From: John Tsiombikas <nuclear@member.fsf.org>
Date: Sat, 13 Jun 2020 10:59:00 +0000 (+0300)
Subject: backport builds
X-Git-Url: http://git.mutantstargoat.com/user/nuclear/?p=dosdemo;a=commitdiff_plain;h=b49854e0980a030cc8338a7f4ee779ecde1aae38

backport builds
---

diff --git a/GNUmakefile b/GNUmakefile
index 6e91ee8..49fb3a1 100644
--- a/GNUmakefile
+++ b/GNUmakefile
@@ -7,7 +7,7 @@ bin = demo
 asmsrc += cspr/dbgfont.asm cspr/confont.asm
 bindata = data/loading.img
 
-inc = -I/usr/local/include -Isrc -Isrc/scr -Isrc/sdl -Ilibs -Ilibs/imago/src -Ilibs/mikmod/include
+inc = -I/usr/local/include -Isrc -Isrc/3dgfx -Isrc/scr -Isrc/sdl -Ilibs -Ilibs/imago/src -Ilibs/mikmod/include
 warn = -pedantic -Wall -Wno-unused-variable -Wno-unused-function
 
 CFLAGS = $(arch) $(warn) -g $(inc) `sdl-config --cflags`
diff --git a/Makefile b/Makefile
index 8a2db49..0b6a460 100644
--- a/Makefile
+++ b/Makefile
@@ -1,19 +1,19 @@
 !ifdef __UNIX__
 dosobj = src/dos/audos.obj src/dos/djdpmi.obj src/dos/gfx.obj src/dos/keyb.obj &
 	src/dos/logger.obj src/dos/main.obj src/dos/sball.obj src/dos/timer.obj &
-	src/dos/vbe.obj src/dos/vga.obj src/dos/watdpmi.obj
+	src/dos/vbe.obj src/dos/vga.obj src/dos/watdpmi.obj src/dos/mouse.obj
 3dobj = src/3dgfx/3dgfx.obj src/3dgfx/mesh.obj src/3dgfx/meshload.obj &
 	src/3dgfx/polyclip.obj src/3dgfx/polyfill.obj
-srcobj = src/bsptree.obj src/cfgopt.obj src/console.obj &
-	src/demo.obj src/dynarr.obj src/gfxutil.obj &
-	src/metasurf.obj src/noise.obj &
+srcobj = src/bsptree.obj src/cfgopt.obj src/console.obj src/demo.obj &
+	src/dynarr.obj src/gfxutil.obj src/metasurf.obj src/noise.obj &
 	src/rbtree.obj src/screen.obj src/tinyfps.obj src/treestor.obj &
-	src/ts_text.obj src/util.obj
+	src/ts_text.obj src/util.obj src/data.obj
 scrobj = src/scr/bump.obj src/scr/fract.obj src/scr/greets.obj &
 	src/scr/grise.obj src/scr/hairball.obj src/scr/infcubes.obj &
 	src/scr/metaball.obj src/scr/plasma.obj src/scr/polytest.obj &
 	src/scr/smoketxt.obj src/scr/thunder.obj src/scr/tilemaze.obj &
 	src/scr/tunnel.obj
+csprobj = cspr/dbgfont.obj cspr/confont.obj
 
 incpath = -Isrc -Isrc/dos -Isrc/3dgfx -Ilibs -Ilibs/imago/src -Ilibs/anim/src &
 	-Ilibs/midas
@@ -22,26 +22,26 @@ libpath = libpath libs/imago libpath libs/anim libpath libs/midas
 
