fix bgr2rgb16 and reduce mode change glitching

5 files changed, 62 insertions, 38 deletions

diff --git a/frontend/arm_utils.s b/frontend/arm_utils.s
index b0585b7..4736380 100644
--- a/frontend/arm_utils.s
+++ b/frontend/arm_utils.s
@@ -13,38 +13,44 @@
 
 .global bgr555_to_rgb565
 bgr555_to_rgb565:
-    mov         r3, #0x03e0
+    mov         r3, #0x07c0
     vdup.16     q15, r3
-    mov         r2, r2, lsr #6
+    sub         r2, r2, #64
 0:
     vldmia      r1!, {q0-q3}
-    vshr.u16    q4, q0, #10
-    vshr.u16    q5, q1, #10
-    vshr.u16    q6, q2, #10
-    vshr.u16    q7, q3, #10
-    vshl.u16    q8, q0, #11
-    vshl.u16    q9, q1, #11
-    vshl.u16    q10, q2, #11
-    vshl.u16    q11, q3, #11
-    vand        q0, q0, q15
-    vand        q1, q1, q15
-    vand        q2, q2, q15
-    vand        q3, q3, q15
+    vshl.u16    q4, q0, #11
+    vshl.u16    q5, q1, #11
+    vshl.u16    q6, q2, #11
+    vshl.u16    q7, q3, #11
+    vsri.u16    q4, q0, #10
+    vsri.u16    q5, q1, #10
+    vsri.u16    q6, q2, #10
+    vsri.u16    q7, q3, #10
     vshl.u16    q0, q0, #1
     vshl.u16    q1, q1, #1
     vshl.u16    q2, q2, #1
     vshl.u16    q3, q3, #1
-    vorr        q0, q0, q4
-    vorr        q1, q1, q5
-    vorr        q2, q2, q6
-    vorr        q3, q3, q7
-    vorr        q0, q0, q8
-    vorr        q1, q1, q9
-    vorr        q2, q2, q10
-    vorr        q3, q3, q11
-    vstmia      r0!, {q0-q3}
-    subs        r2, r2, #1
-    bne         0b
+    vbit        q4, q0, q15
+    vbit        q5, q1, q15
+    vbit        q6, q2, q15
+    vbit        q7, q3, q15
+    vstmia      r0!, {q4-q7}
+    subs        r2, r2, #64
+    bge         0b
+
+    adds        r2, r2, #64
+    bxeq        lr
+
+    @ handle the remainder
+0:
+    vld1.16     {q0}, [r1, :64]!
+    vshl.u16    q1, q0, #11
+    vshl.u16    q2, q0, #1
+    vsri.u16    q1, q0, #10
+    vbit        q1, q2, q15
+    subs        r2, r2, #16
+    vst1.16     {q1}, [r0, :64]!
+    bgt         0b
 
     bx          lr
 
diff --git a/frontend/linux/fbdev.c b/frontend/linux/fbdev.c
index b4889ea..bcef0fc 100644
--- a/frontend/linux/fbdev.c
+++ b/frontend/linux/fbdev.c
@@ -61,7 +61,8 @@ void vout_fbdev_wait_vsync(struct vout_fbdev *fbdev)
 	ioctl(fbdev->fd, FBIO_WAITFORVSYNC, &arg);
 }
 
-int vout_fbdev_resize(struct vout_fbdev *fbdev, int w, int h, int bpp,
+/* it is recommended to call vout_fbdev_clear() before this */
+void *vout_fbdev_resize(struct vout_fbdev *fbdev, int w, int h, int bpp,
 		      int left_border, int right_border, int top_border, int bottom_border, int buffer_cnt)
 {
 	int w_total = left_border + w + right_border;
@@ -83,13 +84,20 @@ int vout_fbdev_resize(struct vout_fbdev *fbdev, int w, int h, int bpp,
 		fbdev->fbvar_new.xres_virtual = w_total;
 		fbdev->fbvar_new.yres_virtual = h_total;
 		fbdev->fbvar_new.xoffset = left_border;
+		fbdev->fbvar_new.yoffset = top_border;
 		fbdev->fbvar_new.bits_per_pixel = bpp;
 		printf(" switching to %dx%d@%d\n", w, h, bpp);
+
+		// seems to help a bit to avoid glitches
+		vout_fbdev_wait_vsync(fbdev);
+
 		ret = ioctl(fbdev->fd, FBIOPUT_VSCREENINFO, &fbdev->fbvar_new);
 		if (ret == -1) {
 			perror("FBIOPUT_VSCREENINFO ioctl");
-			return -1;
+			return NULL;
 		}
+
+		fbdev->buffer_write = 1;
 	}
 
 	fbdev->buffer_count = buffer_cnt;
@@ -99,6 +107,7 @@ int vout_fbdev_resize(struct vout_fbdev *fbdev, int w, int h, int bpp,
 		ret = ioctl(fbdev->fd, FBIOPUT_VSCREENINFO, &fbdev->fbvar_new);
 		if (ret == -1) {
 			fbdev->buffer_count = 1;
+			fbdev->buffer_write = 0;
 			fprintf(stderr, "Warning: failed to increase virtual resolution, "
 					"doublebuffering disabled\n");
 		}
@@ -110,7 +119,7 @@ int vout_fbdev_resize(struct vout_fbdev *fbdev, int w, int h, int bpp,
 
