = rpipe_new(dev_file);
-		buf = abuf_new(nfr, aparams_bpf(dipar));
+		buf = abuf_new(nfr, dipar);
 		aproc_setout(dev_rec, buf);
 		ibufsz += nfr;
 
 		/*
 		 * append a converter, if needed
 		 */
-		if (!aparams_eq(dipar, &ipar)) {
+		if (!aparams_eqenc(dipar, &ipar)) {
 			if (debug_level > 0) {
 				fprintf(stderr, "%s: ", devpath);
 				aparams_print2(dipar, &ipar);
@@ -153,7 +153,7 @@ dev_init(char *devpath,
 			}
 			conv = conv_new("subconv", dipar, &ipar);
 			aproc_setin(conv, buf);
-			buf = abuf_new(nfr, aparams_bpf(&ipar));
+			buf = abuf_new(nfr, &ipar);
 			aproc_setout(conv, buf);
 			ibufsz += nfr;
 		}
@@ -178,14 +178,14 @@ dev_init(char *devpath,
 		 * create the write end
 		 */
 		dev_play = wpipe_new(dev_file);
-		buf = abuf_new(nfr, aparams_bpf(dopar));
+		buf = abuf_new(nfr, dopar);
 		aproc_setin(dev_play, buf);
 		obufsz += nfr;
 		
 		/*
 		 * append a converter, if needed
 		 */
-		if (!aparams_eq(&opar, dopar)) {
+		if (!aparams_eqenc(&opar, dopar)) {
 			if (debug_level > 0) {
 				fprintf(stderr, "%s: ", devpath);
 				aparams_print2(&opar, dopar);
@@ -193,7 +193,7 @@ dev_init(char *devpath,
 			}
 			conv = conv_new("mixconv", &opar, dopar);
 			aproc_setout(conv, buf);
-			buf = abuf_new(nfr, aparams_bpf(&opar));
+			buf = abuf_new(nfr, &opar);
 			aproc_setin(conv, buf);
 			obufsz += nfr;
 		}
@@ -382,7 +382,9 @@ dev_attach(char *name,
 	
 	if (ibuf) {
 		pbuf = LIST_FIRST(&dev_mix->obuflist);		
-		if (!aparams_eq(ipar, &dev_opar)) {
+		if (!aparams_eqenc(ipar, &dev_opar) ||
+		    !aparams_eqrate(ipar, &dev_opar) ||
+		    !aparams_subset(ipar, &dev_opar)) {
 			if (debug_level > 1) {
 				fprintf(stderr, "dev_attach: %s: ", name);
 				aparams_print2(ipar, &dev_opar);
@@ -392,7 +394,7 @@ dev_attach(char *name,
 			nfr -= nfr % dev_round;
 			conv = conv_new(name, ipar, &dev_opar);
 			aproc_setin(conv, ibuf);
-			ibuf = abuf_new(nfr, aparams_bpf(&dev_opar));
+			ibuf = abuf_new(nfr, &dev_opar);
 			aproc_setout(conv, ibuf);
 			/* XXX: call abuf_fill() here ? */
 		}
@@ -402,7 +404,9 @@ dev_attach(char *name,
 	}
 	if (obuf) {
 		rbuf = LIST_FIRST(&dev_sub->ibuflist);
-		if (!aparams_eq(opar, &dev_ipar)) {
+		if (!aparams_eqenc(opar, &dev_ipar) ||
+		    !aparams_eqrate(opar, &dev_ipar) ||
+		    !aparams_subset(opar, &dev_ipar)) {
 			if (debug_level > 1) {
 				fprintf(stderr, "dev_attach: %s: ", name);
 				aparams_print2(&dev_ipar, opar);
@@ -412,7 +416,7 @@ dev_attach(char *name,
 			nfr -= nfr % dev_round;
 			conv = conv_new(name, &dev_ipar, opar);
 			aproc_setout(conv, obuf);
-			obuf = abuf_new(nfr, aparams_bpf(&dev_ipar));
+			obuf = abuf_new(nfr, &dev_ipar);
 			aproc_setin(conv, obuf);
 		}
 		aproc_setout(dev_sub, obuf);
diff --git a/usr.bin/aucat/sock.c b/usr.bin/aucat/sock.c
index 7c9aecce3ea..76b3f3764a9 100644
--- a/usr.bin/aucat/sock.c
+++ b/usr.bin/aucat/sock.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: sock.c,v 1.1 2008/10/26 08:49:44 ratchov Exp $	*/
+/*	$OpenBSD: sock.c,v 1.2 2008/11/03 22:25:13 ratchov Exp $	*/
 /*
  * Copyright (c) 2008 Alexandre Ratchov <alex@caoua.org>
  *
@@ -324,13 +324,13 @@ sock_allocbuf(struct sock *f)
 
 	if (f->mode & AMSG_PLAY) {
 		nfr = f->bufsz - dev_bufsz * f->rpar.rate / dev_rate;
-		rbuf = abuf_new(nfr, aparams_bpf(&f->rpar));
+		rbuf = abuf_new(nfr, &f->rpar);
 		aproc_setout(f->pipe.file.rproc, rbuf);
 		f->odelta = 0;
 	}
 	if (f->mode & AMSG_REC) {
 		nfr = f->bufsz - dev_bufsz * f->wpar.rate / dev_rate;
-		wbuf = abuf_new(nfr, aparams_bpf(&f->wpar));
+		wbuf = abuf_new(nfr, &f->wpar);
 		aproc_setin(f->pipe.file.wproc, wbuf);
 		f->idelta = 0;
 	}