fxsdk/fxconv/fxconv.py

"""
Convert data files into gint formats or object files
"""

import os
import tempfile
import subprocess
import collections
import fnmatch
import re

from PIL import Image

__all__ = [
	# Color names
	"FX_BLACK", "FX_DARK", "FX_LIGHT", "FX_WHITE", "FX_ALPHA",
	# Conversion mechanisms
	"ObjectData", "u8", "u16", "u32", "ref", "sym",
	# Functions
	"quantize", "convert", "elf",
	# Reusable classes
	"Area", "Grid",
	# Reusable converters
	"convert_bopti_fx", "convert_bopti_cg",
	"convert_topti",
	"convert_libimg_fx", "convert_libimg_cg",
	# Meta API to use fxconv-metadata.txt files
	"Metadata", "parse_parameters",
]

#
#  Constants
#

# Colors
FX_BLACK  = (  0,   0,   0, 255)
FX_DARK   = ( 85,  85,  85, 255)
FX_LIGHT  = (170, 170, 170, 255)
FX_WHITE  = (255, 255, 255, 255)
FX_ALPHA  = (  0,   0,   0,   0)

# fx-9860G profiles
class FxProfile:
	def __init__(self, id, name, colors, layers):
		"""
		Construct an FxProfile object.
		* [id] is the profile ID in bopti
		* [name] is the profile's name as seen in the "profile" key
		* [colors] is the set of supported colors
		* [layers] is a list of layer functions
		"""

		self.id = id
		self.name = name
		self.gray = FX_LIGHT in colors or FX_DARK in colors
		self.colors = colors
		self.layers = layers

	@staticmethod
	def find(name):
		"""Find a profile by name."""
		for profile in FX_PROFILES:
			if profile.name == name:
				return profile
		return None

FX_PROFILES = [
	# Usual black-and-white bitmaps without transparency, as in MonochromeLib
	FxProfile(0x0, "mono", { FX_BLACK, FX_WHITE }, [
		lambda c: (c == FX_BLACK),
	]),
    # Black-and-white with transparency, equivalent of two bitmaps in ML
	FxProfile(0x1, "mono_alpha", { FX_BLACK, FX_WHITE, FX_ALPHA }, [
		lambda c: (c != FX_ALPHA),
		lambda c: (c == FX_BLACK),
	]),
	# Gray engine bitmaps, reference could have been Eiyeron's Gray Lib
	FxProfile(0x2, "gray", { FX_BLACK, FX_DARK, FX_LIGHT, FX_WHITE }, [
		lambda c: (c in [FX_BLACK, FX_LIGHT]),
		lambda c: (c in [FX_BLACK, FX_DARK]),
	]),
	# Gray images with transparency, unfortunately 3 layers since 5 colors
	FxProfile(0x3, "gray_alpha",
		{ FX_BLACK, FX_DARK, FX_LIGHT, FX_WHITE, FX_ALPHA }, [
		lambda c: (c != FX_ALPHA),
		lambda c: (c in [FX_BLACK, FX_LIGHT]),
		lambda c: (c in [FX_BLACK, FX_DARK]),
	]),
]

# fx-CG 50 profiles
class CgProfile:
	def __init__(self, id, name, alpha):
		"""
		Construct a CgProfile object.
		* [id] is the profile ID in bopti
		* [name] is the profile name as found in the specification key
		* [alpha] is True if this profile supports alpha, False otherwise
		"""

		self.id = id
		self.name = name
		self.supports_alpha = alpha

	@staticmethod
	def find(name):
		"""Find a profile by name."""
		for profile in CG_PROFILES:
			if profile.name == name:
				return profile
		return None

CG_PROFILES = [
	# 16-bit RGB565 and RGB565 with alpha
	CgProfile(0x0, "rgb565", False),
	CgProfile(0x1, "rgb565a", True),
	# 8-bit palette for RGB565 and RGB565A
	CgProfile(0x4, "p8_rgb565", False),
	CgProfile(0x5, "p8_rgb565a", True),
	# 4-bit palette for RGB565 and RGB565A
	CgProfile(0x6, "p4_rgb565", False),
	CgProfile(0x3, "p4_rgb565a", True),