 	mem_size = fbdev->fb_size * fbdev->buffer_count;
 	if (fbdev->mem_size >= mem_size)
-		return 0;
+		goto out;
 
 	if (fbdev->mem != NULL)
 		munmap(fbdev->mem, fbdev->mem_size);
@@ -126,11 +135,13 @@ int vout_fbdev_resize(struct vout_fbdev *fbdev, int w, int h, int bpp,
 		fbdev->mem = NULL;
 		fbdev->mem_size = 0;
 		perror("mmap framebuffer");
-		return -1;
+		return NULL;
 	}
 
 	fbdev->mem_size = mem_size;
-	return 0;
+
+out:
+	return (char *)fbdev->mem + fbdev->fb_size * fbdev->buffer_write;
 }
 
 void vout_fbdev_clear(struct vout_fbdev *fbdev)
@@ -160,6 +171,7 @@ struct vout_fbdev *vout_fbdev_init(const char *fbdev_name, int *w, int *h, int b
 {
 	struct vout_fbdev *fbdev;
 	int req_w, req_h;
+	void *pret;
 	int ret;
 
 	fbdev = calloc(1, sizeof(*fbdev));
@@ -188,8 +200,8 @@ struct vout_fbdev *vout_fbdev_init(const char *fbdev_name, int *w, int *h, int b
 	if (*h != 0)
 		req_h = *h;
 
-	ret = vout_fbdev_resize(fbdev, req_w, req_h, bpp, 0, 0, 0, 0, buffer_cnt);
-	if (ret != 0)
+	pret = vout_fbdev_resize(fbdev, req_w, req_h, bpp, 0, 0, 0, 0, buffer_cnt);
+	if (pret == NULL)
 		goto fail;
 
 	printf("%s: %ix%i@%d\n", fbdev_name, fbdev->fbvar_new.xres, fbdev->fbvar_new.yres,
diff --git a/frontend/linux/fbdev.h b/frontend/linux/fbdev.h
index ddbfdab..2db9163 100644
--- a/frontend/linux/fbdev.h
+++ b/frontend/linux/fbdev.h
@@ -3,7 +3,7 @@ struct vout_fbdev;
 struct vout_fbdev *vout_fbdev_init(const char *fbdev_name, int *w, int *h, int bpp, int buffer_count);
 void *vout_fbdev_flip(struct vout_fbdev *fbdev);
 void  vout_fbdev_wait_vsync(struct vout_fbdev *fbdev);
-int   vout_fbdev_resize(struct vout_fbdev *fbdev, int w, int h, int bpp,
+void *vout_fbdev_resize(struct vout_fbdev *fbdev, int w, int h, int bpp,
 			int left_border, int right_border, int top_border, int bottom_border,
 			int buffer_count);
 void  vout_fbdev_clear(struct vout_fbdev *fbdev);
diff --git a/frontend/plugin_lib.c b/frontend/plugin_lib.c
index e2a0d58..35a23c1 100644
--- a/frontend/plugin_lib.c
+++ b/frontend/plugin_lib.c
@@ -28,14 +28,18 @@ int pl_fbdev_init(void)
 
 int pl_fbdev_set_mode(int w, int h, int bpp)
 {
-	int ret;
+	void *ret;
 
 	pl_fbdev_w = w;
-	printf("set mode %dx%d@%d\n", w, h, bpp);
+
+	vout_fbdev_clear(layer_fb);
 	ret = vout_fbdev_resize(layer_fb, w, h, bpp, 0, 0, 0, 0, 3);
-	if (ret)
+	if (ret == NULL)
 		fprintf(stderr, "failed to set mode\n");
-	return ret;
+	else
+		pl_fbdev_buf = ret;
+
+	return (ret != NULL) ? 0 : -1;
 }
 
 void *pl_fbdev_flip(void)
diff --git a/plugins/dfxvideo/draw_fb.c b/plugins/dfxvideo/draw_fb.c
index f3d0cf2..b39e2ee 100644
--- a/plugins/dfxvideo/draw_fb.c
+++ b/plugins/dfxvideo/draw_fb.c
@@ -48,7 +48,7 @@ static void blit(void)
 {
  extern void bgr555_to_rgb565(void *dst, void *src, int bytes);
  extern void bgr888_to_rgb888(void *dst, void *src, int bytes);
- int x = PSXDisplay.DisplayPosition.x;
+ int x = PSXDisplay.DisplayPosition.x & ~3; // XXX: align needed by bgr*_to_...
  int y = PSXDisplay.DisplayPosition.y;
  int w = PreviousPSXDisplay.Range.x1;
  int h = PreviousPSXDisplay.DisplayMode.y;
@@ -93,6 +93,8 @@ void DoBufferSwap(void)
  if (PSXDisplay.DisplayMode.x == 0 || PSXDisplay.DisplayMode.y == 0)
   return;
 
+ /* careful if rearranging this code, we try to set mode and flip
+  * to get the hardware apply both changes at the same time */
  if (PSXDisplay.DisplayMode.x != fbw || PSXDisplay.DisplayMode.y != fbh
      || PSXDisplay.RGB24 != fb24bpp) {
   fbw = PSXDisplay.DisplayMode.x;