diff --git a/compiler/res/prog8lib/cx16/gfx2.p8 b/compiler/res/prog8lib/cx16/gfx2.p8
index 46cf47276..0c259672d 100644
--- a/compiler/res/prog8lib/cx16/gfx2.p8
+++ b/compiler/res/prog8lib/cx16/gfx2.p8
@@ -70,21 +70,23 @@ gfx2 {
 
         active_mode = mode
         if bpp
-            clear_screen()
+            clear_screen(0)
     }
 
-    sub clear_screen() {
+    sub clear_screen(ubyte color) {
         position(0, 0)
         when active_mode {
             1 -> {
                 ; lores 256c
                 repeat 240/2
-                    cs_innerloop640()
+                    cs_innerloop640(color)
             }
             2 -> {
                 ; highres 4c
+                ubyte[] colors = [%00000000, %01010101, %10101010, %11111111]
+                color = colors[color&3]
                 repeat 480/4
-                    cs_innerloop640()
+                    cs_innerloop640(color)
             }
         }
         position(0, 0)
@@ -104,6 +106,8 @@ gfx2 {
     }
 
     sub fillrect(uword xx, uword yy, uword rwidth, uword rheight, ubyte color) {
+        ; Draw a filled rectangle of the given size and color.
+        ; To fill the whole screen, use clear_screen(color) instead - it is much faster.
         if rwidth==0
             return
         repeat rheight {
@@ -784,17 +788,17 @@ skip:
         }
     }
 
-    asmsub cs_innerloop640() clobbers(Y) {
+    asmsub cs_innerloop640(ubyte color @A) clobbers(Y) {
         %asm {{
             ldy  #80
--           stz  cx16.VERA_DATA0
-            stz  cx16.VERA_DATA0
-            stz  cx16.VERA_DATA0
-            stz  cx16.VERA_DATA0
-            stz  cx16.VERA_DATA0
-            stz  cx16.VERA_DATA0
-            stz  cx16.VERA_DATA0
-            stz  cx16.VERA_DATA0
+-           sta  cx16.VERA_DATA0
+            sta  cx16.VERA_DATA0
+            sta  cx16.VERA_DATA0
+            sta  cx16.VERA_DATA0
+            sta  cx16.VERA_DATA0
+            sta  cx16.VERA_DATA0
+            sta  cx16.VERA_DATA0
+            sta  cx16.VERA_DATA0
             dey
             bne  -
             rts
diff --git a/compiler/res/prog8lib/cx16/monogfx.p8 b/compiler/res/prog8lib/cx16/monogfx.p8
index c97f1a4fa..4e5e4d12e 100644
--- a/compiler/res/prog8lib/cx16/monogfx.p8
+++ b/compiler/res/prog8lib/cx16/monogfx.p8
@@ -26,7 +26,7 @@ monogfx {
         cx16.VERA_L1_TILEBASE = 0
         width = 320
         height = 240
-        clear_screen()
+        clear_screen(0)
     }
 
     sub hires() {
@@ -40,7 +40,7 @@ monogfx {
         cx16.VERA_L1_TILEBASE = %00000001
         width = 640
         height = 480
-        clear_screen()
+        clear_screen(0)
     }
 
     sub textmode() {
@@ -50,17 +50,17 @@ monogfx {
         cx16.VERA_DC_VIDEO = (cx16.VERA_DC_VIDEO & %11111000) | cx16.r15L
     }
 
-    sub clear_screen() {
+    sub clear_screen(ubyte color) {
         stipple(false)
         position(0, 0)
         when width {
             320 -> {
                 repeat 240/2/8
-                    cs_innerloop640()
+                    cs_innerloop640(color)
             }
             640 -> {
                 repeat 480/8
-                    cs_innerloop640()
+                    cs_innerloop640(color)
             }
         }
         position(0, 0)
@@ -84,6 +84,8 @@ monogfx {
     }
 
     sub fillrect(uword xx, uword yy, uword rwidth, uword rheight, bool draw) {
+        ; Draw a filled rectangle of the given size and color.
+        ; To fill the whole screen, use clear_screen(color) instead - it is much faster.
         if rwidth==0
             return
         repeat rheight {
@@ -765,17 +767,20 @@ skip:
         }
     }
 
-    asmsub cs_innerloop640() clobbers(Y) {
+    asmsub cs_innerloop640(ubyte color @A) clobbers(Y) {
         %asm {{
-            ldy  #80
--           stz  cx16.VERA_DATA0
-            stz  cx16.VERA_DATA0
-            stz  cx16.VERA_DATA0
-            stz  cx16.VERA_DATA0
-            stz  cx16.VERA_DATA0
-            stz  cx16.VERA_DATA0
-            stz  cx16.VERA_DATA0
-            stz  cx16.VERA_DATA0
+            cmp  #0
+            beq  +
+            lda  #255
++           ldy  #80
+-           sta  cx16.VERA_DATA0
+            sta  cx16.VERA_DATA0
+            sta  cx16.VERA_DATA0
+            sta  cx16.VERA_DATA0
+            sta  cx16.VERA_DATA0
+            sta  cx16.VERA_DATA0
+            sta  cx16.VERA_DATA0
+            sta  cx16.VERA_DATA0
             dey
             bne  -
             rts
diff --git a/examples/cx16/testgfx2.p8 b/examples/cx16/testgfx2.p8
index 2e405693a..50f8ba2ba 100644
--- a/examples/cx16/testgfx2.p8
+++ b/examples/cx16/testgfx2.p8
@@ -78,7 +78,8 @@ main {
         }
 
         sys.wait(60)
-        gfx2.clear_screen()
+        gfx2.clear_screen(2)
+        gfx2.clear_screen(0)
 
         ubyte radius
 
diff --git a/examples/cx16/testmonogfx.p8 b/examples/cx16/testmonogfx.p8
index 4c00f7833..00fa408c9 100644
--- a/examples/cx16/testmonogfx.p8
+++ b/examples/cx16/testmonogfx.p8
@@ -189,7 +189,8 @@ main {
         }
 
         sys.wait(60)
-        monogfx.clear_screen()
+        monogfx.clear_screen(1)
+        monogfx.clear_screen(0)
 
         ubyte radius