	# Original names for RGB565 and RGB565A
	CgProfile(0x0, "r5g6b5", False),
	CgProfile(0x1, "r5g6b5a", True),
]

# Libimg flags
LIBIMG_FLAG_OWN = 1
LIBIMG_FLAG_RO  = 2

#
#  Character sets
#

FX_CHARSETS = {
	# Digits 0...9
	"numeric": [ (ord('0'), 10) ],
	# Uppercase letters A...Z
	"upper":   [ (ord('A'), 26) ],
	# Upper and lowercase letters A..Z, a..z
	"alpha":   [ (ord('A'), 26), (ord('a'), 26) ],
	# Letters and digits A..Z, a..z, 0..9
	"alnum":   [ (ord('A'), 26), (ord('a'), 26), (ord('0'), 10) ],
	# All printable characters from 0x20 to 0x7e
	"print":   [ (0x20, 95) ],
	# All 128 ASII characters
	"ascii":   [ (0x00, 128) ],
	# Custom Unicode block intervals
	"unicode": [],
}

#
# Conversion mechanisms
#

def u8(x):
    return bytes([ x & 255 ])
def u16(x):
    return bytes([ (x >> 8) & 255, x & 255 ])
def u32(x):
    return bytes([ (x >> 24) & 255, (x >> 16) & 255, (x >> 8) & 255, x & 255 ])

def ref(base, offset=None, padding=None, prefix_underscore=True):
	if isinstance(base, bytes) or isinstance(base, bytearray):
		base = bytes(base)
		if offset is not None:
			raise FxconvError(f"reference to bytes does not allow offset")
		if padding and len(base) % padding != 0:
			base += bytes(padding - len(base) % padding)
		return Ref("bytes", base)

	elif isinstance(base, str):
		if padding is not None:
			raise FxconvError(f"reference to name does not allow padding")
		if prefix_underscore:
			base = "_" + base
		if offset is not None:
			offset = int(offset)
			base = f"{base} + {offset}"
		return Ref("name", base)

	elif isinstance(base, ObjectData):
		if offset is not None or padding is not None:
			raise FxconvError("reference to structure does not allow offset " +
				"or padding")
		return Ref("struct", base)

	else:
		raise FxconvError(f"invalid type {type(base)} for ref()")

def ptr(base):
	return ref(base)

def chars(text, length, require_final_nul=True):
	btext = bytes(text, 'utf-8')
	if len(btext) >= length and require_final_nul:
		raise FxconvError(f"'{text}' does not fit within {length} bytes")
	return btext + bytes(length - len(btext))

def string(text):
	return ref(bytes(text, 'utf-8') + bytes([0]))

def sym(name):
	return Sym("_" + name)

# There are 3 kinds of Refs:
#   "bytes"    -> target is a bytes(), we point to that data
#   "name"     -> target is an str like "_sym+2", we point to that
#   "struct"   -> target is an ObjectData
Ref = collections.namedtuple("Ref", ["kind", "target"])

Sym = collections.namedtuple("Sym", ["name"])

class ObjectData:
	"""
	A sequence of bytes that can contain pointers to external variables or
	other data generated along the output structure.
	"""

	def __init__(self, alignment=4):
		"""Construct an empty ObjectData sequence."""

		if alignment & (alignment - 1) != 0:
			raise FxconvError(f"invalid ObjectData alignment {align} (not a " +
				"power of 2)")
		self.alignment = alignment

		# Elements in the structure: bytes, Ref, Sym, ObjectData
		self.inner = []

	def __add__(self, other):
		if isinstance(other, bytes) or isinstance(other, bytearray):
			self.inner.append(bytes(other))
		elif isinstance(other, Ref):
			self.inner.append(other)
		elif isinstance(other, Sym):
			self.inner.append(other)
		elif isinstance(other, ObjectData):
			self.inner.append(other)
		return self

