optimize rendering

Makefile

@@ -1,12 +1,12 @@
 CC = 9c
 LD = 9l
-CFLAGS = -Wall -Wextra -O2
+CFLAGS = -Wall -Wextra -O2 -g
 PROG = strans
 
 SRCS = $(wildcard *.c)
 OBJS = $(SRCS:.c=.o)
 
-all: $(PROG)
+all: $(PROG) xim bench 
 
 $(PROG): $(OBJS)
 	$(LD) -o $@ $(OBJS) -lthread -lString -lbio -lxcb -lm
@@ -15,5 +15,13 @@ $(OBJS): dat.h fn.h ipc.h
 
 clean:
 	rm -f $(OBJS) $(PROG)
+	make -C xim/ clean
+	make -C bench/ clean
 
-.PHONY: all clean
+xim:
+	make -C xim/
+
+bench:
+	make -C bench/
+
+.PHONY: all clean xim bench

bench.sh

@@ -0,0 +1,23 @@
+#!/bin/sh
+
+pkill strans
+pkill strans-xim
+sleep 1
+
+./strans map font &
+sleep 1
+
+# warm up glyph cache
+./bench/bench bench/bench.keys 100
+echo "cache warmed up"
+
+STRANS_PID=$(pgrep -x strans)
+perf record -g -o bench/perf.data -p "$STRANS_PID" &
+PERF_PID=$!
+sleep 1
+
+./bench/bench bench/bench.keys 100
+
+kill -INT $PERF_PID
+wait $PERF_PID 2>/dev/null
+pkill strans

bench/Makefile

@@ -0,0 +1,8 @@
+CC = cc
+CFLAGS = -Wall -O2
+
+bench: main.c
+	$(CC) $(CFLAGS) -o $@ main.c
+
+clean:
+	rm -f bench

bench/bench.keys

@@ -0,0 +1,30 @@
+^ntoukyounotenki^Ryoidesu^R
+konpyuutanopuroguramu^B^B^Bgramu^R
+nihongowobenkyoushiteimasu^R
+sakuranoshanohana^B^B^B^Bhanagasaiteiru^R
+kaigi^Ekyounokaigi^R
+ashitahaiitenkininarudeshou^R
+^sdkssudgktpdy^Rgksrmfekfqnxm^R
+vmfhrrh^Btprtmxm^Rdufrlavldml^R
+dmlrlwkekfwprtm^R
+gksrnl^B^B^Bdlfp^R
+answkd^Eanswkd^R
+qkrtlwkf^B^B^Bwkfgkf^R
+^thello world^R
+the quick brown fox jumps over the lazy dog^R
+programming is fun^R
+^vVietNam^Rxinchaobancokhoekhong^R
+toidilambaitap^B^B^Bbaitap^R
+hoctiengviet^R
+^nkaishanitsuutomeshiteimasu^R
+tanaboraguukoudesu^B^B^B^B^Bkoudesu^R
+^srhksrnrdlTkfgkwl^R
+answkddkssud^B^Bdkssud^R
+wnsdydgktpdy^R
+^vtoidenlopcuahang^R
+motngaymoiconduong^R
+^thigh performance input method^R
+^nnyuuryokuhouhou^R
+tesutonosuuretsu^R
+^sxkfldzmsdml^R
+qhfmfjxm^R

