Blit video memory in three-scanline-chunks for extra speed.

Author: davidgiven

arch/wp2450ds/tty.z80

@@ -169,6 +169,7 @@ clear_bytes:
     jr nz, .1
     ret
 
+    label XXX
 tty_delete_line:
     ld a, (tty_cursory)
     cp SCREEN_HEIGHT - 1
@@ -185,22 +186,20 @@ tty_delete_line:
     sub b                   ; calculate number of text lines to move
     ld b, a
                             ; 1
-    add a                   ; 0
-    add a                   ; 0
     add a                   
     add b                   ; 1
-    ld d, a                 ; b = number of scanlines to move
+    ld d, a                 ; b = number of three-line chunks to move
 
-    ld bc, -SCREEN_WIDTH
+    ld bc, -(SCREEN_WIDTH*3)
     add hl, bc              ; compensate for preincrement in loop
 .1
-    ld bc, SCREEN_WIDTH*(CHAR_HEIGHT+1)
+    ld bc, SCREEN_WIDTH*(CHAR_HEIGHT+3)
     add hl, bc
-    call read_scanline
+    call read_chunk
 
     ld bc, -(SCREEN_WIDTH*CHAR_HEIGHT)
     add hl, bc
-    call write_scanline
+    call write_chunk
 
     dec d
     jr nz, .1
@@ -217,27 +216,26 @@ tty_insert_line:
     cp SCREEN_HEIGHT - 1
     jr z, on_last_line      ; skip scroll if there's nothing to do
 
-    ld hl, SCREEN_WIDTH*SCREEN_HEIGHT*CHAR_HEIGHT - SCREEN_WIDTH
+    ; Source vdeo memory address.
+    ld hl, SCREEN_WIDTH*SCREEN_HEIGHT*CHAR_HEIGHT - SCREEN_WIDTH*3
 
     ld b, a
     ld a, SCREEN_HEIGHT - 1
     sub b                   ; calculate number of text lines to move
     ld b, a
                             ; 1
-    add a                   ; 0
-    add a                   ; 0
     add a                   
     add b                   ; 1
-    ld d, a                 ; b = number of scanlines to move
+    ld d, a                 ; b = number of three-line chunks to move
 
 .1
-    call read_scanline
+    call read_chunk
 
     ld bc, -(SCREEN_WIDTH * CHAR_HEIGHT)
     add hl, bc
-    call write_scanline
+    call write_chunk
 
-    ld bc, SCREEN_WIDTH * (CHAR_HEIGHT-1)
+    ld bc, SCREEN_WIDTH * (CHAR_HEIGHT-3)
     add hl, bc
 
     dec d
@@ -252,63 +250,49 @@ tty_insert_line:
     ld hl, SCREEN_WIDTH * 1 * CHAR_HEIGHT
     jp clear_bytes
 
-; Loads the scanline at video memory address HL into the buffer.
-read_scanline:
-    push hl
-    out0l PORT_VIDEO_ADDR_LO
-    out0h PORT_VIDEO_ADDR_HI
-    ld hl, video_buffer
-    out0l MAR1L
-    out0h MAR1H
-    ld a, 6
-    out0a MAR1B
-    ld hl, SCREEN_WIDTH
-    out0l BCR1L
-    out0h BCR1H
-    ld hl, PORT_VIDEO_DATA_R
-    out0l IAR1L
-    out0h IAR1H
+; Loads the chunk at video memory address HL into the buffer.
+read_chunk:
+    ld a, PORT_VIDEO_DATA_R
+    call setup_dma
     ld a, 01000110b         ; One memory wait state, 0 I/O wait state, I/O->memory
     out0a DCNTL
     ld a, 10010000b         ; DE1 enable, DWE0 disable
     out0a DSTAT             ; perform the transfer
-
-.1
+wait_for_dma_end:
     in0a DSTAT
     and 10000000b           ; test for DMA completion
-    jr nz, .1
-
-    pop hl
+    jr nz, wait_for_dma_end
     ret
 
-; Writes the scanline at video memory address HL from the buffer.
-write_scanline:
-    push hl
-    out0l PORT_VIDEO_ADDR_LO
-    out0h PORT_VIDEO_ADDR_HI
-    ld hl, video_buffer
-    out0l MAR1L
-    out0h MAR1H
-    ld a, 6
-    out0a MAR1B
-    ld hl, SCREEN_WIDTH
-    out0l BCR1L
-    out0h BCR1H
-    ld hl, PORT_VIDEO_DATA_W
-    out0l IAR1L
-    out0h IAR1H
+; Writes the chunk at video memory address HL from the buffer.
+write_chunk:
+    ld a, PORT_VIDEO_DATA_W
+    call setup_dma
     ld a, 01000100b         ; One memory wait state, 0 I/O wait state, memory->I/O
     out0a DCNTL
     ld a, 10010000b         ; DE1 enable, DWE0 disable
     out0a DSTAT             ; perform the transfer
+    jr wait_for_dma_end
 
-.1
-    in0a DSTAT
-    and 10000000b           ; test for DMA completion
-    jr nz, .1
-
+; On entry: HL is the video memory address, and A is the I/O port.
+setup_dma:
+    ld (scanline_dma + IAR1L - MAR1L), a
+    out0l PORT_VIDEO_ADDR_LO
+    out0h PORT_VIDEO_ADDR_HI
+    push hl
+    ld b, scanline_dma.end - scanline_dma
+    ld c, MAR1L
+    ld hl, scanline_dma
+    otimr
     pop hl
     ret
+scanline_dma:
+    dw video_buffer         ; MAR1
+    db 0x06                 ; MAR1B
+    dw PORT_VIDEO_DATA_W    ; IAR1
+    db 0                    ; unused
+    dw SCREEN_WIDTH*3       ; BCR1
+scanline_dma.end:
 
 toggle_cursor:
     ld de, (tty_cursorx)
@@ -348,7 +332,7 @@ CUROFF:
     ret
 
 video_buffer:
-    ds 91
+    ds 91*3
 
 font_table:
     include "font.inc"