yaft ( http://uobikiemukot.github.io/yaft/ ) on OpenBSD/luna88k, take 2. 8bpp version.

yaft-luna88k-8bpp-20140810.diff

diff --git a/draw.h b/draw.h
index 9871ba3..9d69618 100644
--- a/draw.h
+++ b/draw.h
@@ -1,6 +1,8 @@
 /* See LICENSE for licence details. */
 static inline void draw_sixel(struct framebuffer *fb, int line, int col, uint8_t *bitmap)
 {
+	/* XXX: Maybe need to re-write on LUNA */
+
 	int h, w, src_offset, dst_offset;
 	uint32_t pixel, color = 0;
 
@@ -16,16 +18,33 @@ static inline void draw_sixel(struct framebuffer *fb, int line, int col, uint8_t
 	}
 }
 
+/* Based on OpenBSD:src/sys/arch/luna88k/dev/omrasops.c */
+
+#define	ALL1BITS	(~0U)
+#define	BLITWIDTH	(32)
+#define	ALIGNMASK	(0x1f)
+#define	BYTESDONE	(4)
+#define P0(addr) ((u_int32_t *)(addr))
+#define P1(addr) ((u_int32_t *)((u_int8_t *)(addr) +  0x40000))
+#define P2(addr) ((u_int32_t *)((u_int8_t *)(addr) +  0x80000))
+#define P3(addr) ((u_int32_t *)((u_int8_t *)(addr) +  0xC0000))
+#define P4(addr) ((u_int32_t *)((u_int8_t *)(addr) + 0x100000))
+#define P5(addr) ((u_int32_t *)((u_int8_t *)(addr) + 0x140000))
+#define P6(addr) ((u_int32_t *)((u_int8_t *)(addr) + 0x180000))
+#define P7(addr) ((u_int32_t *)((u_int8_t *)(addr) + 0x1C0000))
+
 static inline void draw_line(struct framebuffer *fb, struct terminal *term, int line)
 {
-	int pos, size, bdf_padding, glyph_width, margin_right;
-	int col, w, h;
-	uint32_t pixel;
+	int pos, size, col;
 	struct color_pair_t color_pair;
 	struct cell_t *cellp;
 
+	u_int8_t *p;
+	int x, y, width, height, align;
+	u_int32_t lmask, rmask, glyph, glyphbg, fgpat, bgpat;
+	int fg, bg, plane;
+
 	for (col = term->cols - 1; col >= 0; col--) {
-		margin_right = (term->cols - 1 - col) * CELL_WIDTH;
 
 		/* target cell */
 		cellp = &term->cells[col + line * term->cols];
@@ -39,11 +58,16 @@ static inline void draw_line(struct framebuffer *fb, struct terminal *term, int
 		/* copy current color_pair (maybe changed) */
 		color_pair = cellp->color_pair;
 
+#if 0
 		/* check wide character or not */
 		glyph_width = (cellp->width == HALF) ? CELL_WIDTH: CELL_WIDTH * 2;
 		bdf_padding = my_ceil(glyph_width, BITS_PER_BYTE) * BITS_PER_BYTE - glyph_width;
 		if (cellp->width == WIDE)
 			bdf_padding += CELL_WIDTH;
+#endif
+		/* check wide character or not */
+		if (cellp->width == NEXT_TO_WIDE)
+			continue;
 
 		/* check cursor positon */
 		if ((term->mode & MODE_CURSOR && line == term->cursor.y)
@@ -54,40 +78,159 @@ static inline void draw_line(struct framebuffer *fb, struct terminal *term, int
 			color_pair.bg = (!tty.visible && BACKGROUND_DRAW) ? PASSIVE_CURSOR_COLOR: ACTIVE_CURSOR_COLOR;
 		}
 
