Add a 2bpp CGB boot ROM logo, pending palettes

BootROMs/cgb_boot.asm

@@ -87,20 +87,24 @@ ELSE
 
 .tilemapRowLoop
 
+    call .write_with_palette
+
+    ; Repeat the 3 tiles common between E and B. This saves 27 bytes after
+    ; compression, with a cost of 17 bytes of code.
     push af
-    ; Switch to second VRAM Bank
+    sub $20
-    ld a, 1
+    sub $3
-    ldh [$4F], a
+    jr nc, .notspecial
-    ld [hl], 8
+    add $20
-    ; Switch to back first VRAM Bank
+    call .write_with_palette
-    xor a
+    dec c
-    ldh [$4F], a
+.notspecial
     pop af
-    ldi [hl], a
+
-    add d
+    add d ; d = 3 for SameBoy logo, d = 1 for Nintendo logo
     dec c
     jr nz, .tilemapRowLoop
-    sub 47
+    sub 44
     push de
     ld de, $10
     add hl, de
@@ -116,6 +120,19 @@ ELSE
     ld l, $a7
     ld bc, $0107
     jr .tilemapRowLoop
+
+.write_with_palette
+    push af
+    ; Switch to second VRAM Bank
+    ld a, 1
+    ldh [$4F], a
+    ld [hl], 8
+    ; Switch to back first VRAM Bank
+    xor a
+    ldh [$4F], a
+    pop af
+    ldi [hl], a
+    ret
 .endTilemap
 ENDC
 
@@ -532,7 +549,7 @@ TrademarkSymbol:
     db $3c,$42,$b9,$a5,$b9,$a5,$42,$3c
 
 SameBoyLogo:
-    incbin "SameBoyLogo.pb8"
+    incbin "SameBoyLogo.pb12"
 
 AnimationColors:
     dw $7FFF ; White
@@ -634,30 +651,28 @@ ReadCGBLogoHalfTile:
     ld a, e
     ret
 
-; LoadTileset using PB8 codec, 2019 Damian Yerrick
+; LoadTileset using PB12 codec, 2020 Jakub Kądziołka
-;
+; (based on PB8 codec, 2019 Damian Yerrick)
-; The logo is compressed using PB8, a form of RLE with unary-coded
-; run lengths.  Each block representing 8 bytes consists of a control
-; byte, where each bit (MSB to LSB) is 0 for literal or 1 for repeat
-; previous, followed by the literals in that block.
 
 SameBoyLogo_dst = $8080
 SameBoyLogo_length = (128 * 24) / 64
 
 LoadTileset:
     ld hl, SameBoyLogo
-    ld de, SameBoyLogo_dst
+    ld de, SameBoyLogo_dst - 1
     ld c, SameBoyLogo_length
-.pb8BlockLoop:
+.refill
-    ; Register map for PB8 decompression
+    ; Register map for PB12 decompression
     ; HL: source address in boot ROM
     ; DE: destination address in VRAM
     ; A: Current literal value
     ; B: Repeat bits, terminated by 1000...
-    ; C: Number of 8-byte blocks left in this block
     ; Source address in HL lets the repeat bits go straight to B,
     ; bypassing A and avoiding spilling registers to the stack.
     ld b, [hl]
+    dec b
+    jr z, .sameboyLogoEnd
+    inc b
     inc hl
 
     ; Shift a 1 into lower bit of shift value.  Once this bit
@@ -665,26 +680,53 @@ LoadTileset:
     scf
     rl b
 
-.pb8BitLoop:
+.loop
     ; If not a repeat, load a literal byte
-    jr c,.pb8Repeat
+    jr c, .simple_repeat
-    ld a, [hli]
-.pb8Repeat:
-    ; Decompressed data uses colors 0 and 1, so write once, inc twice
-    ld [de], a
-    inc de
-    inc de
     sla b
-    jr nz, .pb8BitLoop
+    jr c, .shifty_repeat
-
+    ld a, [hli]
-    dec c
+    jr .got_byte
-    jr nz, .pb8BlockLoop
+.shifty_repeat
-
+    sla b
-; End PB8 decoding.  The rest uses HL as the destination
+    jr nz, .no_refill_during_shift
-    ld h, d
+    ld b, [hl] ; see above. Also, no, factoring it out into a callable
-    ld l, e
+    inc hl ; routine doesn't save bytes, even with conditional calls
+    scf
+    rl b
+.no_refill_during_shift
+    ld c, a
+    jr nc, .shift_left
+    srl a
+    db $fe ; eat the add a with cp d8
+.shift_left
+    add a
+    sla b
+    jr c, .go_and
+    or c
+    db $fe ; eat the and c with cp d8
+.go_and
+    and c
+    jr .got_byte
+.simple_repeat
+    sla b
+    jr c, .got_byte
+    ; far repeat
+    dec de
+    ld a, [de]
+    inc de
+.got_byte
+    inc de
+    ld [de], a
+    sla b
+    jr nz, .loop
+    jr .refill
 