 dosobj = src\dos\audos.obj src\dos\djdpmi.obj src\dos\gfx.obj src\dos\keyb.obj &
 	src\dos\logger.obj src\dos\main.obj src\dos\sball.obj src\dos\timer.obj &
-	src\dos\vbe.obj src\dos\vga.obj src\dos\watdpmi.obj
+	src\dos\vbe.obj src\dos\vga.obj src\dos\watdpmi.obj src\dos\mouse.obj
 3dobj = src\3dgfx\3dgfx.obj src\3dgfx\mesh.obj src\3dgfx\meshload.obj &
 	src\3dgfx\polyclip.obj src\3dgfx\polyfill.obj
-srcobj = src\3dgfx.obj src\bsptree.obj src\cfgopt.obj src\console.obj &
-	src\demo.obj src\dynarr.obj src\gfxutil.obj src\mesh.obj src\meshload.obj &
-	src\metasurf.obj src\noise.obj src\polyclip.obj src\polyfill.obj &
+srcobj = src\bsptree.obj src\cfgopt.obj src\console.obj src\demo.obj &
+	src\dynarr.obj src\gfxutil.obj src\metasurf.obj src\noise.obj &
 	src\rbtree.obj src\screen.obj src\tinyfps.obj src\treestor.obj &
-	src\ts_text.obj src\util.obj
+	src\ts_text.obj src\util.obj src\data.obj
 scrobj = src\scr\bump.obj src\scr\fract.obj src\scr\greets.obj &
 	src\scr\grise.obj src\scr\hairball.obj src\scr\infcubes.obj &
 	src\scr\metaball.obj src\scr\plasma.obj src\scr\polytest.obj &
 	src\scr\smoketxt.obj src\scr\thunder.obj src\scr\tilemaze.obj &
 	src\scr\tunnel.obj
+csprobj = cspr\dbgfont.obj cspr\confont.obj
 
 incpath = -Isrc -Isrc\dos -Isrc\3dgfx -Ilibs -Ilibs\imago\src -Ilibs\anim\src &
 	-Ilibs\midas
 libpath = libpath libs\imago libpath libs\anim libpath libs\midas
 !endif
 
-obj = $(dosobj) $(3dobj) $(scrobj) $(srcobj)
+obj = $(dosobj) $(3dobj) $(scrobj) $(srcobj) $(csprobj)
 bin = demo.exe
 
 opt = -otexan
@@ -61,8 +61,8 @@ $(bin): cflags.occ $(obj) libs/imago/imago.lib libs/anim/anim.lib
 	%write ldflags.lnk $(LDFLAGS)
 	$(LD) debug all name $@ system dos4g file { @objects } @ldflags
 
-.c: src;src/dos;src/3dgfx;src/scr
-.asm: src;src/dos;src/3dgfx;src/scr
+.c: src;src/dos;src/3dgfx;src/scr;cspr
+.asm: src;src/dos;src/3dgfx;src/scr;cspr
 
 cflags.occ: Makefile
 	%write $@ $(CFLAGS)
diff --git a/libs/anim/Makefile b/libs/anim/Makefile
index 657e083..5da7a19 100644
--- a/libs/anim/Makefile
+++ b/libs/anim/Makefile
@@ -2,17 +2,19 @@ obj = anim.obj track.obj
 alib = anim.lib
 
 def = -dM_PI=3.141592653589793
-opt = -5 -fp5 -otexan -I.. -I../../src $(def)
+opt = -5 -fp5 -otexan
 dbg = -d1
 
 !ifdef __UNIX__
+inc =  -I.. -I../../src
 RM = rm -f
 !else
+inc =  -I.. -I..\..\src
 RM = del
 !endif
 
 CC = wcc386
-CFLAGS = $(dbg) $(opt) $(def) -zq -bt=dos
+CFLAGS = $(dbg) $(opt) $(inc) $(def) -zq -bt=dos
 
 $(alib): $(obj)
 	wlib -b -n $@ $(obj)
diff --git a/libs/anim/src/anim.c b/libs/anim/src/anim.c
index a2f1944..6468d31 100644
--- a/libs/anim/src/anim.c
+++ b/libs/anim/src/anim.c
@@ -617,11 +617,8 @@ void anm_set_rotation4f(struct anm_node *node, float x, float y, float z, float
 void anm_set_rotation_axis(struct anm_node *node, float angle, float x, float y, float z, anm_time_t tm)
 {
 	cgm_quat q;
-	cgm_vec3 axis;
-
-	cgm_vcons(&axis, x, y, z);
-	cgm_qrotation(&q, &axis, angle);
 