-		for (h = 0; h < CELL_HEIGHT; h++) {
-			/* if UNDERLINE attribute on, swap bg/fg */
-			if ((h == (CELL_HEIGHT - 1)) && (cellp->attribute & attr_mask[ATTR_UNDERLINE]))
-				color_pair.bg = color_pair.fg;
-
-			for (w = 0; w < CELL_WIDTH; w++) {
-				pos = (term->width - 1 - margin_right - w) * fb->bytes_per_pixel
-					+ (line * CELL_HEIGHT + h) * fb->line_length;
-
-				/* set color palette */
-				if (cellp->glyphp->bitmap[h] & (0x01 << (bdf_padding + w)))
-					pixel = term->color_palette[color_pair.fg];
-				else if (fb->wall && color_pair.bg == DEFAULT_BG) /* wallpaper */
-					memcpy(&pixel, fb->wall + pos, fb->bytes_per_pixel);
-				else
-					pixel = term->color_palette[color_pair.bg];
-
-				/* update copy buffer only */
-				memcpy(fb->buf + pos, &pixel, fb->bytes_per_pixel);
+		/* color palette */
+		fg = term->color_palette[color_pair.fg];
+		bg = term->color_palette[color_pair.bg];
+#if 0
+		if (fb->wall && color_pair.bg == DEFAULT_BG) /* wallpaper */
+			memcpy(&pixel, fb->wall + pos, fb->bytes_per_pixel);
+#endif
+
+		x = CELL_WIDTH * col;
+		y = CELL_HEIGHT * line;
+
+		p = (u_int8_t *)fb->buf + y * fb->line_length + ((x / 32) * 4);
+		align = x & ALIGNMASK;
+		width = cellp->glyphp->width * CELL_WIDTH + align;
+		lmask = ALL1BITS >> align;
+		rmask = ALL1BITS << (-width & ALIGNMASK);
+		height = 0;
+
+		if (width <= BLITWIDTH) {
+			lmask &= rmask;
+			while (height < CELL_HEIGHT) {
+				/* if UNDERLINE attribute on, swap bg/fg */
+				if ((height == (CELL_HEIGHT - 1))
+					&& (cellp->attribute & attr_mask[ATTR_UNDERLINE]))
+						bg = fg;
+				glyph = (uint32_t)cellp->glyphp->bitmap[height];
+				/* shift leftmost */
+				glyph <<= 32 - cellp->glyphp->width * CELL_WIDTH;
+
+				glyph = glyph >> align;
+				glyphbg = glyph ^ ALL1BITS;
+
+				fgpat = (fg & 0x01) ? glyph : 0;
+				bgpat = (bg & 0x01) ? glyphbg : 0;
+				*P0(p) = (*P0(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x02) ? glyph : 0;
+				bgpat = (bg & 0x02) ? glyphbg : 0;
+				*P1(p) = (*P1(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x04) ? glyph : 0;
+				bgpat = (bg & 0x04) ? glyphbg : 0;
+				*P2(p) = (*P2(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x08) ? glyph : 0;
+				bgpat = (bg & 0x08) ? glyphbg : 0;
+				*P3(p) = (*P3(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x10) ? glyph : 0;
+				bgpat = (bg & 0x10) ? glyphbg : 0;
+				*P4(p) = (*P4(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x20) ? glyph : 0;
+				bgpat = (bg & 0x20) ? glyphbg : 0;
+				*P5(p) = (*P5(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x40) ? glyph : 0;
+				bgpat = (bg & 0x40) ? glyphbg : 0;
+				*P6(p) = (*P6(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x80) ? glyph : 0;
+				bgpat = (bg & 0x80) ? glyphbg : 0;
+				*P7(p) = (*P7(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+
+				p += fb->line_length;
+				height++;
+			}
+		} else {
+			u_int8_t *q = p;
+			u_int32_t lhalf, rhalf;
+			u_int32_t lhalfbg, rhalfbg;
+
+			while (height < CELL_HEIGHT) {
+				/* if UNDERLINE attribute on, swap bg/fg */
+				if ((height == (CELL_HEIGHT - 1))
+					&& (cellp->attribute & attr_mask[ATTR_UNDERLINE]))
+						bg = fg;
+				glyph = (uint32_t)cellp->glyphp->bitmap[height];
+				/* shift leftmost */
+				glyph <<= 32 - cellp->glyphp->width * CELL_WIDTH;
+
+				lhalf = glyph >> align;
+				lhalfbg = lhalf ^ ALL1BITS;
+
+				fgpat = (fg & 0x01) ? lhalf : 0;
+				bgpat = (bg & 0x01) ? lhalfbg : 0;
+				*P0(p) = (*P0(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x02) ? lhalf : 0;
+				bgpat = (bg & 0x02) ? lhalfbg : 0;
+				*P1(p) = (*P1(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x04) ? lhalf : 0;
+				bgpat = (bg & 0x04) ? lhalfbg : 0;
+				*P2(p) = (*P2(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x08) ? lhalf : 0;
+				bgpat = (bg & 0x08) ? lhalfbg : 0;
+				*P3(p) = (*P3(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x10) ? lhalf : 0;
+				bgpat = (bg & 0x10) ? lhalfbg : 0;
+				*P4(p) = (*P4(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x20) ? lhalf : 0;
+				bgpat = (bg & 0x20) ? lhalfbg : 0;
+				*P5(p) = (*P5(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x40) ? lhalf : 0;
+				bgpat = (bg & 0x40) ? lhalfbg : 0;
+				*P6(p) = (*P6(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x80) ? lhalf : 0;
+				bgpat = (bg & 0x80) ? lhalfbg : 0;
+				*P7(p) = (*P7(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+
+				p += BYTESDONE;
+
+				rhalf = glyph << (BLITWIDTH - align);
+				rhalfbg = rhalf ^ ALL1BITS;
+
+				fgpat = (fg & 0x01) ? rhalf : 0;
+				bgpat = (bg & 0x01) ? rhalfbg : 0;
+				*P0(p) = ((fgpat | bgpat) & rmask) | (*P0(p) & ~rmask);
+				fgpat = (fg & 0x02) ? rhalf : 0;
+				bgpat = (bg & 0x02) ? rhalfbg : 0;
+				*P1(p) = ((fgpat | bgpat) & rmask) | (*P1(p) & ~rmask);
+				fgpat = (fg & 0x04) ? rhalf : 0;
+				bgpat = (bg & 0x04) ? rhalfbg : 0;
+				*P2(p) = ((fgpat | bgpat) & rmask) | (*P2(p) & ~rmask);
+				fgpat = (fg & 0x08) ? rhalf : 0;
+				bgpat = (bg & 0x08) ? rhalfbg : 0;
+				*P3(p) = ((fgpat | bgpat) & rmask) | (*P3(p) & ~rmask);
+				fgpat = (fg & 0x10) ? rhalf : 0;
+				bgpat = (bg & 0x10) ? rhalfbg : 0;
+				*P4(p) = ((fgpat | bgpat) & rmask) | (*P4(p) & ~rmask);
+				fgpat = (fg & 0x20) ? rhalf : 0;
+				bgpat = (bg & 0x20) ? rhalfbg : 0;
+				*P5(p) = ((fgpat | bgpat) & rmask) | (*P5(p) & ~rmask);
+				fgpat = (fg & 0x40) ? rhalf : 0;
+				bgpat = (bg & 0x40) ? rhalfbg : 0;
+				*P6(p) = ((fgpat | bgpat) & rmask) | (*P6(p) & ~rmask);
+				fgpat = (fg & 0x80) ? rhalf : 0;
+				bgpat = (bg & 0x80) ? rhalfbg : 0;
+				*P7(p) = ((fgpat | bgpat) & rmask) | (*P7(p) & ~rmask);
+
+				p = (q += fb->line_length);
+				height++;
 			}
 		}
 	}
 
