diff --git a/host/src/STARTUP.s b/host/src/STARTUP.s
index 0852267..bc9ebc8 100755
--- a/host/src/STARTUP.s
+++ b/host/src/STARTUP.s
@@ -58,17 +58,14 @@
         ORA Screen_RAM_settings
         STA Screen_RAM_settings
 
-        ;; Paint the bitmap black. More bitmap: https://www.c64-wiki.com/wiki/53272, https://www.c64-wiki.com/wiki/Screen_RAM#Moving_of_screen_RAM
-        Mov_16 A_start, A_start + 1, #<VIC_bank, #>VIC_bank
-        Mov_16 length, length + 1, #<$1f40,  #>$1f40
-        LDA #$00
-        jsr memset
+        ;; clear the screen
+        JSR clear_screen
 
         ;; Sets the screen color to black and white
         Mov_16 A_start, A_start + 1, #<Screen_RAM, #>Screen_RAM
         Mov_16 length, length + 1, #<$03E8, #>$03E8
         LDA #%11110000
-        jsr memset
+        JSR memset
 
         ;; Disable maskeble interups (not all)
         SEI
diff --git a/host/src/routines/circle/circle_test_position.s b/host/src/routines/circle/circle_test_position.s
index 1c3ea74..ffa296f 100644
--- a/host/src/routines/circle/circle_test_position.s
+++ b/host/src/routines/circle/circle_test_position.s
@@ -21,12 +21,8 @@ loop:
 ;        DEY
 ;        BNE delay_point
 
-        ;;clear screen
-        VIC_bank = $4000
-        Mov_16 A_start, A_start + 1, #<VIC_bank, #>VIC_bank
-        Mov_16 length, length + 1, #<$1f40,  #>$1f40
-        LDA #$00
-        jsr memset
+
+        JSR clear_screen
 
         ;;move circle
         INC $AD
diff --git a/host/src/routines/circle/circle_test_size.s b/host/src/routines/circle/circle_test_size.s
index 02f901b..1c9093a 100644
--- a/host/src/routines/circle/circle_test_size.s
+++ b/host/src/routines/circle/circle_test_size.s
@@ -30,12 +30,7 @@ loop:
         LDA $AD
         STA radius
 
-        ;; clean the screen
-        VIC_bank = $4000
-        Mov_16 A_start, A_start + 1, #<VIC_bank, #>VIC_bank
-        Mov_16 length, length + 1, #<$1f40,  #>$1f40
-        LDA #$00
-        jsr memset
+        JSR clear_screen
 
         ;; draw the circle
         JSR circle
diff --git a/host/src/routines/line/line.s b/host/src/routines/line/line.s
index 9768771..89a50de 100644
--- a/host/src/routines/line/line.s
+++ b/host/src/routines/line/line.s
@@ -5,26 +5,24 @@
         .include "line.inc"
 
 ;; Fix line that is too long
-        LDA Y_pos
-        CMP #$C8 ;;y_max = $C8
-        BCC do_not_fix_y_pos
-        LDA #$C8
+        LDA #$C8 ;y_max = $C8
+        CMP Y_pos
+        BCS do_not_fix_y_pos
         STA Y_pos
 do_not_fix_y_pos:
-        LDA Y_end
-        CMP #$C8 ;;y_max = $C8
-        BCC do_not_fix_y_end
         LDA #$C8
+        CMP Y_pos ;y_max = $C8
+        BCS do_not_fix_y_end
         STA Y_end
 do_not_fix_y_end:
 
