From: notaz <notaz@pixelinis>
Date: Sat, 31 Jul 2010 16:12:26 +0000 (+0300)
Subject: wiz blitters, fb restore, tweaks
X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=6ca083930098ee075c8f61cf2c04d616349959c3;p=ginge.git

wiz blitters, fb restore, tweaks
---

diff --git a/common/host_fb.c b/common/host_fb.c
index 46467d4..94910ec 100644
--- a/common/host_fb.c
+++ b/common/host_fb.c
@@ -14,6 +14,7 @@ static int host_stride;
 #include "fbdev.c"
 
 static struct vout_fbdev *fbdev;
+static unsigned short host_pal[256];
 
 void *host_video_flip(void)
 {
@@ -42,39 +43,13 @@ int host_video_init(int *stride, int no_dblbuf)
   return 0;
 }
 
-#elif defined(WIZ)
-
-#include "warm.c"
-#include "wiz_video.c"
-
-void *host_video_flip(void)
+void host_video_finish(void)
 {
-  vout_gp2x_flip();
-  host_screen = g_screen_ptr;
-  return host_screen;
+  vout_fbdev_finish(fbdev);
+  fbdev = NULL;
 }
 
-int host_video_init(int *stride, int no_dblbuf)
-{
-  int ret;
-
-  host_stride = 320 * 2;
-  if (stride != 0)
-    *stride = host_stride;
-
-  ret = vout_gp2x_init(no_dblbuf);
-  if (ret != 0)
-    return ret;
-
-  host_video_flip();
-  return 0;
-}
-
-#endif
-
-static unsigned short host_pal[256];
-
-static void host_update_pal(unsigned int *pal)
+void host_video_update_pal(unsigned int *pal)
 {
   unsigned short *dstp = host_pal;
   int i;
@@ -85,15 +60,16 @@ static void host_update_pal(unsigned int *pal)
   }
 }
 
-void host_video_blit4(const unsigned char *src, int w, int h, unsigned int *pal)
+void host_video_change_bpp(int bpp)
+{
+}
+
+void host_video_blit4(const unsigned char *src, int w, int h)
 {
   unsigned short *dst = host_screen;
   unsigned short *hpal = host_pal;
   int i, u;
 
-  if (pal != NULL)
-    host_update_pal(pal);
-
   for (i = 0; i < 240; i++, dst += host_stride / 2 - 320) {
     for (u = 320 / 2; u > 0; u--, src++) {
       *dst++ = hpal[*src >> 4];
@@ -104,15 +80,12 @@ void host_video_blit4(const unsigned char *src, int w, int h, unsigned int *pal)
   host_video_flip();
 }
 
-void host_video_blit8(const unsigned char *src, int w, int h, unsigned int *pal)
+void host_video_blit8(const unsigned char *src, int w, int h)
 {
   unsigned short *dst = host_screen;
   unsigned short *hpal = host_pal;
   int i, u;
 
-  if (pal != NULL)
-    host_update_pal(pal);
-
   for (i = 0; i < 240; i++, dst += host_stride / 2 - 320) {
     for (u = 320 / 4; u > 0; u--) {
       *dst++ = hpal[*src++];
@@ -136,3 +109,73 @@ void host_video_blit16(const unsigned short *src, int w, int h)
   host_video_flip();
 }
 
+#elif defined(WIZ)
+
+#include "warm.c"
+#include "wiz_video.c"
+
+void *host_video_flip(void)
+{
+  vout_gp2x_flip();
+  host_screen = g_screen_ptr;
+  return host_screen;
+}
+
+int host_video_init(int *stride, int no_dblbuf)
+{
+  int ret;
+
+  host_stride = 320 * 2;
+  if (stride != 0)
+    *stride = host_stride;
+
+  ret = vout_gp2x_init(no_dblbuf);
+  if (ret != 0)
+    return ret;
+
+  vout_gp2x_set_mode(16, !no_dblbuf);
+  host_video_flip();
+  return 0;
+}
+
+void host_video_finish(void)
+{
+  vout_gp2x_finish();
+}
+
+void host_video_update_pal(unsigned int *pal)
+{
+  vout_gp2x_set_palette(pal, 256);
+}
+
+void host_video_change_bpp(int bpp)
+{
+  vout_gp2x_set_mode(bpp, 1);
+}
+
+#ifdef LOADER
+void host_video_blit4(const unsigned char *src, int w, int h)
+{
+  memcpy(host_screen, src, 320*240/2); // FIXME
+  host_video_flip();
+}
+
+void host_video_blit8(const unsigned char *src, int w, int h)
+{
+  extern void rotated_blit8(void *dst, const void *linesx4);
+
+  rotated_blit8(host_screen, src);
+  host_video_flip();
+}
+
+void host_video_blit16(const unsigned short *src, int w, int h)
+{
+  extern void rotated_blit16(void *dst, const void *linesx4);
+
+  rotated_blit16(host_screen, src);
+  host_video_flip();
+}
+#endif // LOADER
+
+#endif // WIZ
+
diff --git a/common/host_fb.h b/common/host_fb.h
index edb931b..1fcfdc3 100644
--- a/common/host_fb.h
+++ b/common/host_fb.h
@@ -1,6 +1,9 @@
 int   host_video_init(int *stride, int no_dblbuf);
+void  host_video_finish(void);
 void *host_video_flip(void);
+void  host_video_change_bpp(int bpp);
+void  host_video_update_pal(unsigned int *pal);
 