 	/* actual display update (bit blit) */
+#if 0
 	pos = (line * CELL_HEIGHT) * fb->line_length;
 	size = CELL_HEIGHT * fb->line_length;
 	memcpy(fb->fp + pos, fb->buf + pos, size);
-
-	/* TODO: page flip
-		if fb_fix_screeninfo.ypanstep > 0, we can use hardware panning.
-		set fb_fix_screeninfo.{yres_virtual,yoffset} and call ioctl(FBIOPAN_DISPLAY)
-		but drivers  of recent hardware (inteldrmfb, nouveaufb, radeonfb) don't support...
-		(we can use this by using libdrm)
-	*/
+#else
+	size = TERM_WIDTH / 8;
+	for (height = 0; height < CELL_HEIGHT; height++) {
+		for (plane = 0; plane < DEPTH; plane++) {
+			pos = (line * CELL_HEIGHT + height) * fb->line_length
+				+ 0x40000 * plane;
+			memcpy(fb->fp + pos, fb->buf + pos, size);
+		}
+	}
+#endif
 
 	term->line_dirty[line] = ((term->mode & MODE_CURSOR) && term->cursor.y == line) ? true: false;
 }
diff --git a/fb/openbsd.h b/fb/openbsd.h
index 5a11596..276f017 100644
--- a/fb/openbsd.h
+++ b/fb/openbsd.h
@@ -13,11 +13,17 @@ typedef unsigned long   u_long;
 #include  <dev/wscons/wsconsio.h>
 #include  <dev/wscons/wsksymdef.h>
 
-/* some structs for NetBSD */
+/* some structs for OpenBSD */
 enum term_size {
+#if 0
 	TERM_WIDTH  = 640,
 	TERM_HEIGHT = 480,
 	DEPTH = 8,
+#else
+	TERM_WIDTH  = 1280,
+	TERM_HEIGHT = 1024,
+	DEPTH = 8,
+#endif
 };
 
 enum fbtype_t {
@@ -45,6 +51,7 @@ struct fbinfo_t {
 
 struct framebuffer {
 	uint8_t *fp;          /* pointer of framebuffer(read only) */
+	uint8_t *fp_orig;     /* pointer of framebuffer(original) */
 	uint8_t *wall;        /* buffer for wallpaper */
 	uint8_t *buf;         /* copy of framebuffer */
 	int fd;               /* file descriptor of framebuffer */
@@ -83,7 +90,7 @@ uint8_t *load_wallpaper(struct framebuffer *fb)
 	return ptr;
 }
 
-/* some functions for NetBSD framebuffer */
+/* some functions for OpenBSD framebuffer */
 void cmap_create(struct wsdisplay_cmap **cmap, int size)
 {
 	*cmap                = (struct wsdisplay_cmap *) ecalloc(1, sizeof(struct wsdisplay_cmap));
@@ -197,12 +204,19 @@ void fb_init(struct framebuffer *fb, uint32_t *color_palette)
 		goto fb_init_error;
 	}
 