+	cgm_qrotation(&q, angle, x, y, z);
 	anm_set_rotation(node, (float*)&q, tm);
 }
 
diff --git a/src/data.asm b/src/data.asm
index 26234fc..e9cd436 100644
--- a/src/data.asm
+++ b/src/data.asm
@@ -1,4 +1,4 @@
-	section .data
+	section .data USE32
 
 	global loading_pixels
 	global _loading_pixels
diff --git a/src/dos/audos.c b/src/dos/audos.c
index 513bf89..6b7709a 100644
--- a/src/dos/audos.c
+++ b/src/dos/audos.c
@@ -4,6 +4,7 @@
 #include <ctype.h>
 #include "audio.h"
 #include "midasdll.h"
+#include "util.h"
 
 #define SET_MUS_VOL(vol) \
 	do { \
@@ -209,3 +210,13 @@ unsigned long get_msec(void)
 {
 	return ticks * tick_interval;
 }
+
+void sleep_msec(unsigned long msec)
+{
+	unsigned long wakeup_time = ticks + msec / tick_interval;
+	while(ticks < wakeup_time) {
+#ifdef USE_HLT
+		halt();
+#endif
+	}
+}
diff --git a/src/dos/main.c b/src/dos/main.c
index 0de90ed..b415bb4 100644
--- a/src/dos/main.c
+++ b/src/dos/main.c
@@ -69,7 +69,7 @@ int main(int argc, char **argv)
 		}
 	}
 
-	if(init(argc, argv) == -1) {
+	if(demo_init(argc, argv) == -1) {
 		status = -1;
 		goto break_evloop;
 	}
@@ -83,7 +83,7 @@ int main(int argc, char **argv)
 	for(;;) {
 		int key;
 		while((key = kb_getkey()) != -1) {
-			demo_key(key, 1);
+			demo_keyboard(key, 1);
 			if(quit) goto break_evloop;
 		}
 
@@ -99,11 +99,11 @@ int main(int argc, char **argv)
 		}
 
 		time_msec = get_msec();
-		draw();
+		demo_draw();
 	}
 
 break_evloop:
-	cleanup();
+	demo_cleanup();
 	set_text_mode();
 	cleanup_video();
 	kb_shutdown();
diff --git a/src/gfxutil.c b/src/gfxutil.c
index 259b56d..d410766 100644
--- a/src/gfxutil.c
+++ b/src/gfxutil.c
@@ -1,6 +1,6 @@
 #include <string.h>
-#include "gfxutil.h"
 #include "demo.h"
+#include "gfxutil.h"
 
 enum {
 	IN		= 0,
@@ -147,34 +147,58 @@ void draw_line(int x0, int y0, int x1, int y1, unsigned short color)
 