-void host_video_blit4(const unsigned char *src, int w, int h, unsigned int *pal);
-void host_video_blit8(const unsigned char *src, int w, int h, unsigned int *pal);
+void host_video_blit4(const unsigned char *src, int w, int h);
+void host_video_blit8(const unsigned char *src, int w, int h);
 void host_video_blit16(const unsigned short *src, int w, int h);
diff --git a/common/wiz_video.c b/common/wiz_video.c
index 7996ee3..999fc06 100644
--- a/common/wiz_video.c
+++ b/common/wiz_video.c
@@ -100,6 +100,59 @@ static int vout_gp2x_init(int no_dblbuf)
 	return 0;
 }
 
+static void vout_gp2x_set_mode(int bpp, int rot)
+{
+	int rot_cmd[2] = { 0, 0 };
+	int code = 0, bytes = 2;
+	unsigned int r;
+	int ret;
+
+	rot_cmd[0] = rot ? 6 : 5;
+	ret = ioctl(fbdev, _IOW('D', 90, int[2]), rot_cmd);
+	if (ret < 0)
+		perror("rot ioctl failed");
+
+	memregl[0x4004>>2] = rot ? 0x013f00ef : 0x00ef013f;
+	memregl[0x4000>>2] |= 1 << 3;
+
+	switch (bpp)
+	{
+		case 8:
+			code = 0x443a;
+			bytes = 1;
+			break;
+
+		case 15:
+		case 16:
+			code = 0x4432;
+			bytes = 2;
+			break;
+
+		default:
+			fprintf(stderr, "unhandled bpp request: %d\n", abs(bpp));
+			return;
+	}
+
+	memregl[0x405c>>2] = bytes;
+	memregl[0x4060>>2] = bytes * (rot ? 240 : 320);
+
+	r = memregl[0x4058>>2];
+	r = (r & 0xffff) | (code << 16) | 0x10;
+	memregl[0x4058>>2] = r;
+}
+
+static void vout_gp2x_set_palette(unsigned int *pal, int len)
+{
+	/* pollux palette is 16bpp only.. */
+	int i;
+	for (i = 0; i < len; i++)
+	{
+		int c = pal[i];
+		c = ((c >> 8) & 0xf800) | ((c >> 5) & 0x07c0) | ((c >> 3) & 0x001f);
+		memregl[0x4070>>2] = (i << 24) | c;
+	}
+}
+
 void vout_gp2x_finish(void)
 {
 	memregl[0x406C>>2] = fb_paddr[0];
diff --git a/common/wiz_video_arm.s b/common/wiz_video_arm.s
new file mode 100644
index 0000000..d538ad1
--- /dev/null
+++ b/common/wiz_video_arm.s
@@ -0,0 +1,139 @@
+@ vim:filetype=armasm
+
+@ input: r2-r5
+@ output: r7,r8
+@ trash: r6
+.macro rb_line_low
+    mov     r6, r2, lsl #16
+    mov     r7, r3, lsl #16
+    orr     r7, r7, r6, lsr #16
+    mov     r6, r4, lsl #16
+    mov     r8, r5, lsl #16
+    orr     r8, r8, r6, lsr #16
+.endm
+
+.macro rb_line_hi
+    mov     r6, r2, lsr #16
+    mov     r7, r3, lsr #16
+    orr     r7, r6, r7, lsl #16
+    mov     r6, r4, lsr #16
+    mov     r8, r5, lsr #16
+    orr     r8, r6, r8, lsl #16
+.endm
+
+.global rotated_blit16 @ void *dst, void *linesx4
+rotated_blit16:
+    stmfd   sp!,{r4-r8,lr}
+
+    sub     r0, r0, #240*2 @ adjust
+    mov     lr, #240/4
+
+rotated_blit_loop16_o:
+    orr     lr, lr, #((320/4)-1) << 16
+    add     r0, r0, #(240*320)*2
+