+	/* XXX: Should be check if WSDISPLAYIO_TYPE_LUNA ? */
+
 	fb->width  = TERM_WIDTH;
 	fb->height = TERM_HEIGHT;
 	fb->bytes_per_pixel = my_ceil(DEPTH, BITS_PER_BYTE);
 
+#if 0
 	fb->line_length = fb->bytes_per_pixel * fb->width;
 	fb->screen_size = fb->height * fb->line_length;
+#else
+	fb->line_length = 2048 / 8;
+	fb->screen_size = fb->height * fb->line_length * DEPTH;
+#endif
 	fb->vinfo = bpp_table[DEPTH];
 
 	if (DEBUG)
@@ -214,7 +228,7 @@ void fb_init(struct framebuffer *fb, uint32_t *color_palette)
 		fb->cmap = fb->cmap_org = NULL;
 		fb->vinfo.fbtype = FBTYPE_RGB;
 	}
-	else if (DEPTH == 8) {
+	else if (DEPTH == 8 || DEPTH == 4 || DEPTH == 1 ) {
 		cmap_create(&fb->cmap, COLORS);
 		cmap_create(&fb->cmap_org, finfo.cmsize);
 		cmap_init(fb);
@@ -226,7 +240,10 @@ void fb_init(struct framebuffer *fb, uint32_t *color_palette)
 	for (i = 0; i < COLORS; i++) /* init color palette */
 		color_palette[i] = (fb->bytes_per_pixel == 1) ? (uint32_t) i: color2pixel(&fb->vinfo, color_list[i]);
 
-	fb->fp    = (uint8_t *) emmap(0, fb->screen_size, PROT_WRITE | PROT_READ, MAP_SHARED, fb->fd, 0);
+	fb->fp_orig = (uint8_t *) emmap(0, fb->screen_size, PROT_WRITE | PROT_READ, MAP_SHARED, fb->fd, 0);
+
+	fb->fp = fb->fp_orig + 8; /* XXX: LUNA quirk; need 8 byte offset */
+
 	fb->buf   = (uint8_t *) ecalloc(1, fb->screen_size);
 	//fb->wall  = (WALLPAPER && fb->bytes_per_pixel > 1) ? load_wallpaper(fb): NULL;
 
@@ -251,6 +268,6 @@ void fb_die(struct framebuffer *fb)
 	}
 	free(fb->buf);
 	free(fb->wall);
-	emunmap(fb->fp, fb->screen_size);
+	emunmap(fb->fp_orig, fb->screen_size);
 	eclose(fb->fd);
 }
diff --git a/makefile b/makefile
index caf405f..8fc5679 100644
--- a/makefile
+++ b/makefile
@@ -1,7 +1,8 @@
 CC ?= gcc
 #CC ?= clang
 
-CFLAGS  ?= -std=c99 -pedantic -Wall -Wextra -O3 -s -pipe
+#CFLAGS  ?= -std=c99 -pedantic -Wall -Wextra -O3 -s -pipe
+CFLAGS  = -std=c99 -pedantic -Wall -Wextra -O3 -s -pipe
 LDFLAGS ?=
 
 XCFLAGS  ?= -std=c99 -pedantic -Wall -Wextra -I/usr/include/X11/ -O3 -s -pipe
diff --git a/yaft.c b/yaft.c
index 9acbad6..cbeaf9f 100644
--- a/yaft.c
+++ b/yaft.c
@@ -77,11 +77,13 @@ void tty_init(struct termios *save_tm)
 	vtm.mode   = VT_PROCESS;
 	vtm.waitv  = 0;
 	vtm.relsig = vtm.acqsig = vtm.frsig = SIGUSR1;
+#if 0
 	if (ioctl(STDIN_FILENO, VT_SETMODE, &vtm))
 		fatal("ioctl: VT_SETMODE failed (maybe here is not console)");
 
 	if (ioctl(STDIN_FILENO, KDSETMODE, KD_GRAPHICS))
 		fatal("ioctl: KDSETMODE failed (maybe here is not console)");
+#endif
 
 	etcgetattr(STDIN_FILENO, save_tm);
 	set_rawmode(STDIN_FILENO, save_tm);
@@ -102,8 +104,10 @@ void tty_die(struct termios *save_tm)
 	vtm.mode   = VT_AUTO;
 	vtm.waitv  = 0;
 	vtm.relsig = vtm.acqsig = vtm.frsig = 0;
+#if 0
 	ioctl(STDIN_FILENO, VT_SETMODE, &vtm);
 	ioctl(STDIN_FILENO, KDSETMODE, KD_TEXT);
+#endif
 
 	tcsetattr(STDIN_FILENO, TCSAFLUSH, save_tm);
 	fflush(stdout);

yaft-luna88k-8bpp-20140810a.diff

diff --git a/draw.h b/draw.h
index 9871ba3..b31bd02 100644
--- a/draw.h
+++ b/draw.h
@@ -1,6 +1,8 @@
 /* See LICENSE for licence details. */
 static inline void draw_sixel(struct framebuffer *fb, int line, int col, uint8_t *bitmap)
 {
+	/* XXX: Maybe need to re-write on LUNA */
+
 	int h, w, src_offset, dst_offset;
 	uint32_t pixel, color = 0;
 