 #define BLUR(w, h, pstep, sstep) \
 	for(i=0; i<h; i++) { \
-		int sum = sptr[0] * (rad + 1); \
+		int r, g, b; \
+		int rsum = UNPACK_R16(sptr[0]) * (rad + 1); \
+		int gsum = UNPACK_G16(sptr[0]) * (rad + 1); \
+		int bsum = UNPACK_B16(sptr[0]) * (rad + 1); \
 		int count = (rad * 2 + 1) << 8; \
 		int midsize = w - rad * 2; \
-		int firstpix = sptr[0]; \
-		int lastpix = sptr[pstep * (w - 1)]; \
+		int rfirstpix = UNPACK_R16(sptr[0]); \
+		int rlastpix = UNPACK_R16(sptr[pstep * (w - 1)]); \
+		int gfirstpix = UNPACK_G16(sptr[0]); \
+		int glastpix = UNPACK_G16(sptr[pstep * (w - 1)]); \
+		int bfirstpix = UNPACK_B16(sptr[0]); \
+		int blastpix = UNPACK_B16(sptr[pstep * (w - 1)]); \
 		/* add up the contributions for the -1 pixel */ \
 		for(j=0; j<rad; j++) { \
-			sum += sptr[pstep * j]; \
+			rsum += UNPACK_R16(sptr[pstep * j]); \
+			gsum += UNPACK_G16(sptr[pstep * j]); \
+			bsum += UNPACK_B16(sptr[pstep * j]); \
 		} \
 		/* first part adding sptr[rad] and subtracting sptr[0] */ \
 		for(j=0; j<=rad; j++) { \
-			sum += (int)sptr[pstep * rad] - firstpix; \
+			rsum += UNPACK_R16((int)sptr[pstep * rad]) - rfirstpix; \
+			gsum += UNPACK_G16((int)sptr[pstep * rad]) - gfirstpix; \
+			bsum += UNPACK_B16((int)sptr[pstep * rad]) - bfirstpix; \
 			sptr += pstep; \
-			*dptr = scale * sum / count; \
+			r = scale * rsum / count; \
+			g = scale * gsum / count; \
+			b = scale * bsum / count; \
+			*dptr = PACK_RGB16(r, g, b); \
 			dptr += pstep; \
 		} \
 		/* middle part adding sptr[rad] and subtracting sptr[-(rad+1)] */ \
 		for(j=1; j<midsize; j++) { \
-			sum += (int)sptr[pstep * rad] - (int)sptr[-(rad + 1) * pstep]; \
+			rsum += UNPACK_R16((int)sptr[pstep * rad]) - UNPACK_R16((int)sptr[-(rad + 1) * pstep]); \
+			gsum += UNPACK_G16((int)sptr[pstep * rad]) - UNPACK_G16((int)sptr[-(rad + 1) * pstep]); \
+			bsum += UNPACK_B16((int)sptr[pstep * rad]) - UNPACK_B16((int)sptr[-(rad + 1) * pstep]); \
 			sptr += pstep; \
-			*dptr = scale * sum / count; \
+			r = scale * rsum / count; \
+			g = scale * gsum / count; \
+			b = scale * bsum / count; \
+			*dptr = PACK_RGB16(r, g, b); \
 			dptr += pstep; \
 		} \
 		/* last part adding lastpix and subtracting sptr[-(rad+1)] */ \
 		for(j=0; j<rad; j++) { \
-			sum += lastpix - (int)sptr[-(rad + 1) * pstep]; \
+			rsum += rlastpix - UNPACK_R16((int)sptr[-(rad + 1) * pstep]); \
+			gsum += glastpix - UNPACK_G16((int)sptr[-(rad + 1) * pstep]); \
+			bsum += blastpix - UNPACK_B16((int)sptr[-(rad + 1) * pstep]); \
 			sptr += pstep; \
-			*dptr = scale * sum / count; \
+			r = scale * rsum / count; \
+			g = scale * gsum / count; \
+			b = scale * bsum / count; \
+			*dptr = PACK_RGB16(r, g, b); \
 			dptr += pstep; \
 		} \
 		sptr += sstep; \
@@ -183,23 +207,23 @@ void draw_line(int x0, int y0, int x1, int y1, unsigned short color)
 
 /* TODO bound blur rad to image size to avoid inner loop conditionals */
 /* TODO make version with pow2 (rad*2+1) to avoid div with count everywhere */
-void blur_grey_horiz(uint16_t *dest, uint16_t *src, int xsz, int ysz, int rad, int scale)
+void blur_horiz(uint16_t *dest, uint16_t *src, int xsz, int ysz, int rad, int scale)
 {
 	int i, j;
-	unsigned char *dptr = (unsigned char*)dest;
-	unsigned char *sptr = (unsigned char*)src;
+	uint16_t *dptr = dest;
+	uint16_t *sptr = src;
 