bench/main.c

@@ -0,0 +1,168 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+#include <time.h>
+#include <sys/socket.h>
+#include <sys/un.h>
+
+enum {
+	Kback	= 0xf008,
+	Kret	= 0xf00d,
+	Kesc	= 0xf01b,
+	Mctrl	= 1<<2,
+};
+
+typedef struct Key Key;
+struct Key {
+	int	k;
+	int	mod;
+};
+
+static Key *keys;
+static int nkeys;
+static int fd;
+
+static void
+die(char *msg)
+{
+	perror(msg);
+	exit(1);
+}
+
+static void
+addkey(int k, int mod)
+{
+	static int cap;
+
+	if(nkeys >= cap){
+		cap = cap ? cap * 2 : 256;
+		keys = realloc(keys, cap * sizeof(Key));
+		if(!keys)
+			die("realloc");
+	}
+	keys[nkeys].k = k;
+	keys[nkeys].mod = mod;
+	nkeys++;
+}
+
+static void
+loadkeys(char *file)
+{
+	FILE *f;
+	int c;
+
+	f = fopen(file, "r");
+	if(!f)
+		die(file);
+	while((c = fgetc(f)) != EOF){
+		if(c == '\n' || c == '\r')
+			continue;
+		if(c != '^'){
+			addkey(c, 0);
+			continue;
+		}
+		c = fgetc(f);
+		if(c == EOF)
+			break;
+		switch(c){
+		case 'B':
+			addkey(Kback, 0);
+			break;
+		case 'R':
+			addkey(Kret, 0);
+			break;
+		case 'E':
+			addkey(Kesc, 0);
+			break;
+		default:
+			addkey(c, Mctrl);
+			break;
+		}
+	}
+	fclose(f);
+}
+
+static void
+dial(void)
+{
+	struct sockaddr_un addr;
+
+	fd = socket(AF_UNIX, SOCK_STREAM, 0);
+	if(fd < 0)
+		die("socket");
+	memset(&addr, 0, sizeof(addr));
+	addr.sun_family = AF_UNIX;
+	snprintf(addr.sun_path, sizeof(addr.sun_path),
+		"/tmp/strans.%d", getuid());
+	if(connect(fd, (struct sockaddr*)&addr, sizeof(addr)) < 0)
+		die("connect");
+}
+
+static void
+sendkey(int key, int mod)
+{
+	unsigned char req[4];
+
+	req[0] = 0;
+	req[1] = mod;
+	req[2] = key & 0xff;
+	req[3] = (key >> 8) & 0xff;
+	if(write(fd, req, 4) != 4)
+		die("write");
+}
+
+static int
+readresp(void)
+{
+	unsigned char buf[256];
+	int n;
+
+	if(read(fd, buf, 2) != 2)
+		return -1;
+	n = buf[1];
+	if(n > 0 && read(fd, buf + 2, n) != n)
+		return -1;
+	return 0;
+}
+
+static double
+now(void)
+{
+	struct timespec ts;
+
+	clock_gettime(CLOCK_MONOTONIC, &ts);
+	return ts.tv_sec + ts.tv_nsec * 1e-9;
+}
+
+int
+main(int argc, char **argv)
+{
+	int i, j, niter;
+	double t0, t1, dt;
+
+	if(argc < 2){
+		fprintf(stderr, "usage: bench file [niter]\n");
+		exit(1);
+	}
+
+	loadkeys(argv[1]);
+	niter = argc > 2 ? atoi(argv[2]) : 1000;
+	dial();
+	t0 = now();
+	for(i = 0; i < niter; i++)
+		for(j = 0; j < nkeys; j++){
+			sendkey(keys[j].k, keys[j].mod);
+			if(readresp() < 0){
+				fprintf(stderr, "failed at iter %d key %d\n", i, j);
+				exit(1);
+			}
+		}
+	t1 = now();
+	dt = t1 - t0;
+	printf("%d iters x %d keys = %d keys\n", niter, nkeys, niter * nkeys);
+	printf("%.3f ms total, %.3f us/key, %.3f us/iter\n",
+		dt * 1000, dt * 1e6 / (niter * nkeys), dt * 1e6 / niter);
+	close(fd);
+	return 0;
+}

dat.h

@@ -117,7 +117,7 @@ typedef struct Drawcmd Drawcmd;
 struct Drawcmd
 {
 	Str	pre;
-	Str	kouho[Maxkouho];
+	Str	kouho[Maxdisp];
 	int	nkouho;
 	int	sel;
 };

fn.h

@@ -28,6 +28,7 @@ void	drawthread(void*);
 void	imthread(void*);
 Emit	transmap(Im*, Rune);
 Emit	transko(Im*, Rune);
+Emit	transvi(Im*, Rune);
 void	backko(Im*);
 void	dictsend(Im*, Str*);
 

font.c

@@ -99,7 +99,8 @@ void
 putfont(u32int *buf, int w, int h, int px, int py, Rune r)
 {
 	Glyph *g;
-	int i, j, x, y, a, sel, f;
+	int i, j, a, sel, f;
+	int y0, j0, j1, x0, i0, i1;
 	u32int *p;
 