+; End PB12 decoding.  The rest uses HL as the destination
 .sameboyLogoEnd
+    ld h, d
+    ld l, $80
+
 ; Copy (unresized) ROM logo
     ld de, $104
 .CGBROMLogoLoop

BootROMs/pb12.py

@@ -0,0 +1,68 @@
+import sys
+
+def opts(byte):
+    # top bit: 0 = left, 1 = right
+    # bottom bit: 0 = or, 1 = and
+    if byte is None: return []
+    return [
+        byte | (byte << 1) & 0xff,
+        byte & (byte << 1),
+        byte | (byte >> 1) & 0xff,
+        byte & (byte >> 1),
+    ]
+
+def pb12(data):
+    data = iter(data)
+
+    literals = bytearray()
+    bits = 0
+    control = 0
+    prev = [None, None]
+    gotta_end = False
+
+    chunk = bytearray()
+    while True:
+        try:
+            byte = next(data)
+        except StopIteration:
+            if bits == 0: break
+            byte = 0
+        chunk.append(byte)
+
+        if byte in prev:
+            bits += 2
+            control <<= 1
+            control |= 1
+            control <<= 1
+            if prev[1] == byte:
+                control |= 1 # 10 = out[-2], 11 = out[-1]
+        else:
+            bits += 2
+            control <<= 2
+            options = opts(prev[1])
+            if byte in options:
+                # 01 = modify
+                control |= 1
+
+                bits += 2
+                control <<= 2
+                control |= options.index(byte)
+            else:
+                # 00 = literal
+                literals.append(byte)
+        prev = [prev[1], byte]
+        if bits >= 8:
+            outctl = control >> (bits - 8)
+            assert outctl != 1 # that's the end byte
+            yield bytes([outctl]) + literals
+            bits -= 8
+            control &= (1 << bits) - 1
+            literals = bytearray()
+            chunk = bytearray()
+    yield b'\x01'
+
+_, infile, outfile = sys.argv
+with open(infile, 'rb') as f:
+    data = f.read()
+with open(outfile, 'wb') as f:
+    f.writelines(pb12(data))

Makefile

@@ -382,21 +382,18 @@ $(BIN)/SDL/Shaders: Shaders
 
 # Boot ROMs
 
-$(OBJ)/%.1bpp: %.png
+$(OBJ)/%.2bpp: %.png
 	-@$(MKDIR) -p $(dir $@)
-	rgbgfx -d 1 -h -o $@ $<
+	rgbgfx -h -u -o $@ $<
 
-$(OBJ)/BootROMs/SameBoyLogo.pb8: $(OBJ)/BootROMs/SameBoyLogo.1bpp $(PB8_COMPRESS)
+$(OBJ)/BootROMs/SameBoyLogo.pb12: $(OBJ)/BootROMs/SameBoyLogo.2bpp BootROMs/pb12.py
-	$(realpath $(PB8_COMPRESS)) -l 384 $< $@
+	python3 BootROMs/pb12.py $< $@
-
-$(PB8_COMPRESS): BootROMs/pb8.c
-	$(CC) $< -o $@
 
 $(BIN)/BootROMs/agb_boot.bin: BootROMs/cgb_boot.asm
 $(BIN)/BootROMs/cgb_boot_fast.bin: BootROMs/cgb_boot.asm
 $(BIN)/BootROMs/sgb2_boot: BootROMs/sgb_boot.asm
 
-$(BIN)/BootROMs/%.bin: BootROMs/%.asm $(OBJ)/BootROMs/SameBoyLogo.pb8
+$(BIN)/BootROMs/%.bin: BootROMs/%.asm $(OBJ)/BootROMs/SameBoyLogo.pb12
 	-@$(MKDIR) -p $(dir $@)
 	rgbasm -i $(OBJ)/BootROMs/ -i BootROMs/ -o $@.tmp $<
 	rgblink -o $@.tmp2 $@.tmp