-	BLUR(xsz, ysz, 2, 0);
+	BLUR(xsz, ysz, 1, 0);
 }
 
 
-void blur_grey_vert(uint16_t *dest, uint16_t *src, int xsz, int ysz, int rad, int scale)
+void blur_vert(uint16_t *dest, uint16_t *src, int xsz, int ysz, int rad, int scale)
 {
 	int i, j;
-	unsigned char *dptr = (unsigned char*)dest;
-	unsigned char *sptr = (unsigned char*)src;
-	int pixel_step = xsz * 2;
-	int scanline_step = 2 - ysz * pixel_step;
+	uint16_t *dptr = dest;
+	uint16_t *sptr = src;
+	int pixel_step = xsz;
+	int scanline_step = 1 - ysz * pixel_step;
 
 	BLUR(ysz, xsz, pixel_step, scanline_step);
 }
@@ -221,15 +245,26 @@ void blitfb(uint16_t *dest, uint16_t *src, int width, int height, int pitch_pix)
 {
 	int i;
 	for(i=0; i<height; i++) {
-		memcpy(dest, src, width * 2);
+		memcpy(dest, src, width << 1);
 		dest += 320;
+	}
+}
+
+void blit(uint16_t *dest, int destwidth, uint16_t *src, int width, int height, int pitch_pix)
+{
+	int i, spansz = width << 1;
+	for(i=0; i<height; i++) {
+		memcpy(dest, src, spansz);
+		dest += destwidth;
 		src += pitch_pix;
 	}
 }
 
-void blitfb_key(uint16_t *dest, uint16_t *src, int width, int height, int pitch_pix, uint16_t key)
+void blit_key(uint16_t *dest, int destwidth, uint16_t *src, int width, int height, int pitch_pix, uint16_t key)
 {
-	int i, j, dadv = 320 - width;
+	int i, j;
+	int dadv = destwidth - width;
+	int sadv = pitch_pix - width;
 
 	for(i=0; i<height; i++) {
 		for(j=0; j<width; j++) {
@@ -238,6 +273,7 @@ void blitfb_key(uint16_t *dest, uint16_t *src, int width, int height, int pitch_
 			dest++;
 		}
 		dest += dadv;
+		src += sadv;
 	}
 
 }
diff --git a/src/gfxutil.h b/src/gfxutil.h
index b46f813..d4f06c6 100644
--- a/src/gfxutil.h
+++ b/src/gfxutil.h
@@ -3,6 +3,7 @@
 
 #include "inttypes.h"
 
+
 #define PACK_RGB16(r, g, b) \
 	((((uint16_t)(r) << 8) & 0xf800) | \
 	 (((uint16_t)(g) << 3) & 0x7e0) | \
@@ -12,23 +13,37 @@
 #define UNPACK_G16(c)	(((c) >> 3) & 0xfc)
 #define UNPACK_B16(c)	(((c) << 3) & 0xf8)
 
+
+#ifdef BUILD_BIGENDIAN
+
+#define PACK_RGB32(r, g, b) \
+	((((r) & 0xff) << 8) | (((g) & 0xff) << 16) | (((b) & 0xff) << 24) | 0xff)
+
+#define UNPACK_R32(c)	(((c) >> 8) & 0xff)
+#define UNPACK_G32(c)	(((c) >> 16) & 0xff)
+#define UNPACK_B32(c)	(((c) >> 24) & 0xff)
+
+#else	/* LITTLE_ENDIAN */
+
 #define PACK_RGB32(r, g, b) \
-	((((r) & 0xff) << 16) | (((g) & 0xff) << 8) | ((b) & 0xff))
+	((((r) & 0xff) << 16) | (((g) & 0xff) << 8) | ((b) & 0xff) | 0xff000000)
 
 #define UNPACK_R32(c)	(((c) >> 16) & 0xff)
 #define UNPACK_G32(c)	(((c) >> 8) & 0xff)
 #define UNPACK_B32(c)	((c) & 0xff)
+#endif
 
 int clip_line(int *x0, int *y0, int *x1, int *y1, int xmin, int ymin, int xmax, int ymax);
 void draw_line(int x0, int y0, int x1, int y1, unsigned short color);
 