	@staticmethod
	def element_size(el):
		if isinstance(el, bytes):
			return len(el)
		elif isinstance(el, Ref):
			return 4
		elif isinstance(el, Sym):
			return 0
		elif isinstance(el, tuple): # linked sub-ObjectData
			return el[1]
		else:
			raise Exception(f"invalid _element_length: {el}")

	def align(self, size, alignment, elements):
		padding = (alignment - size) % alignment
		if padding != 0:
			elements.append(bytes(padding))
		return padding

	def link(self, symbol):
		inner = []
		outer = []
		elements = []
		size = 0

		# First unfold all structures within [inner] as we accumulate the total
		# size of the inner data
		for el in self.inner:
			if isinstance(el, ObjectData):
				size += self.align(size, el.alignment, inner)
				code, code_size = el.link(f"{symbol} + {size}")
				inner.append((code, code_size))
				size += code_size
			else:
				inner.append(el)
				size += self.element_size(el)

		# Then replace complex references with unfolded data appended at the
		# end of the structure
		for el in inner:
			if isinstance(el, Ref) and el.kind == "bytes":
				elements.append(Ref("name", f"{symbol} + {size}"))
				outer.append(el.target)
				size += self.element_size(el.target)

			elif isinstance(el, Ref) and el.kind == "struct":
				size += self.align(size, el.target.alignment, outer)
				elements.append(Ref("name", f"{symbol} + {size}"))
				code, code_size = el.target.link(f"{symbol} + {size}")
				outer.append((code, code_size))
				size += code_size

			else:
				elements.append(el)

		elements += outer

		# Make sure the whole structure is properly aligned
		size += self.align(size, self.alignment, elements)

		# Finally, generate actual assembler code based on all elements
		asm = ""

		for el in elements:
			if isinstance(el, bytes):
				asm += ".byte " + ",".join(hex(x) for x in el) + "\n"
			elif isinstance(el, Ref) and el.kind == "name":
				asm += f".long {el.target}\n"
			elif isinstance(el, Sym):
				asm += f".global {el.name}\n"
				asm += f"{el.name}:\n"
			elif isinstance(el, tuple): # linked ObjectData
				asm += el[0]

		return asm, size

# User-friendly synonym
Structure = ObjectData

#
#  Area specifications
#

class Area:
	"""
	A subrectangle of an image, typically used for pre-conversion cropping.
	"""

	def __init__(self, area, img):
		"""
		Construct an Area object from a dict specification. The following keys
		may be used to specific the position and size of the rectangle:

		* "x", "y"          (int strings, default to 0)
		* "width", "height" (int strings, default to image dimensions)
		* "size"            ("WxH" where W and H are the width and height)

		The Area objects has attributes "x", "y", "w" and "h". Both positions
		default to 0 and both sizes to the corresponding image dimensions.
		"""

		self.x = int(area.get("x", 0))
		self.y = int(area.get("y", 0))
		self.w = int(area.get("width", img.width))
		self.h = int(area.get("height", img.height))

		if "size" in area:
			self.w, self.h = map(int, area["size"].split("x"))

	def tuple(self):
		"""Return the tuple representation (x,y,w,h), suitable for .crop(). """
		return (self.x, self.y, self.w, self.h)

#
#  Grid specifications
#

class Grid:
	"""
	A grid over an image, used to isolate glyphs in fonts and tiles in maps.
	Supports several types of spacing. To apply an outer border, use crop
	through an Area before using the Grid.
	"""

	def __init__(self, grid):
		"""
		Construct a Grid object from a dict specification. The following keys
		may be used to specify the dimension and spacing of the cells:

		* "border"           (int string, defaults to 0)
		* "padding"          (int string, defaults to 0)
		* "width", "height"  (int strings, mandatory if "size" not set)
		* "size"             ("WxH" where W and H are the cell width/height)