@@ -16,16 +18,33 @@ static inline void draw_sixel(struct framebuffer *fb, int line, int col, uint8_t
 	}
 }
 
+/* Based on OpenBSD:src/sys/arch/luna88k/dev/omrasops.c */
+
+#define	ALL1BITS	(~0U)
+#define	BLITWIDTH	(32)
+#define	ALIGNMASK	(0x1f)
+#define	BYTESDONE	(4)
+#define P0(addr) ((u_int32_t *)(addr))
+#define P1(addr) ((u_int32_t *)((u_int8_t *)(addr) +  0x40000))
+#define P2(addr) ((u_int32_t *)((u_int8_t *)(addr) +  0x80000))
+#define P3(addr) ((u_int32_t *)((u_int8_t *)(addr) +  0xC0000))
+#define P4(addr) ((u_int32_t *)((u_int8_t *)(addr) + 0x100000))
+#define P5(addr) ((u_int32_t *)((u_int8_t *)(addr) + 0x140000))
+#define P6(addr) ((u_int32_t *)((u_int8_t *)(addr) + 0x180000))
+#define P7(addr) ((u_int32_t *)((u_int8_t *)(addr) + 0x1C0000))
+
 static inline void draw_line(struct framebuffer *fb, struct terminal *term, int line)
 {
-	int pos, size, bdf_padding, glyph_width, margin_right;
-	int col, w, h;
-	uint32_t pixel;
+	int pos, size, col, glyph_width, bdf_shift;
 	struct color_pair_t color_pair;
 	struct cell_t *cellp;
 
+	u_int8_t *p;
+	int x, y, width, height, align;
+	u_int32_t lmask, rmask, glyph, glyphbg, fgpat, bgpat;
+	int fg, bg, plane;
+
 	for (col = term->cols - 1; col >= 0; col--) {
-		margin_right = (term->cols - 1 - col) * CELL_WIDTH;
 
 		/* target cell */
 		cellp = &term->cells[col + line * term->cols];
@@ -39,12 +58,19 @@ static inline void draw_line(struct framebuffer *fb, struct terminal *term, int
 		/* copy current color_pair (maybe changed) */
 		color_pair = cellp->color_pair;
 
+#if 0
 		/* check wide character or not */
 		glyph_width = (cellp->width == HALF) ? CELL_WIDTH: CELL_WIDTH * 2;
 		bdf_padding = my_ceil(glyph_width, BITS_PER_BYTE) * BITS_PER_BYTE - glyph_width;
 		if (cellp->width == WIDE)
 			bdf_padding += CELL_WIDTH;
+#endif
+		/* check wide character or not */
+		if (cellp->width == NEXT_TO_WIDE)
+			continue;
 
+		glyph_width = (cellp->width == HALF) ? CELL_WIDTH: CELL_WIDTH * 2;
+		bdf_shift = 32 - my_ceil(glyph_width, BITS_PER_BYTE) * BITS_PER_BYTE;
 		/* check cursor positon */
 		if ((term->mode & MODE_CURSOR && line == term->cursor.y)
 			&& (col == term->cursor.x
@@ -54,40 +80,159 @@ static inline void draw_line(struct framebuffer *fb, struct terminal *term, int
 			color_pair.bg = (!tty.visible && BACKGROUND_DRAW) ? PASSIVE_CURSOR_COLOR: ACTIVE_CURSOR_COLOR;
 		}
 