 /* scale in 24.8 fixed point */
-void blur_grey_horiz(uint16_t *dest, uint16_t *src, int xsz, int ysz, int radius, int scale);
-void blur_grey_vert(uint16_t *dest, uint16_t *src, int xsz, int ysz, int radius, int scale);
+void blur_horiz(uint16_t *dest, uint16_t *src, int xsz, int ysz, int radius, int scale);
+void blur_vert(uint16_t *dest, uint16_t *src, int xsz, int ysz, int radius, int scale);
 
 void convimg_rgb24_rgb16(uint16_t *dest, unsigned char *src, int xsz, int ysz);
 
-void blitfb(uint16_t *dest, uint16_t *src, int xsz, int ysz, int pitch_pix);
-void blitfb_key(uint16_t *dest, uint16_t *src, int xsz, int ysz, int pitch_pix, uint16_t key);
+void blitfb(uint16_t *dest, uint16_t *src, int width, int height, int pitch_pix);
+void blit(uint16_t *dest, int destwidth, uint16_t *src, int xsz, int ysz, int pitch_pix);
+void blit_key(uint16_t *dest, int destwidth, uint16_t *src, int xsz, int ysz, int pitch_pix, uint16_t key);
 
 #endif	/* GFXUTIL_H_ */
diff --git a/src/scr/greets.c b/src/scr/greets.c
index 49a5b50..9a49527 100644
--- a/src/scr/greets.c
+++ b/src/scr/greets.c
@@ -154,22 +154,15 @@ static void draw(void)
 	draw_smktxt(stx);
 	g3d_framebuffer(FB_WIDTH, FB_HEIGHT, fb_pixels);
 
-	dest = fb_pixels;
-	src = cur_smokebuf;
-	for(i=0; i<FB_HEIGHT; i++) {
-		for(j=0; j<FB_WIDTH; j++) {
-			unsigned int alpha = *src++;
-			*dest++ = PACK_RGB16(alpha, alpha, alpha);
-		}
-	}
+	memcpy(fb_pixels, cur_smokebuf, smokebuf_size);
 
 	/*perf_start();*/
-	blur_grey_horiz(prev_smokebuf, cur_smokebuf, FB_WIDTH, FB_HEIGHT, BLUR_RAD, 240);
+	blur_horiz(prev_smokebuf, cur_smokebuf, FB_WIDTH, FB_HEIGHT, BLUR_RAD, 240);
 	/*
 	perf_end();
 	printf("blur perf: %lu\n", (unsigned long)perf_interval_count);
 	*/
-	blur_grey_vert(cur_smokebuf, prev_smokebuf, FB_WIDTH, FB_HEIGHT, BLUR_RAD, 240);
+	blur_vert(cur_smokebuf, prev_smokebuf, FB_WIDTH, FB_HEIGHT, BLUR_RAD, 240);
 	swap_smoke_buffers();
 
 	msec = get_msec();
diff --git a/src/screen.c b/src/screen.c
index 5167d96..8069ce8 100644
--- a/src/screen.c
+++ b/src/screen.c
@@ -204,7 +204,7 @@ void start_loadscr(void)
 void end_loadscr(void)
 {
 	blitfb(loading_pixels + SPLAT_Y * 320 + SPLAT_X, loading_pixels + 320 * 240, 32, 72, 32);
-	blitfb_key(loading_pixels + FING_Y * 320 + FING_LAST_X, loading_pixels + 247 * 320 + 64,
+	blit_key(loading_pixels + FING_Y * 320 + FING_LAST_X, 320, loading_pixels + 247 * 320 + 64,
 			FING_W, FING_H, FING_W, 0);
 	swap_buffers(loading_pixels);
 	sleep_msec(300);
@@ -220,7 +220,7 @@ void loadscr(int n, int count)
 	dptr = loading_pixels + FING_Y * 320 + FING_X + prev_xoffs;
 
 	while(prev_xoffs < xoffs) {
-		blitfb_key(dptr, sptr, FING_W, FING_H, FING_W, 0);
+		blit_key(dptr, 320, sptr, FING_W, FING_H, FING_W, 0);
 		dptr++;
 		prev_xoffs++;
 	}
diff --git a/src/util.h b/src/util.h
index b305d78..8a46036 100644
--- a/src/util.h
+++ b/src/util.h
@@ -16,6 +16,14 @@
 #define PACKED
 #endif
 