		The Grid object has attributes "border", "padding", "w" and "h". Each
		cell is of size "(w,h)" and has "padding" pixels of proper padding
		around it. Additionally, cells are separated by a border of size
		"border"; this includes an outer border.
		"""

		self.border = int(grid.get("border", 0))
		self.padding = int(grid.get("padding", 0))

		self.w = int(grid.get("width", -1))
		self.h = int(grid.get("height", -1))

		if "size" in grid:
			self.w, self.h = map(int, grid["size"].split("x"))

		if self.w <= 0 or self.h <= 0:
			raise FxconvError("size of grid unspecified or invalid")

	def size(self, img):
		"""Count the number of elements in the grid."""
		b, p, w, h =  self.border, self.padding, self.w, self.h

		# Padding-extended parameters
		W = w + 2 * p
		H = h + 2 * p

		columns = (img.width  - b) // (W + b)
		rows    = (img.height - b) // (H + b)
		return columns * rows


	def iter(self, img):
		"""Yields subrectangles of the grid as tuples (x,y,w,h)."""
		b, p, w, h =  self.border, self.padding, self.w, self.h

		# Padding-extended parameters
		W = w + 2 * p
		H = h + 2 * p

		columns = (img.width  - b) // (W + b)
		rows    = (img.height - b) // (H + b)

		for r in range(rows):
			for c in range(columns):
				x = b + c * (W + b) + p
				y = b + r * (H + b) + p
				yield (x, y, x + w, y + h)

#
# Binary conversion
#

def convert_binary(input, params):
	return open(input, "rb").read()

#
# Image conversion for fx-9860G
#

def convert_bopti_fx(input, params):
	if isinstance(input, Image.Image):
		img = input.copy()
	else:
		img = Image.open(input)
	if img.width >= 4096 or img.height >= 4096:
		raise FxconvError(f"'{input}' is too large (max. 4095x4095)")

	# Expand area.size and get the defaults. Crop image to resulting area.
	area = Area(params.get("area", {}), img)
	img = img.crop(area.tuple())

	# Quantize the image and check the profile
	img = quantize(img, dither=False)

	# If profile is provided, check its validity, otherwise use the smallest
	# compatible profile

	colors = { y for (x,y) in img.getcolors() }

	if "profile" in params:
		name = params["profile"]
		p = FxProfile.find(name)

		if p is None:
			raise FxconvError(f"unknown profile {name} in '{input}'")
		if colors - p.colors:
			raise FxconvError(f"{name} has too few colors for '{input}'")
	else:
		name = "gray" if FX_LIGHT in colors or FX_DARK in colors else "mono"
		if FX_ALPHA in colors: name += "_alpha"
		p = FxProfile.find(name)

	# Make the image header

	header = bytes ([(0x80 if p.gray else 0) + p.id])
	encode24bit = lambda x: bytes([ x >> 16, (x & 0xff00) >> 8, x & 0xff ])
	header += encode24bit((img.size[0] << 12) + img.size[1])

	# Split the image into layers depending on the profile and zip them all

	layers = [ _image_project(img, layer) for layer in p.layers ]
	count  = len(layers)
	size   = len(layers[0])

	data = bytearray(count * size)
	n = 0

	for longword in range(size // 4):
		for layer in layers:
			for i in range(4):
				data[n] = layer[4 * longword + i]
				n += 1

	# Generate the object file
	return header + data

def _image_project(img, f):
	# New width and height
	w = (img.size[0] + 31) // 32
	h = (img.size[1])

	data = bytearray(4 * w * h)
	im = img.load()