-		for (h = 0; h < CELL_HEIGHT; h++) {
-			/* if UNDERLINE attribute on, swap bg/fg */
-			if ((h == (CELL_HEIGHT - 1)) && (cellp->attribute & attr_mask[ATTR_UNDERLINE]))
-				color_pair.bg = color_pair.fg;
-
-			for (w = 0; w < CELL_WIDTH; w++) {
-				pos = (term->width - 1 - margin_right - w) * fb->bytes_per_pixel
-					+ (line * CELL_HEIGHT + h) * fb->line_length;
-
-				/* set color palette */
-				if (cellp->glyphp->bitmap[h] & (0x01 << (bdf_padding + w)))
-					pixel = term->color_palette[color_pair.fg];
-				else if (fb->wall && color_pair.bg == DEFAULT_BG) /* wallpaper */
-					memcpy(&pixel, fb->wall + pos, fb->bytes_per_pixel);
-				else
-					pixel = term->color_palette[color_pair.bg];
-
-				/* update copy buffer only */
-				memcpy(fb->buf + pos, &pixel, fb->bytes_per_pixel);
+		/* color palette */
+		fg = term->color_palette[color_pair.fg];
+		bg = term->color_palette[color_pair.bg];
+#if 0
+		if (fb->wall && color_pair.bg == DEFAULT_BG) /* wallpaper */
+			memcpy(&pixel, fb->wall + pos, fb->bytes_per_pixel);
+#endif
+
+		x = CELL_WIDTH * col;
+		y = CELL_HEIGHT * line;
+
+		p = (u_int8_t *)fb->buf + y * fb->line_length + ((x / 32) * 4);
+		align = x & ALIGNMASK;
+		width = cellp->glyphp->width * CELL_WIDTH + align;
+		lmask = ALL1BITS >> align;
+		rmask = ALL1BITS << (-width & ALIGNMASK);
+		height = 0;
+
+		if (width <= BLITWIDTH) {
+			lmask &= rmask;
+			while (height < CELL_HEIGHT) {
+				/* if UNDERLINE attribute on, swap bg/fg */
+				if ((height == (CELL_HEIGHT - 1))
+					&& (cellp->attribute & attr_mask[ATTR_UNDERLINE]))
+						bg = fg;
+				glyph = (uint32_t)cellp->glyphp->bitmap[height];
+				/* shift leftmost */
+				glyph <<= bdf_shift;
+
+				glyph = glyph >> align;
+				glyphbg = glyph ^ ALL1BITS;
+
+				fgpat = (fg & 0x01) ? glyph : 0;
+				bgpat = (bg & 0x01) ? glyphbg : 0;
+				*P0(p) = (*P0(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x02) ? glyph : 0;
+				bgpat = (bg & 0x02) ? glyphbg : 0;
+				*P1(p) = (*P1(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x04) ? glyph : 0;
+				bgpat = (bg & 0x04) ? glyphbg : 0;
+				*P2(p) = (*P2(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x08) ? glyph : 0;
+				bgpat = (bg & 0x08) ? glyphbg : 0;
+				*P3(p) = (*P3(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x10) ? glyph : 0;
+				bgpat = (bg & 0x10) ? glyphbg : 0;
+				*P4(p) = (*P4(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x20) ? glyph : 0;
+				bgpat = (bg & 0x20) ? glyphbg : 0;
+				*P5(p) = (*P5(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x40) ? glyph : 0;
+				bgpat = (bg & 0x40) ? glyphbg : 0;
+				*P6(p) = (*P6(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x80) ? glyph : 0;
+				bgpat = (bg & 0x80) ? glyphbg : 0;
+				*P7(p) = (*P7(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+
+				p += fb->line_length;
+				height++;
+			}
+		} else {
+			u_int8_t *q = p;
+			u_int32_t lhalf, rhalf;
+			u_int32_t lhalfbg, rhalfbg;
+
+			while (height < CELL_HEIGHT) {
+				/* if UNDERLINE attribute on, swap bg/fg */
+				if ((height == (CELL_HEIGHT - 1))
+					&& (cellp->attribute & attr_mask[ATTR_UNDERLINE]))
+						bg = fg;
+				glyph = (uint32_t)cellp->glyphp->bitmap[height];
+				/* shift leftmost */
+				glyph <<= bdf_shift;
+
+				lhalf = glyph >> align;
+				lhalfbg = lhalf ^ ALL1BITS;
+
+				fgpat = (fg & 0x01) ? lhalf : 0;
+				bgpat = (bg & 0x01) ? lhalfbg : 0;
+				*P0(p) = (*P0(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x02) ? lhalf : 0;
+				bgpat = (bg & 0x02) ? lhalfbg : 0;
+				*P1(p) = (*P1(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x04) ? lhalf : 0;
+				bgpat = (bg & 0x04) ? lhalfbg : 0;
+				*P2(p) = (*P2(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x08) ? lhalf : 0;
+				bgpat = (bg & 0x08) ? lhalfbg : 0;
+				*P3(p) = (*P3(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x10) ? lhalf : 0;
+				bgpat = (bg & 0x10) ? lhalfbg : 0;
+				*P4(p) = (*P4(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x20) ? lhalf : 0;
+				bgpat = (bg & 0x20) ? lhalfbg : 0;
+				*P5(p) = (*P5(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x40) ? lhalf : 0;
+				bgpat = (bg & 0x40) ? lhalfbg : 0;
+				*P6(p) = (*P6(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+				fgpat = (fg & 0x80) ? lhalf : 0;
+				bgpat = (bg & 0x80) ? lhalfbg : 0;
+				*P7(p) = (*P7(p) & ~lmask) | ((fgpat | bgpat) & lmask);
+
+				p += BYTESDONE;
+
+				rhalf = glyph << (BLITWIDTH - align);
+				rhalfbg = rhalf ^ ALL1BITS;
+
+				fgpat = (fg & 0x01) ? rhalf : 0;
+				bgpat = (bg & 0x01) ? rhalfbg : 0;
+				*P0(p) = ((fgpat | bgpat) & rmask) | (*P0(p) & ~rmask);
+				fgpat = (fg & 0x02) ? rhalf : 0;
+				bgpat = (bg & 0x02) ? rhalfbg : 0;
+				*P1(p) = ((fgpat | bgpat) & rmask) | (*P1(p) & ~rmask);
+				fgpat = (fg & 0x04) ? rhalf : 0;
+				bgpat = (bg & 0x04) ? rhalfbg : 0;
+				*P2(p) = ((fgpat | bgpat) & rmask) | (*P2(p) & ~rmask);
+				fgpat = (fg & 0x08) ? rhalf : 0;
+				bgpat = (bg & 0x08) ? rhalfbg : 0;
+				*P3(p) = ((fgpat | bgpat) & rmask) | (*P3(p) & ~rmask);
+				fgpat = (fg & 0x10) ? rhalf : 0;
+				bgpat = (bg & 0x10) ? rhalfbg : 0;
+				*P4(p) = ((fgpat | bgpat) & rmask) | (*P4(p) & ~rmask);
+				fgpat = (fg & 0x20) ? rhalf : 0;
+				bgpat = (bg & 0x20) ? rhalfbg : 0;
+				*P5(p) = ((fgpat | bgpat) & rmask) | (*P5(p) & ~rmask);
+				fgpat = (fg & 0x40) ? rhalf : 0;
+				bgpat = (bg & 0x40) ? rhalfbg : 0;
+				*P6(p) = ((fgpat | bgpat) & rmask) | (*P6(p) & ~rmask);
+				fgpat = (fg & 0x80) ? rhalf : 0;
+				bgpat = (bg & 0x80) ? rhalfbg : 0;
+				*P7(p) = ((fgpat | bgpat) & rmask) | (*P7(p) & ~rmask);
+
+				p = (q += fb->line_length);
+				height++;
 			}
 		}
 	}
 