-
         ;;dx
         SEC
         LDA X_end
         SBC X_pos
         BCS dx_no_underflow;; X_end >= X_pos
         EOR #$ff ; Fix bit underflow
+        ADC #$01
         STA dx
         ;; line_* expect X_pos < X_end and now its not the case.
         ;; Lets move them around
@@ -36,28 +34,27 @@ do_not_fix_y_end:
         LDY Y_end
         STX Y_end
         STY Y_pos
-
-
 dx_no_underflow:
         STA dx
+
         SEC
-        LDA Y_pos
-        SBC Y_end
-        STA dy
-        BCC down ;normal  Y_pos < Y_end
+        LDA Y_end
+        SBC Y_pos
+        BCS down ;normal  Y_pos < Y_end
 up:; Y_pos > Y_end
+        EOR #$ff ; Fix bit underflow
+        ADC #$01
         STA dy
+
         CMP dx
         BCC shallow; dy < dx
 steep:
         jsr line_up_inv
         RTS
 shallow: ;dy =< dx
-        lda dx
         jsr line_up
         RTS
 down:
-        EOR #$ff ; Fix bit underflow
         STA dy
         CMP dx
         BCC shallow_; dy < dx
diff --git a/host/src/routines/line/line_down.s b/host/src/routines/line/line_down.s
index b81a095..6641a0a 100644
--- a/host/src/routines/line/line_down.s
+++ b/host/src/routines/line/line_down.s
@@ -67,6 +67,7 @@ end_selfmod:
         STA byte_to_paint
         ;; X = X_end - X_pos
         LDX dx
+        INX ; okay if it overflow to $00 because it will go back to $ff in loop.
         Sub_16 btp_mem_pos, btp_mem_pos + 1, #$00, #$00,! ;; Y has always a offset of at least 1 = C + $0000
         INY
 for_x:
diff --git a/host/src/routines/line/line_down_inv.s b/host/src/routines/line/line_down_inv.s
index 465ffdf..4446379 100644
--- a/host/src/routines/line/line_down_inv.s
+++ b/host/src/routines/line/line_down_inv.s
@@ -55,6 +55,7 @@ end_selfmod:
         INY
         Sub_16 btp_mem_pos, btp_mem_pos + 1, #$00, #$00, !
         LDX dy
+        INX
         CLC
 for_y: ; C =0
         LDA byte_to_paint
diff --git a/host/src/routines/line/line_test_extensive.s b/host/src/routines/line/line_test_extensive.s
index 4f79005..8438af5 100644
--- a/host/src/routines/line/line_test_extensive.s
+++ b/host/src/routines/line/line_test_extensive.s
@@ -8,6 +8,8 @@
         X_pos_ = $0E
         Y_end_ = $0F
         X_end_ = $0C
+
+full_angle_test:
         LDA #$d2
         STA X_pos_
         LDA #$62
@@ -34,23 +36,45 @@
         CMP #$ff
         bne @loop
 
-clear_screen:
-        ;;Lets clear bitmap
-        VIC_bank = $4000
-        ;;Paint the bitmap black. More bitmap: https://www.c64-wiki.com/wiki/53272, https://www.c64-wiki.com/wiki/Screen_RAM#Moving_of_screen_RAM
-        Mov_16 A_start, A_start + 1, #<VIC_bank, #>VIC_bank
-        Mov_16 length, length + 1, #<$1f40,  #>$1f40
-        LDA #$00
-        jsr memset
+jsr clear_screen
+
+full_angle_test_pos_end_swapped:
+LDA #$d2
+STA X_end_
+LDA #$62
+STA Y_end_
+LDA #$ff
+STA X_pos_
+LDA #$0
+STA Y_pos_
+
+;; Full angle test
+@loop:
+LDA Y_end_
+STA Y_end
+LDA X_end_
+STA X_end
+LDA X_pos_
+STA X_pos
+LDA Y_pos_
+STA Y_pos
+
+jsr line
+INC Y_pos_
+LDA Y_pos_
+CMP #$ff
+bne @loop
+
+jsr clear_screen
 
 long_line_test_a:
-        LDA #$00
+        LDA #$01
         STA X_pos_
         LDA #$60
         STA Y_pos_
         LDA #$ff
         STA X_end_
-        LDA #$0
+        LDA #$00
         STA Y_end_
 @loop:
         LDA Y_pos_
@@ -68,15 +92,10 @@ long_line_test_a:
         CMP #$ff
         BNE @loop
 