	# Now generate a 32-bit byte sequence
	for y in range(img.size[1]):
		for x in range(img.size[0]):
			bit = int(f(im[x, y]))
			data[4 * y * w + (x >> 3)] |= (bit << (~x & 7))

	return data

#
# Image conversion for fx-CG 50
#

def convert_bopti_cg(input, params):
	if isinstance(input, Image.Image):
		img = input.copy()
	else:
		img = Image.open(input)
	if img.width >= 65536 or img.height >= 65536:
		raise FxconvError(f"'{input}' is too large (max. 65535x65535)")

	# Crop image to key "area"
	area = Area(params.get("area", {}), img)
	img = img.crop(area.tuple())

	# If no profile is specified, fall back to rgb565 or rgb565a later on
	name = params.get("profile", None)
	if name is not None:
		profile = CgProfile.find(name)

	if name in ["r5g6b5", "r5g6b5a", "rgb565", "rgb565a", None]:
		# Encode the image into the 16-bit format
		encoded, alpha = r5g6b5(img)

		if name is None:
			name = "rgb565" if alpha is None else "rgb565a"
		profile = CgProfile.find(name)

	elif name.startswith("p"):
		force_alpha = name.endswith("_rgb565a")
		if name.startswith("p8"):
			trim_palette = True
			palette_base = 0x80
			color_count = 256
		elif name.startswith("p4"):
			trim_palette = False
			palette_base = 0x00
			color_count = 16
		else:
			raise FxconvError(f"unknown palette format {profile}")

		# Encode the image into 16-bit with a palette of 16 or 256 entries
		encoded, palette, alpha = r5g6b5(img, color_count=color_count,
			trim_palette=trim_palette, palette_base=palette_base,
			force_alpha=force_alpha)

		color_count = len(palette) // 2
		encoded = palette + encoded

	else:
		raise FxconvError(f"unknown color profile '{name}'")

	# Resolve "p8" and "p4" to their optimal variation
	if name == "p8":
		name = "p8_rgb565" if alpha is None else "p8_rgb565a"
		profile = CgProfile.find(name)
	elif name == "p4":
		name = "p4_rgb565" if alpha is None else "p4_rgb565a"
		profile = CgProfile.find(name)

	if alpha is not None and not profile.supports_alpha:
		raise FxconvError(f"'{input}' has transparency; use rgb565a, " +
			"p8_rgb565a or p4_rgb565a")

	header = bytes()
	header += u16(profile.id)
	header += u16(alpha if alpha is not None else 0xffff)
	header += u16(img.width) + u16(img.height)

	if name in ["p8_rgb565", "p8_rgb565a"]:
		header += u16(color_count)

	if len(encoded) % 4 != 0:
		encoded += bytes(4 - (len(encoded) % 4))
	return header + encoded

#
# Font conversion
#

def _trim(img):
	def blank(x):
		return all(px[x,y] == FX_WHITE for y in range(img.height))

	left = 0
	right = img.width
	px = img.load()

	while left + 1 < right and blank(left):
		left += 1
	while right - 1 > left and blank(right - 1):
		right -= 1

	return img.crop((left, 0, right, img.height))

def _blockstart(name):
	m = re.match(r'(?:U\+)?([0-9A-Fa-f]+)\.', name)

	if m is None:
		return None
	try:
		return int(m[1], base=16)
	except Exception as e:
		return None

def convert_topti(input, params):

	#--
	# Character set
	#--

	if "charset" not in params:
		raise FxconvError("'charset' attribute is required and missing")

	charset = params["charset"]
	blocks = FX_CHARSETS.get(charset, None)
	if blocks is None:
		raise FxconvError(f"unknown character set '{charset}'")

	# Will be recomputed later for Unicode fonts
	glyph_count = sum(length for start, length in blocks)

	#--
	# Image input
	#--

	grid = Grid(params.get("grid", {}))

	# When using predefined charsets with a single image, apply the area and
	# check that the number of glyphs is correct
	if charset != "unicode":
		if isinstance(input, Image.Image):
			img = input.copy()
		else:
			img = Image.open(input)
		area = Area(params.get("area", {}), img)
		img = img.crop(area.tuple())