 	if(r >= Nglyphs)
@@ -116,17 +117,18 @@ putfont(u32int *buf, int w, int h, int px, int py, Rune r)
 		if(g->bmp == nil)
 			return;
 	}
-	for(j = 0; j < g->h; j++){
-		y = py + j + g->oy + Fontsz - Fontbase;
-		if(y < 0 || y >= h)
-			continue;
-		for(i = 0; i < g->w; i++){
-			x = px + i + g->ox;
-			if(x < 0 || x >= w)
-				continue;
+
+	y0 = py + g->oy + Fontsz - Fontbase;
+	j0 = y0 < 0 ? -y0 : 0;
+	j1 = y0 + g->h > h ? h - y0 : g->h;
+	x0 = px + g->ox;
+	i0 = x0 < 0 ? -x0 : 0;
+	i1 = x0 + g->w > w ? w - x0 : g->w;
+	for(j = j0; j < j1; j++){
+		for(i = i0; i < i1; i++){
 			a = g->bmp[j * g->w + i];
 			if(a > 0){
-				p = &buf[y * w + x];
+				p = &buf[(y0 + j) * w + x0 + i];
 				sel = (*p == Colsel) ? 1 : 0;
 				*p = blendtab[sel][a];
 			}

main.c

@@ -61,7 +61,7 @@ threadmain(int argc, char **argv)
 		usage();
 
 	fontdir = argv[2];
-	drawc = chancreate(sizeof(Drawcmd), 0);
+	drawc = chancreate(sizeof(Drawcmd), 4);
 	keyc = chancreate(sizeof(Keyreq), 0);
 	dictreqc = chancreate(sizeof(Dictreq), 4);
 	dictresc = chancreate(sizeof(Dictres), 0);

strans.c

@@ -16,7 +16,7 @@ Lang langs[] = {
 	{LangJPK,   "kata",    "kanji",  transmap, backmap, dictqmap, nil, nil},
 	{LangKO,    "hangul",  nil,      transko,  backko,  dictqmap, nil, nil},
 	{LangEMOJI, "emoji",   "emoji",  transmap, backmap, dictqmap, nil, nil},
-	{LangVI,    "telex",   nil,      transmap, backmap, dictqmap, nil, nil},
+	{LangVI,    "telex",   nil,      transvi,  backmap, dictqmap, nil, nil},
 };
 int nlang = nelem(langs);
 
@@ -24,7 +24,7 @@ static void
 clearkouho(void)
 {
 	im.nkouho = 0;
-	im.sel = 0;
+	im.sel = -1;
 }
 
 static void
@@ -43,15 +43,18 @@ static void
 show(void)
 {
 	Drawcmd dc;
-	int i;
+	int i, first, n;
 
 	sclear(&dc.pre);
 	if(!mapget(im.l->map, &im.pre, &dc.pre))
 		dc.pre = im.pre;
-	dc.nkouho = im.nkouho;
-	dc.sel = im.sel;
-	for(i = 0; i < dc.nkouho; i++)
-		dc.kouho[i] = im.kouho[i];
+	first = im.sel >= Maxdisp ? im.sel - Maxdisp + 1 : 0;
+	n = im.nkouho - first;
+	if(n > Maxdisp) n = Maxdisp;
+	dc.nkouho = n;
+	dc.sel = im.sel >= 0 ? im.sel - first : -1;
+	for(i = 0; i < n; i++)
+		dc.kouho[i] = im.kouho[first + i];
 	chansend(drawc, &dc);
 }
 
@@ -200,7 +203,7 @@ keystroke(u32int ks, u32int mod, Str *com)
 			return 0;
 		if(ks == Kdown && im.sel < im.nkouho - 1)
 			im.sel++;
-		if(ks == Kup && im.sel > 0)
+		if(ks == Kup && im.sel >= 0)
 			im.sel--;
 		show();
 		return 1;
@@ -257,6 +260,11 @@ keystroke(u32int ks, u32int mod, Str *com)
 			return 1;
 		return 0;
 	}
+	if(ks == '0' && im.nkouho > 0){
+		commit(com);
+		reset();
+		return 1;
+	}
 	if(ks > 0x7f || ks == ' '){
 		commit(com);
 		sputr(com, ks);