+#define BSWAP16(x)	((((x) >> 8) & 0xff) | (((x) & 0xff) << 8))
+#define BSWAP32(x)	\
+	((((x) >> 24) & 0xff) | \
+	 (((x) >> 8) & 0xff00) | \
+	 (((x) << 8) & 0xff0000) | \
+	 ((x) << 24))
+
+#if defined(__i386__) || defined(__x86_64__) || defined(__386__) || defined(MSDOS)
 /* fast conversion of double -> 32bit int
  * for details see:
  *  - http://chrishecker.com/images/f/fb/Gdmfp.pdf
@@ -26,10 +34,56 @@ static INLINE int32_t cround64(double val)
 	val += 6755399441055744.0;
 	return *(int32_t*)&val;
 }
+#else
+#define cround64(x)	((int32_t)(x))
+#endif
+
+static INLINE float rsqrt(float x)
+{
+	float xhalf = x * 0.5f;
+	int32_t i = *(int32_t*)&x;
+	i = 0x5f3759df - (i >> 1);
+	x = *(float*)&i;
+	x = x * (1.5f - xhalf * x * x);
+	return x;
+}
 
 extern uint32_t perf_start_count, perf_interval_count;
 
 #ifdef __WATCOMC__
+void memset16(void *dest, uint16_t val, int count);
+#pragma aux memset16 = \
+	"cld" \
+	"test ecx, 1" \
+	"jz memset16_dwords" \
+	"rep stosw" \
+	"jmp memset16_done" \
+	"memset16_dwords:" \
+	"shr ecx, 1" \
+	"push ax" \
+	"shl eax, 16" \
+	"pop ax" \
+	"rep stosd" \
+	"memset16_done:" \
+	parm[edi][ax][ecx];
+
+#ifdef USE_MMX
+void memcpy64(void *dest, void *src, int count);
+#pragma aux memcpy64 = \
+	"cploop:" \
+	"movq mm0, [edx]" \
+	"movq [ebx], mm0" \
+	"add edx, 8" \
+	"add ebx, 8" \
+	"dec ecx" \
+	"jnz cploop" \
+	"emms" \
+	parm[ebx][edx][ecx] \
+	modify[8087];
+#else
+#define memcpy64(dest, src, count)	memcpy(dest, src, (count) << 3)
+#endif
+
 void perf_start(void);
 #pragma aux perf_start = \
 	"xor eax, eax" \
@@ -52,14 +106,49 @@ void debug_break(void);
 
 void halt(void);
 #pragma aux halt = "hlt";
-
-void memset16(void *ptr, int val, int count);
-#pragma aux memset16 = \
-	"rep stosw" \
-	parm[edi][eax][ecx];
 #endif
 