+rotated_blit_loop16:
+    ldr     r2, [r1, #320*0*2]
+    ldr     r3, [r1, #320*1*2]
+    ldr     r4, [r1, #320*2*2]
+    ldr     r5, [r1, #320*3*2]
+    rb_line_low
+    stmia   r0, {r7,r8}
+    sub     r0, r0, #240*2
+    rb_line_hi
+    stmia   r0, {r7,r8}
+    sub     r0, r0, #240*2
+
+    ldr     r2, [r1, #320*0*2+4]
+    ldr     r3, [r1, #320*1*2+4]
+    ldr     r4, [r1, #320*2*2+4]
+    ldr     r5, [r1, #320*3*2+4]
+    rb_line_low
+    stmia   r0, {r7,r8}
+    sub     r0, r0, #240*2
+    rb_line_hi
+    stmia   r0, {r7,r8}
+    sub     r0, r0, #240*2
+
+    subs    lr, lr, #1<<16
+    add     r1, r1, #8
+    bpl     rotated_blit_loop16
+
+    add     lr, lr, #1<<16
+    subs    lr, lr, #1
+
+    add     r0, r0, #4*2
+    add     r1, r1, #(320*3)*2
+    bgt     rotated_blit_loop16_o
+
+    ldmfd   sp!,{r4-r8,pc}
+
+
+.global rotated_blit8 @ void *dst, void *linesx4
+rotated_blit8:
+    stmfd   sp!,{r4-r8,lr}
+
+    mov     r8, #320
+    sub     r0, r0, #240	@ adjust
+    mov     lr, #240/4
+
+rotated_blit8_loop_o:
+    orr     lr, lr, #((320/4)-1) << 16
+    add     r0, r0, #(240*320)
+
+rotated_blit8_loop:
+    mov     r6, r1
+    ldr     r2, [r6], r8
+    ldr     r3, [r6], r8
+    ldr     r4, [r6], r8
+    ldr     r5, [r6], r8
+
+    mov     r6, r2, lsl #24
+    mov     r6, r6, lsr #8
+    orr     r6, r6, r3, lsl #24
+    mov     r6, r6, lsr #8
+    orr     r6, r6, r4, lsl #24
+    mov     r6, r6, lsr #8
+    orr     r6, r6, r5, lsl #24
+    str     r6, [r0], #-240
+
+    and     r6, r3, #0xff00
+    and     r7, r2, #0xff00
+    orr     r6, r6, r7, lsr #8
+    and     r7, r4, #0xff00
+    orr     r6, r6, r7, lsl #8
+    and     r7, r5, #0xff00
+    orr     r6, r6, r7, lsl #16
+    str     r6, [r0], #-240
+
+    and     r6, r4, #0xff0000
+    and     r7, r2, #0xff0000
+    orr     r6, r6, r7, lsr #16
+    and     r7, r3, #0xff0000
+    orr     r6, r6, r7, lsr #8
+    and     r7, r5, #0xff0000
+    orr     r6, r6, r7, lsl #8
+    str     r6, [r0], #-240
+
+    mov     r6, r5, lsr #24
+    mov     r6, r6, lsl #8
+    orr     r6, r6, r4, lsr #24
+    mov     r6, r6, lsl #8
+    orr     r6, r6, r3, lsr #24
+    mov     r6, r6, lsl #8
+    orr     r6, r6, r2, lsr #24
+    str     r6, [r0], #-240
+
+    subs    lr, lr, #1<<16
+    add     r1, r1, #4
+    bpl     rotated_blit8_loop
+
+    add     lr, lr, #1<<16
+    subs    lr, lr, #1
+
+    add     r0, r0, #4
+    add     r1, r1, #320*3
+    bgt     rotated_blit8_loop_o
+
+    ldmfd   sp!,{r4-r8,pc}
+
diff --git a/dist/ginge.gpe b/dist/ginge.gpe
index 27521e5..0ca9ad1 100755
--- a/dist/ginge.gpe
+++ b/dist/ginge.gpe
@@ -1,7 +1,9 @@
 #!/bin/sh
 