vi.c

@@ -0,0 +1,160 @@
+#include "dat.h"
+#include "fn.h"
+
+static struct {
+	Rune	base;
+	Rune	tone[5];	/* s f r x j */
+} vitone[] = {
+	{L'a', {L'á', L'à', L'ả', L'ã', L'ạ'}},
+	{L'â', {L'ấ', L'ầ', L'ẩ', L'ẫ', L'ậ'}},
+	{L'ă', {L'ắ', L'ằ', L'ẳ', L'ẵ', L'ặ'}},
+	{L'e', {L'é', L'è', L'ẻ', L'ẽ', L'ẹ'}},
+	{L'ê', {L'ế', L'ề', L'ể', L'ễ', L'ệ'}},
+	{L'i', {L'í', L'ì', L'ỉ', L'ĩ', L'ị'}},
+	{L'o', {L'ó', L'ò', L'ỏ', L'õ', L'ọ'}},
+	{L'ô', {L'ố', L'ồ', L'ổ', L'ỗ', L'ộ'}},
+	{L'ơ', {L'ớ', L'ờ', L'ở', L'ỡ', L'ợ'}},
+	{L'u', {L'ú', L'ù', L'ủ', L'ũ', L'ụ'}},
+	{L'ư', {L'ứ', L'ừ', L'ử', L'ữ', L'ự'}},
+	{L'y', {L'ý', L'ỳ', L'ỷ', L'ỹ', L'ỵ'}},
+	{L'A', {L'Á', L'À', L'Ả', L'Ã', L'Ạ'}},
+	{L'Â', {L'Ấ', L'Ầ', L'Ẩ', L'Ẫ', L'Ậ'}},
+	{L'Ă', {L'Ắ', L'Ằ', L'Ẳ', L'Ẵ', L'Ặ'}},
+	{L'E', {L'É', L'È', L'Ẻ', L'Ẽ', L'Ẹ'}},
+	{L'Ê', {L'Ế', L'Ề', L'Ể', L'Ễ', L'Ệ'}},
+	{L'I', {L'Í', L'Ì', L'Ỉ', L'Ĩ', L'Ị'}},
+	{L'O', {L'Ó', L'Ò', L'Ỏ', L'Õ', L'Ọ'}},
+	{L'Ô', {L'Ố', L'Ồ', L'Ổ', L'Ỗ', L'Ộ'}},
+	{L'Ơ', {L'Ớ', L'Ờ', L'Ở', L'Ỡ', L'Ợ'}},
+	{L'U', {L'Ú', L'Ù', L'Ủ', L'Ũ', L'Ụ'}},
+	{L'Ư', {L'Ứ', L'Ừ', L'Ử', L'Ữ', L'Ự'}},
+	{L'Y', {L'Ý', L'Ỳ', L'Ỷ', L'Ỹ', L'Ỵ'}},
+};
+
+static int
+istone(Rune c)
+{
+	return c == 's' || c == 'f' || c == 'r' || c == 'x' || c == 'j';
+}
+
+static int
+toneidx(Rune c)
+{
+	switch(c){
+	case 's': return 0;
+	case 'f': return 1;
+	case 'r': return 2;
+	case 'x': return 3;
+	case 'j': return 4;
+	}
+	return -1;
+}
+
+static int
+isvowel(Rune c)
+{
+	int i;
+
+	for(i = 0; i < nelem(vitone); i++)
+		if(vitone[i].base == c)
+			return 1;
+	return 0;
+}
+
+static Rune
+removetone(Rune c)
+{
+	int i, j;
+
+	for(i = 0; i < nelem(vitone); i++){
+		if(vitone[i].base == c)
+			return c;
+		for(j = 0; j < 5; j++)
+			if(vitone[i].tone[j] == c)
+				return vitone[i].base;
+	}
+	return c;
+}
+
+static Rune
+applytone(Rune c, int tidx)
+{
+	int i;
+	Rune base;
+
+	base = removetone(c);
+	for(i = 0; i < nelem(vitone); i++)
+		if(vitone[i].base == base)
+			return vitone[i].tone[tidx];
+	return c;
+}
+
+Emit
+transvi(Im *im, Rune c)
+{
+	Emit e;
+	Str mapped, pre;
+	int i, tidx, vi, last, penult;
+	Rune v, b1, b2;
+
+	if(!istone(c) && c != 'z')
+		return transmap(im, c);
+
+	memset(&e, 0, sizeof e);
+	if(im->pre.n == 0){
+		sputr(&e.s, c);
+		return e;
+	}
+	if(im->pre.r[im->pre.n - 1] == '\\'){
+		pre = im->pre;
+		pre.n--;
+		if(!mapget(im->l->map, &pre, &mapped))
+			mapped = pre;
+		sputr(&mapped, c);
+		e.eat = 1;
+		e.s = mapped;
+		return e;
+	}
+	if(!mapget(im->l->map, &im->pre, &mapped))
+		mapped = im->pre;
+	last = -1;
+	penult = -1;
+	for(i = 0; i < mapped.n; i++){
+		v = removetone(mapped.r[i]);
+		if(isvowel(v)){
+			penult = last;
+			last = i;
+		}
+	}
+	vi = -1;
+	if(last >= 0){
+		if(last < mapped.n - 1 || penult < 0)
+			vi = last;
+		else{
+			b1 = removetone(mapped.r[penult]);
+			b2 = removetone(mapped.r[last]);
+			if((b1 == 'o' || b1 == 'O') && (b2 == 'a' || b2 == 'A' || b2 == 'e' || b2 == 'E'))
+				vi = last;
+			else if((b1 == 'u' || b1 == 'U') && (b2 == 'y' || b2 == 'Y'))
+				vi = last;
+			else
+				vi = penult;
+		}
+	}
+	if(vi < 0){
+		e.eat = 1;
+		e.s = mapped;
+		sputr(&e.s, c);
+		return e;
+	}
+
+	if(c == 'z')
+		mapped.r[vi] = removetone(mapped.r[vi]);
+	else{
+		tidx = toneidx(c);
+		mapped.r[vi] = applytone(mapped.r[vi], tidx);
+	}
+	e.eat = 1;
+	e.next = mapped;
+	return e;
+}