 #ifdef __GNUC__
+#if defined(__i386__) || defined(__x86_64__)
+#define memset16(dest, val, count) asm volatile ( \
+	"cld\n\t" \
+	"test $1, %2\n\t" \
+	"jz 0f\n\t" \
+	"rep stosw\n\t" \
+	"jmp 1f\n\t" \
+	"0:\n\t" \
+	"shr $1, %2\n\t" \
+	"push %%ax\n\t" \
+	"shl $16, %%eax\n\t" \
+	"pop %%ax\n\t" \
+	"rep stosl\n\t" \
+	"1:\n\t"\
+	:: "D"(dest), "a"((uint16_t)(val)), "c"(count) \
+	: "memory")
+#else
+static void INLINE memset16(void *dest, uint16_t val, int count)
+{
+	uint16_t *ptr = dest;
+	while(count--) *ptr++ = val;
+}
+#endif
+
+#ifdef USE_MMX
+#define memcpy64(dest, src, count) asm volatile ( \
+	"0:\n\t" \
+	"movq (%1), %%mm0\n\t" \
+	"movq %%mm0, (%0)\n\t" \
+	"add $8, %1\n\t" \
+	"add $8, %0\n\t" \
+	"dec %2\n\t" \
+	"jnz 0b\n\t" \
+	"emms\n\t" \
+	:: "r"(dest), "r"(src), "r"(count) \
+	: "%mm0")
+#else
+#define memcpy64(dest, src, count)	memcpy(dest, src, (count) << 3)
+#endif
+
 #define perf_start()  asm volatile ( \
 	"xor %%eax, %%eax\n" \
 	"cpuid\n" \
@@ -83,13 +172,30 @@ void memset16(void *ptr, int val, int count);
 
 #define halt() \
 	asm volatile("hlt")
-
-#define memset16(ptr, val, count) asm volatile ( \
-	"rep stosw\n\t" \
-	:: "D"(ptr), "a"(val), "c"(count))
 #endif
 
 #ifdef _MSC_VER
+void __inline memset16(void *dest, uint16_t val, int count)
+{
+	__asm {
+		cld
+		mov ax, val
+		mov edi, dest
+		mov ecx, count
+		test ecx, 1
+		jz memset16_dwords
+		rep stosw
+		jmp memset16_done
+		memset16_dwords:
+		shr ecx, 1
+		push ax
+		shl eax, 16
+		pop ax
+		rep stosd
+		memset16_done:
+	}
+}
+
 #define perf_start() \
 	do { \
 		__asm { \
@@ -115,16 +221,6 @@ void memset16(void *ptr, int val, int count);
 	do { \
 		__asm { int 3 } \
 	} while(0)
-
-#define memset16(ptr, val, count) \
-	do { \
-		__asm { \
-			mov edi, ptr \
-			mov ecx, count \
-			mov eax, val \
-			rep stosw \
-		} \
-	} while(0)
 #endif
 
 #endif	/* UTIL_H_ */
diff --git a/tools/csprite/src/main.c b/tools/csprite/src/main.c
index fab302b..039a0eb 100644
--- a/tools/csprite/src/main.c
+++ b/tools/csprite/src/main.c
@@ -147,6 +147,7 @@ const char *prefixfmt[] = {
 	"\tsection .text\n"
 	"\tglobal %s\n"
 	"\tglobal _%s\n"
+	"\tglobal %s_\n"
 	"%s:\n"
 	"_%s:\n"
 	"\tmov eax, [esp + 12]\n"
@@ -162,6 +163,17 @@ const char *prefixfmt[] = {
 	"\tmov edx, eax\n"
 	"\tmov eax, [esp + 16]\n"
 	"\tjmp [titletab + eax * 4]\n\n"
+	"%s_:\n"
+	"\tpush eax\n"
+	"\tmov eax, ebx\n"
+	"\tshl eax, 9\n"
+	"\tshl ebx, 7\n"
+	"\tadd eax, ebx\n"
+	"\tshl edx, 1\n"
+	"\tadd eax, edx\n"
+	"\tpop edx\n"
+	"\tadd edx, eax\n"
+	"\tjmp [titletab + ecx * 4]\n\n"
 	"titletab:\n"
 };
 
@@ -217,7 +229,7 @@ int proc_sheet(const char *fname)
 		ysz = tile_ysz;
 	}
 
-	printf(prefixfmt[asyntax], name, name, name, name, fbpitch);
+	printf(prefixfmt[asyntax], name, name, name, name, name, fbpitch, name);
 	for(i=0; i<num_ytiles*num_xtiles; i++) {
 		if(asyntax == AS_GNU) {
 			printf("\t.long tile%d\n", i);