		# Quantize it (only black pixels will be encoded into glyphs)
		img = quantize(img, dither=False)

		if glyph_count > grid.size(img):
			raise FxconvError(
				f"not enough elements in grid (got {grid.size(img)}, "+
				f"need {glyph_count} for '{charset}')")

		inputs = [ img ]

	# In Unicode mode, load images for the provided directory, but don't apply
	# the area (this makes no sense since the sizes are different)
	else:
		try:
			files = os.listdir(input)
		except Exception as e:
			raise FxconvError(
				f"cannot scan directory '{input}' to discover blocks for the"+
				f"unicode charset: {str(e)}")

		# Keep only files with basenames like "<hexa>" or "U+<hexa>" and sort
		# them by code point order (for consistency)
		files = [e for e in files if _blockstart(e) is not None]
		files = sorted(files, key=_blockstart)

		# Open all images and guess the block size
		inputs = []
		for file in files:
			img = Image.open(os.path.join(input, file))
			img = quantize(img, dither=False)
			inputs.append(img)

		blocks = [(_blockstart(e), grid.size(img))
		          for e, img in zip(files, inputs)]

		# Recompute the total glyph count
		glyph_count = sum(length for start, length in blocks)


	#--
	# Proportionality and metadata
	#--

	proportional = (params.get("proportional", "false") == "true")
	title = bytes(params.get("title", ""), "utf-8") + bytes([0])

	flags = set(params.get("flags", "").split(","))
	flags.remove("")
	flags_std = { "bold", "italic", "serif", "mono" }

	if flags - flags_std:
		raise FxconvError(f"unknown flags: {', '.join(flags - flags_std)}")

	bold   = int("bold" in flags)
	italic = int("italic" in flags)
	serif  = int("serif" in flags)
	mono   = int("mono" in flags)

	flags = (bold << 7) | (italic << 6) | (serif << 5) | (mono << 4) \
			| int(proportional)
	# Default line height to glyph height
	line_height = int(params.get("height", grid.h))

	# Default character spacing to 1
	char_spacing = int(params.get("char-spacing", "1"))

	#--
	# Encoding blocks
	#---

	def encode_block(b):
		start, length = b
		return u32((start << 12) | length)

	data_blocks = b''.join(encode_block(b) for b in blocks)

	#--
	# Encoding glyphs
	#--

	data_glyphs = []
	total_glyphs = 0
	data_width = bytearray()
	data_index  = bytearray()

	for img in inputs:
		for (number, region) in enumerate(grid.iter(img)):
			# Upate index
			if not (number % 8):
				idx = total_glyphs // 4
				data_index += u16(idx)

			# Get glyph area
			glyph = img.crop(region)
			if proportional:
				glyph = _trim(glyph)
				data_width.append(glyph.width)

			length = 4 * ((glyph.width * glyph.height + 31) >> 5)
			bits = bytearray(length)
			offset = 0
			px = glyph.load()

			for y in range(glyph.size[1]):
				for x in range(glyph.size[0]):
					color = (px[x,y] == FX_BLACK)
					bits[offset >> 3] |= ((color * 0x80) >> (offset & 7))
					offset += 1

			data_glyphs.append(bits)
			total_glyphs += length

	data_glyphs = b''.join(data_glyphs)

	#---
	#  Object file generation
	#---

	# Base data: always put the raw data and blocks first since they are
	# 4-aligned, to preserve alignment on the rest of the references.
	o = ObjectData()

	# Make the title pointer NULL if no title is specified
	if len(title) > 1:
		o += ref(title, padding=4)
	else:
		o += u32(0)

	o += u8(flags) + u8(line_height) + u8(grid.h) + u8(len(blocks))
	o += u32(glyph_count)
	o += u8(char_spacing) + bytes(3)
	o += ref(data_blocks)
	o += ref(data_glyphs)