win.c

@@ -72,8 +72,10 @@ drawkouho(Drawcmd *dc, int first, int n, int w, int h)
 
 	memset(img, (uchar)Colbg, w * h * sizeof(u32int));
 	drawstr(img, 0, 0, dc->pre.r, dc->pre.n, w, h);
-	sely = Fontsz + (dc->sel - first) * Fontsz;
+	if(dc->sel >= 0){
+		sely = Fontsz + dc->sel * Fontsz;
 		memset(img + sely * w, (uchar)Colsel, Fontsz * w * sizeof(u32int));
+	}
 	for(i = 0, y = Fontsz; i < n; i++, y += Fontsz){
 		s = &dc->kouho[first+i];
 		putfont(img, w, h, 0, y, '1' + i + Asciitofull);
@@ -99,17 +101,16 @@ winhide(void)
 static void
 winshow(Drawcmd *dc)
 {
-	int px, py, w, h, i, n, first, maxw;
+	int px, py, w, h, i, n, maxw;
 	u32int vals[4];
 	xcb_query_pointer_reply_t *ptr;
 	xcb_query_pointer_cookie_t cookie;
 
 	cookie = xcb_query_pointer(conn, scr->root);
-	first = dc->sel >= Maxdisp ? dc->sel - Maxdisp + 1 : 0;
-	n = min(dc->nkouho - first, Maxdisp);
+	n = dc->nkouho;
 	maxw = dc->pre.n;
 	for(i = 0; i < n; i++)
-		maxw = max(maxw, dc->kouho[first+i].n);
+		maxw = max(maxw, dc->kouho[i].n);
 	ptr = xcb_query_pointer_reply(conn, cookie, nil);
 	if(ptr == nil)
 		die("xcb_query_pointer");
@@ -125,7 +126,7 @@ winshow(Drawcmd *dc)
 		XCB_CONFIG_WINDOW_WIDTH | XCB_CONFIG_WINDOW_HEIGHT,
 		vals);
 	xcb_map_window(conn, win);
-	drawkouho(dc, first, n, w, h);
+	drawkouho(dc, 0, n, w, h);
 	putimage(w, h);
 }
 
@@ -137,6 +138,8 @@ drawthread(void*)
 	threadsetname("draw");
 	wininit();
 	while(chanrecv(drawc, &dc) > 0){
+		while(channbrecv(drawc, &dc) > 0)
+			;
 		if(dc.nkouho == 0 && dc.pre.n == 0)
 			winhide();
 		else