gfx_ScaledSprite_NoClip_Fast:
; Draws a 160x120 sprite scaled 2x to 320x240 at (0,0)
; Optimized for speed.
    push ix                 ; Preserve IX
    ld   iy, 0
    add  iy, sp

    ; --- Setup Source and Destination ---
    ld   hl, (iy + 6)       ; Get sprite pointer (Arg0)
    inc  hl                 ; Skip width byte
    inc  hl                 ; Skip height byte

    ld   de, (CurrentBuffer) ; DE = Top-left of screen (X=0, Y=0)

    ; We use IXL as our row counter (120 rows)
    ld   a, 120
    ld   ixl, a

NcSprRowLoop:
    ; --- Horizontal Scale (160 -> 320 pixels) ---
    ; We process 160 pixels. Unrolled 4x to reduce branch overhead.
    ld   b, 40              ; 40 * 4 = 160 pixels
NcSprPixelLoop:
    repeat 4
        ld   a, (hl)        ; Load 1 pixel
        inc  hl             ; Next source pixel
        ld   (de), a        ; Write pixel once
        inc  de
        ld   (de), a        ; Write pixel twice
        inc  de
    end repeat
    djnz NcSprPixelLoop

    ; --- Vertical Scale (Line Copy) ---
    ; At this point, HL is at the start of the NEXT sprite row.
    ; DE is at the start of the NEXT screen row.
    ; We need to copy the 320 bytes we just wrote to the current screen row.

    push hl                 ; Save sprite pointer

    ; Source for copy = Current DE - 320 bytes
    push de
    pop  hl
    ld   bc, 320
    or   a
    sbc  hl, bc             ; HL = Start of the line we just finished

    ; Destination = DE (Start of the next line)
    ; Count = 320
    ld   bc, 320
    ldir                    ; High-speed copy of the entire line

    ; After LDIR, DE is now at the start of the 3rd row (ready for next loop)
    pop  hl                 ; Restore sprite pointer

    dec  ixl
    jr   nz, NcSprRowLoop

    pop  ix                 ; Restore IX
    ret