 	/* actual display update (bit blit) */
+#if 0
 	pos = (line * CELL_HEIGHT) * fb->line_length;
 	size = CELL_HEIGHT * fb->line_length;
 	memcpy(fb->fp + pos, fb->buf + pos, size);
-
-	/* TODO: page flip
-		if fb_fix_screeninfo.ypanstep > 0, we can use hardware panning.
-		set fb_fix_screeninfo.{yres_virtual,yoffset} and call ioctl(FBIOPAN_DISPLAY)
-		but drivers  of recent hardware (inteldrmfb, nouveaufb, radeonfb) don't support...
-		(we can use this by using libdrm)
-	*/
+#else
+	size = TERM_WIDTH / 8;
+	for (height = 0; height < CELL_HEIGHT; height++) {
+		for (plane = 0; plane < DEPTH; plane++) {
+			pos = (line * CELL_HEIGHT + height) * fb->line_length
+				+ 0x40000 * plane;
+			memcpy(fb->fp + pos, fb->buf + pos, size);
+		}
+	}
+#endif
 
 	term->line_dirty[line] = ((term->mode & MODE_CURSOR) && term->cursor.y == line) ? true: false;
 }
diff --git a/fb/openbsd.h b/fb/openbsd.h
index 5a11596..276f017 100644
--- a/fb/openbsd.h
+++ b/fb/openbsd.h
@@ -13,11 +13,17 @@ typedef unsigned long   u_long;
 #include  <dev/wscons/wsconsio.h>
 #include  <dev/wscons/wsksymdef.h>
 
-/* some structs for NetBSD */
+/* some structs for OpenBSD */
 enum term_size {
+#if 0
 	TERM_WIDTH  = 640,
 	TERM_HEIGHT = 480,
 	DEPTH = 8,
+#else
+	TERM_WIDTH  = 1280,
+	TERM_HEIGHT = 1024,
+	DEPTH = 8,
+#endif
 };
 
 enum fbtype_t {
@@ -45,6 +51,7 @@ struct fbinfo_t {
 
 struct framebuffer {
 	uint8_t *fp;          /* pointer of framebuffer(read only) */
+	uint8_t *fp_orig;     /* pointer of framebuffer(original) */
 	uint8_t *wall;        /* buffer for wallpaper */
 	uint8_t *buf;         /* copy of framebuffer */
 	int fd;               /* file descriptor of framebuffer */
@@ -83,7 +90,7 @@ uint8_t *load_wallpaper(struct framebuffer *fb)
 	return ptr;
 }
 
-/* some functions for NetBSD framebuffer */
+/* some functions for OpenBSD framebuffer */
 void cmap_create(struct wsdisplay_cmap **cmap, int size)
 {
 	*cmap                = (struct wsdisplay_cmap *) ecalloc(1, sizeof(struct wsdisplay_cmap));
@@ -197,12 +204,19 @@ void fb_init(struct framebuffer *fb, uint32_t *color_palette)
 		goto fb_init_error;
 	}
 