-clear_screen_:
-        ;;Paint the bitmap black. More bitmap: https://www.c64-wiki.com/wiki/53272, https://www.c64-wiki.com/wiki/Screen_RAM#Moving_of_screen_RAM
-        Mov_16 A_start, A_start + 1, #<VIC_bank, #>VIC_bank
-        Mov_16 length, length + 1, #<$1f40,  #>$1f40
-        LDA #$00
-        jsr memset
+        jsr clear_screen
 
 long_line_test_b:
-        LDA #$00
+        LDA #$01
         STA X_pos_
         LDA #$00
         STA Y_pos_
@@ -99,15 +118,10 @@ long_line_test_b:
         CMP #$ff
         BNE @loop
 
-clear_screen__:
-        ;;Paint the bitmap black. More bitmap: https://www.c64-wiki.com/wiki/53272, https://www.c64-wiki.com/wiki/Screen_RAM#Moving_of_screen_RAM
-        Mov_16 A_start, A_start + 1, #<VIC_bank, #>VIC_bank
-        Mov_16 length, length + 1, #<$1f40,  #>$1f40
-        LDA #$00
-        jsr memset
+jsr clear_screen
 
 long_line_test_b_pos_end_swapped:
-        LDA #$00
+        LDA #$01
         STA X_end_
         LDA #$00
         STA Y_end_
diff --git a/host/src/routines/line/line_up.s b/host/src/routines/line/line_up.s
index 0edae87..24f1ae1 100644
--- a/host/src/routines/line/line_up.s
+++ b/host/src/routines/line/line_up.s
@@ -50,6 +50,7 @@ end_selfmod:
         ;LDY #$01
         INY
         LDX dx
+        INX
 for_x:
         LDA byte_to_paint
         ORA (btp_mem_pos), Y
diff --git a/host/src/routines/line/line_up_inv.s b/host/src/routines/line/line_up_inv.s
index 5ad7940..44719e8 100644
--- a/host/src/routines/line/line_up_inv.s
+++ b/host/src/routines/line/line_up_inv.s
@@ -34,7 +34,7 @@
         STA D + 1
 
         ;; because C flag is wrong value we let dy_2 be 1 to small
-        Sub_16 dy_2, dy_2 +1, #$01,#$00
+        Sub_16 dx_2, dx_2 +1, #$01,#$00
 