+unset LD_PRELOAD
+
 /sbin/rmmod warm 2> /dev/null
-/sbin/insmod ./warm_2.6.24.ko
+/sbin/insmod ./tools/warm_2.6.24.ko
 
 # theoretically GP2X apps can make use of more RAM, because
 # Wiz has 2.6 kernel (larger memory requirements) and larger
@@ -9,6 +11,7 @@
 mkswap swapfile
 swapon swapfile
 
-./gp2xmenu
+./gp2xmenu --view-game
 
 swapoff swapfile
+/sbin/rmmod warm 2> /dev/null
diff --git a/dist/ginge.sh b/dist/ginge.sh
index 66a99e1..69e4b3a 100755
--- a/dist/ginge.sh
+++ b/dist/ginge.sh
@@ -4,7 +4,7 @@ export FBDEV=/dev/fb1
 ofbset -fb $FBDEV -pos 80 0 -size 640 480 -mem 614400 -en 1
 fbset -fb $FBDEV -g 320 240 320 480 16
 
-# make ir runnable from ssh
+# make it runnable from ssh
 if [ -z "$DISPLAY" ]; then
 export DISPLAY=:0
 fi
diff --git a/loader/Makefile b/loader/Makefile
index 0a28670..b546c7c 100644
--- a/loader/Makefile
+++ b/loader/Makefile
@@ -28,10 +28,11 @@ OBJ += host_pnd.o
 endif
 ifdef WIZ
 CFLAGS += -DWIZ
-OBJ += host_wiz.o
+OBJ += host_wiz.o wiz_video_arm.o
 endif
 
 vpath %.c = ../common/
+vpath %.s = ../common/
 
 TARGET_S = ginge_sloader
 TARGET_D = ginge_dyn
diff --git a/loader/emu.c b/loader/emu.c
index 3ab4b34..993ed55 100644
--- a/loader/emu.c
+++ b/loader/emu.c
@@ -269,21 +269,26 @@ bad_blit:
 // FIXME: pass real dimensions to blitters
 static void mlc_flip(void *src, int bpp)
 {
-  u32 *srcp = NULL;
+  static int old_bpp;
 
   // only pass pal to host if it's dirty
   if (bpp <= 8 && mmsp2.dirty_pal) {
-    srcp = mmsp2.mlc_stl_pallt_d32;
+    host_video_update_pal(mmsp2.mlc_stl_pallt_d32);
     mmsp2.dirty_pal = 0;
   }
 
+  if (bpp != old_bpp) {
+    host_video_change_bpp(bpp);
+    old_bpp = bpp;
+  }
+
   switch (bpp) {
   case  4:
-    host_video_blit4(src, 320, 240, srcp);
+    host_video_blit4(src, 320, 240);
     break;
 
   case  8:
-    host_video_blit8(src, 320, 240, srcp);
+    host_video_blit8(src, 320, 240);
     break;
 
   case 16:
diff --git a/make_wiz.sh b/make_wiz.sh
index 9176fc2..cc8f2b0 100755
--- a/make_wiz.sh
+++ b/make_wiz.sh
@@ -8,5 +8,6 @@ cp dist/ginge.gpe out_wiz/
 cp dist/ginge32.png out_wiz/ginge.png
 cp dist/ginge_dyn_oabi.sh out_wiz/ginge_dyn.sh
 cp tools/cramfsck_oabi out_wiz/tools/cramfsck
+cp tools/warm_2.6.24.ko out_wiz/tools/
 
 dd if=/dev/zero of=out_wiz/swapfile bs=1M count=16
diff --git a/prep/main.c b/prep/main.c
index 8acbef7..a03bc38 100644
--- a/prep/main.c
+++ b/prep/main.c
@@ -258,6 +258,13 @@ int main(int argc, char *argv[])
     return 1;
   }
 
+  if (strcmp(argv[1], "--cleanup") == 0) {
+    // as loader may crash eny time, restore screen for them menu
+    host_video_init(NULL, 0);
+    host_video_finish();
+    return 0;
+  }
+
   if (getcwd(cwd, sizeof(cwd)) == NULL) {
     err(PFX "failed to get cwd\n");
     return 1;
@@ -414,6 +421,10 @@ pass:
   fclose(fin);
 
 no_in_script:
+#ifdef WIZ
+  fprintf(fout, "sync\n");
+  fprintf(fout, "%sginge_prep --cleanup\n", root_path);
+#endif
   if (rerun_gp2xmenu) {
     fprintf(fout, "cd %s\n", root_path);
     fprintf(fout, "exec %s%s\n", root_path, LAUNCHER);