+	/* XXX: Should be check if WSDISPLAYIO_TYPE_LUNA ? */
+
 	fb->width  = TERM_WIDTH;
 	fb->height = TERM_HEIGHT;
 	fb->bytes_per_pixel = my_ceil(DEPTH, BITS_PER_BYTE);
 
+#if 0
 	fb->line_length = fb->bytes_per_pixel * fb->width;
 	fb->screen_size = fb->height * fb->line_length;
+#else
+	fb->line_length = 2048 / 8;
+	fb->screen_size = fb->height * fb->line_length * DEPTH;
+#endif
 	fb->vinfo = bpp_table[DEPTH];
 
 	if (DEBUG)
@@ -214,7 +228,7 @@ void fb_init(struct framebuffer *fb, uint32_t *color_palette)
 		fb->cmap = fb->cmap_org = NULL;
 		fb->vinfo.fbtype = FBTYPE_RGB;
 	}
-	else if (DEPTH == 8) {
+	else if (DEPTH == 8 || DEPTH == 4 || DEPTH == 1 ) {
 		cmap_create(&fb->cmap, COLORS);
 		cmap_create(&fb->cmap_org, finfo.cmsize);
 		cmap_init(fb);
@@ -226,7 +240,10 @@ void fb_init(struct framebuffer *fb, uint32_t *color_palette)
 	for (i = 0; i < COLORS; i++) /* init color palette */
 		color_palette[i] = (fb->bytes_per_pixel == 1) ? (uint32_t) i: color2pixel(&fb->vinfo, color_list[i]);
 
-	fb->fp    = (uint8_t *) emmap(0, fb->screen_size, PROT_WRITE | PROT_READ, MAP_SHARED, fb->fd, 0);
+	fb->fp_orig = (uint8_t *) emmap(0, fb->screen_size, PROT_WRITE | PROT_READ, MAP_SHARED, fb->fd, 0);
+
+	fb->fp = fb->fp_orig + 8; /* XXX: LUNA quirk; need 8 byte offset */
+
 	fb->buf   = (uint8_t *) ecalloc(1, fb->screen_size);
 	//fb->wall  = (WALLPAPER && fb->bytes_per_pixel > 1) ? load_wallpaper(fb): NULL;
 
@@ -251,6 +268,6 @@ void fb_die(struct framebuffer *fb)
 	}
 	free(fb->buf);
 	free(fb->wall);
-	emunmap(fb->fp, fb->screen_size);
+	emunmap(fb->fp_orig, fb->screen_size);
 	eclose(fb->fd);
 }
diff --git a/makefile b/makefile
index caf405f..8fc5679 100644
--- a/makefile
+++ b/makefile
@@ -1,7 +1,8 @@
 CC ?= gcc
 #CC ?= clang
 
-CFLAGS  ?= -std=c99 -pedantic -Wall -Wextra -O3 -s -pipe
+#CFLAGS  ?= -std=c99 -pedantic -Wall -Wextra -O3 -s -pipe
+CFLAGS  = -std=c99 -pedantic -Wall -Wextra -O3 -s -pipe
 LDFLAGS ?=
 
 XCFLAGS  ?= -std=c99 -pedantic -Wall -Wextra -I/usr/include/X11/ -O3 -s -pipe
diff --git a/yaft.c b/yaft.c
index 9acbad6..cbeaf9f 100644
--- a/yaft.c
+++ b/yaft.c
@@ -77,11 +77,13 @@ void tty_init(struct termios *save_tm)
 	vtm.mode   = VT_PROCESS;
 	vtm.waitv  = 0;
 	vtm.relsig = vtm.acqsig = vtm.frsig = SIGUSR1;
+#if 0
 	if (ioctl(STDIN_FILENO, VT_SETMODE, &vtm))
 		fatal("ioctl: VT_SETMODE failed (maybe here is not console)");
 
 	if (ioctl(STDIN_FILENO, KDSETMODE, KD_GRAPHICS))
 		fatal("ioctl: KDSETMODE failed (maybe here is not console)");
+#endif
 
 	etcgetattr(STDIN_FILENO, save_tm);
 	set_rawmode(STDIN_FILENO, save_tm);
@@ -102,8 +104,10 @@ void tty_die(struct termios *save_tm)
 	vtm.mode   = VT_AUTO;
 	vtm.waitv  = 0;
 	vtm.relsig = vtm.acqsig = vtm.frsig = 0;
+#if 0
 	ioctl(STDIN_FILENO, VT_SETMODE, &vtm);
 	ioctl(STDIN_FILENO, KDSETMODE, KD_TEXT);
+#endif
 
 	tcsetattr(STDIN_FILENO, TCSAFLUSH, save_tm);
 	fflush(stdout);