 selfmod:
         LDA dx_2
@@ -52,6 +52,7 @@ end_selfmod:
         Sub_16 btp_mem_pos, btp_mem_pos + 1, #$00, #$00,!
         CLC
         LDX dy
+        INX
 for_y:
         LDA byte_to_paint
         ORA (btp_mem_pos), Y
diff --git a/host/src/routines/memory/clear_screen.s b/host/src/routines/memory/clear_screen.s
new file mode 100644
index 0000000..e55edb8
--- /dev/null
+++ b/host/src/routines/memory/clear_screen.s
@@ -0,0 +1,39 @@
+;;; -*- Mode: asm; indent-tabs-mode: t; tab-width: 8 -*-
+
+;; Sets memory in 'A'-registry to all addresses from 'A_start' until 'A_start' + 'length'
+;; Modifies A, X and A_start
+.proc clear_screen
+        .include "mem.inc"
+        VIC_bank = $4000
+        length_ = $1f40
+        low_length_3 = $C0 + $03 ;; we need one more.
+        Mov_16 A_start, A_start + 1, #<VIC_bank, #>VIC_bank
+        LDA #$00
+        ;; big_set sets the memory in $ff chunks.
+        ;; skipp if length >= $ff
+        LDX #>length_
+        BNE big_set
+        JMP small_set
+big_set: ;sets $ff of memory
+        ;; Y value do not matter, will go through all anyway!
+        .repeat $ff
+        STA (A_start), Y
+        DEY
+        .endrepeat
+        STA (A_start), Y ; dont forget Y =0
+big_set_end:
+        ;;set all hole $ff memory chunks!
+        INC A_start + 1
+        DEX ;; length +1 --
+        BEQ small_set
+        JMP big_set
+
+small_set:
+        LDX big_set + low_length_3
+        LDY #$60
+        STY big_set + low_length_3
+        LDY #<length_
+        JSR big_set
+        STX big_set + low_length_3
+        RTS
+.endproc
diff --git a/host/src/routines/memory/memset.s b/host/src/routines/memory/memset.s
old mode 100755
new mode 100644
index 55700e9..37371ad
--- a/host/src/routines/memory/memset.s
+++ b/host/src/routines/memory/memset.s
@@ -1,40 +1,78 @@
-;;; -*- Mode: asm; indent-tabs-mode: t; tab-width: 8 -*-
-
-;; Sets memory in 'A'-registry to all addresses from 'A_start' until 'A_start' + 'length'
-;; Modifies A, X and A_start
-.proc memset
-        .include "mem.inc"
-
-;; big_set sets the memory in $ff chunks.
-;; skipp if length >= $ff
-LDX length +1
-BNE big_set
-JMP small_set
-
-big_set: ;sets $ff of memory
-        ;; Y value do not matter, will go through all anyway!
-        .repeat $ff
-        STA (A_start), Y
-        DEY
-        .endrepeat
-        STA (A_start), Y ; dont forget Y =0
-big_set_end:
-        ;;set all hole $ff memory chunks!
-        INC A_start + 1
-        DEX ;; length +1 --
-        BEQ small_set
-        JMP big_set
-
-
-
-;;sets the rest of the memory
-;; note that this can use code above (smc) or the same method. may implement later.
-small_set:
-       LDY length
-small_set_loop:
-        STA (A_start), Y
-        DEY
-        BNE small_set_loop
-        STA (A_start), Y
-        RTS
-.endproc
+;;; -*- Mode: asm; indent-tabs-mode: t; tab-width: 8 -*-
+
+;; Sets memory in 'A'-registry to all addresses from 'A_start' until 'A_start' + 'length'
+;; Modifies A, X and A_start
+.proc memset
+        .include "mem.inc"
+
+;; big_set sets the memory in $ff chunks.
+;; skipp if length >= $ff
+LDX length +1
+BNE big_set
+JMP small_set
+
+big_set: ;sets $ff of memory
+        ;; Y value do not matter, will go through all anyway!
+        .repeat $ff
+        STA (A_start), Y
+        DEY
+        .endrepeat
+        STA (A_start), Y ; dont forget Y =0
+big_set_end:
+        ;;set all hole $ff memory chunks!
+        INC A_start + 1
+        DEX ;; length +1 --
+        BEQ small_set
+        JMP big_set
+
+;; Note that cpu cykels total: cy_tot = 66 to 69
+;; But we skipp a BNE (cy = 2*) * [length (mod 255)]
+;; The BNE case has an avrige of 2*255/2 = 255 so this is faster (on avrige.)
+
+small_set:
+        STA data_to_write               ; cy = 3
+        LDA length                      ; cy = 3
+        STA length_copy                 ; cy = 3
+
+        ;; calculate rts-position
+        LDX #$00                        ; cy = 2
+        STX length + 1                  ; cy = 3
+        ;; 3 bytes = STA DEY NOP = seting 1 byte of memory.
+        ;; So we need to calculate: length*3
+        Mult_16 A, length + 1           ; cy = 7
+        ; A=  length
+        ADC length_copy                 ; cy = 3
+        TAY
+        LDA length + 1                  ; cy = 3
+        ADC #$00                        ; cy = 2
+        STA length + 1                  ; cy = 3
+
+        ;; Now RTS_pointer + Y = length*3 + big_set_label
+        LDA #<big_set                   ; cy = 2
+        STA RTS_pointer                 ; cy = 3
+        LDA #>big_set                   ; cy = 2
+        ADC length + 1                  ; cy = 3
+        STA RTS_pointer + 1             ; cy = 3
+
+        ;; read data we will change to RTS
+        STY Y_copy                      ; cy = 3
+        LDA (RTS_pointer), Y            ; cy = 5*
+        TAX                             ; cy = 2
+
+        ;; set RTS in big_set
+        LDA #$60                        ; cy = 2
+        STA (RTS_pointer), Y            ; cy = 5*
+
+        ;; JSR to modified big_set
+        LDY length_copy                 ; cy = 3
+        DEY ; because we want to count to Y=0 :)
+        LDA data_to_write               ; cy = 3
+        JSR big_set                     ; cy = 6
+
+        ;; revert changes
+        LDY Y_copy                      ; cy = 3
+        TXA                             ; cy = 2
+        STA (RTS_pointer), Y            ; cy = 5*
+
+        RTS
+.endproc
diff --git a/host/src/routines/memory/memset_alt.s b/host/src/routines/memory/memset_alt.s
deleted file mode 100644
index b2c50b9..0000000
--- a/host/src/routines/memory/memset_alt.s
+++ /dev/null
@@ -1,75 +0,0 @@
-;;; -*- Mode: asm; indent-tabs-mode: t; tab-width: 8 -*-
-
-;; Sets memory in 'A'-registry to all addresses from 'A_start' until 'A_start' + 'length'
-;; Modifies A, X and A_start
-.proc memset
-        .include "mem.inc"
-
-;; big_set sets the memory in $ff chunks.
-;; skipp if length >= $ff
-LDX length +1
-BNE big_set
-JMP small_set
-
-big_set: ;sets $ff of memory
-        ;; Y value do not matter, will go through all anyway!
-        .repeat $ff
-        STA (A_start), Y
-        DEY
-        .endrepeat
-        STA (A_start), Y ; dont forget Y =0
-big_set_end:
-        ;;set all hole $ff memory chunks!
-        INC A_start + 1
-        DEX ;; length +1 --
-        BEQ small_set
-        JMP big_set
-
-
-small_set:
-        STA data_to_write
-        LDA length
-        STA length_copy
-
-        ;; calculate rts-position
-        LDX #$00
-        STX length + 1
-        ;; 3 bytes = STA DEY NOP = seting 1 byte of memory.
-        ;; So we need to calculate: length*3
-        Mult_16 A, length + 1
-        ; A=  length
-        ADC length_copy
-        TAY
-        LDA length + 1
-        ADC #$00
-        STA length + 1
-
-        ;; Now RTS_pointer + Y = length*3 + big_set_label
-        LDA #<big_set
-        STA RTS_pointer
-        LDA #>big_set
-        ADC length + 1
-        STA RTS_pointer + 1
-
-        ;; read data we will change to RTS
-        STY Y_copy
-        LDA (RTS_pointer), Y
-        TAX
-
-        ;; set RTS in big_set
-        LDA #$60
-        STA (RTS_pointer), Y
-
-        ;; JSR to modified big_set
-        LDY length_copy
-        DEY ; because we want to count to Y=0 :)
-        LDA data_to_write
-        JSR big_set
-
-        ;; revert changes
-        LDY Y_copy
-        TXA
-        STA (RTS_pointer), Y
-
-        RTS
-.endproc
diff --git a/host/src/source.s b/host/src/source.s
index 0d63e58..4768068 100644
--- a/host/src/source.s
+++ b/host/src/source.s
@@ -39,7 +39,8 @@ JMP exit
 .include "routines/pixel/pixel_draw.s"
 .include "routines/pixel/pixel_calc.s"
 .include "routines/text/char_draw.s"
-.include "routines/memory/memset_alt.s"
+.include "routines/memory/memset.s"
+.include "routines/memory/clear_screen.s"
 .include "routines/memory/memcpy.s"
 .include "routines/arithmatic/mult.s"
 .include "routines/arithmatic/div.s"