	# For proportional fonts, add the index (2-aligned) then the glyph size
	# array (1-aligned).
	if proportional:
		o += ref(data_index)
		o += ref(data_width)
	# For fixed-width fonts, add more metrics
	else:
		o += u16(grid.w)
		o += u16((grid.w * grid.h + 31) >> 5)

	return o

#
# libimg conversion for fx-9860G
#

def convert_libimg_fx(input, params):
	if isinstance(input, Image.Image):
		img = input.copy()
	else:
		img = Image.open(input)
	if img.width >= 65536 or img.height >= 65536:
		raise FxconvError(f"'{input}' is too large (max. 65535x65535)")

	# Crop image to area
	area = Area(params.get("area", {}), img)
	img = img.crop(area.tuple())

	# Quantize the image. We don't need to check if there is gray; the VRAM
	# rendering function for mono output will adjust at runetime
	img = quantize(img, dither=False)
	code = { FX_WHITE: 0, FX_LIGHT: 1, FX_DARK: 2, FX_BLACK: 3, FX_ALPHA: 4 }

	# Encode image as a plain series of pixels
	data = bytearray(img.width * img.height)
	im = img.load()
	i = 0

	for y in range(img.height):
		for x in range(img.width):
			data[i] = code[im[x, y]]
			i += 1

	o = ObjectData()
	o += u16(img.width) + u16(img.height)
	o += u16(img.width) + u8(LIBIMG_FLAG_RO) + bytes(1)
	o += ref(data)

	return o


#
# libimg conversion for fx-CG 50
#

def convert_libimg_cg(input, params):
	if isinstance(input, Image.Image):
		img = input.copy()
	else:
		img = Image.open(input)
	if img.width >= 65536 or img.height >= 65536:
		raise FxconvError(f"'{input}' is too large (max. 65535x65535)")

	# Crop image to key "area"
	area = Area(params.get("area", {}), img)
	img = img.crop(area.tuple())

	# Encode the image into 16-bit format and force the alpha to 0x0001
	encoded, alpha = r5g6b5(img, force_alpha=(0x0001,0x0000))

	o = ObjectData()
	o += u16(img.width) + u16(img.height)
	o += u16(img.width) + u8(LIBIMG_FLAG_RO) + bytes(1)
	o += ref(encoded)

	return o

#
# Exceptions
#

class FxconvError(Exception):
	pass

#
# API
#

def quantize(img, dither=False):
	"""
	Convert a PIL.Image.Image into an RGBA image with only these colors:
	* FX_BLACK  = (  0,   0,   0, 255)
	* FX_DARK   = ( 85,  85,  85, 255)
	* FX_LIGHT  = (170, 170, 170, 255)
	* FX_WHITE  = (255, 255, 255, 255)
	* FX_ALPHA  = (  0,   0,   0,   0)

	The alpha channel is first flattened to either opaque of full transparent,
	then all colors are quantized into the 4-shade scale. Floyd-Steinberg
	dithering can be used, although most applications will prefer nearest-
	neighbor coloring.

	Arguments:
	img     -- Input image, in any format
	dither  -- Enable Floyd-Steinberg dithering [default: False]

	Returns a quantized PIL.Image.Image.
	"""

	# Our palette will have only 4 colors for the gray engine
	colors = [ FX_BLACK, FX_DARK, FX_LIGHT, FX_WHITE ]

	# Create the palette
	palette = Image.new("RGBA", (len(colors), 1))
	for (i, c) in enumerate(colors):
		palette.putpixel((i, 0), c)
	palette = palette.convert("P")

	# Make the image RGBA in case it was indexed so that transparent pixels are
	# represented in an alpha channel
	if img.mode == "P":
		img = img.convert("RGBA")

	# Save the alpha channel, and make it either full transparent or opaque
	try:
		alpha_channel = img.getchannel("A").convert("1", dither=Image.NONE)
	except:
		alpha_channel = Image.new("L", img.size, 255)

	# Apply the palette to the original image (transparency removed)
	img = img.convert("RGB")

	# Let's do an equivalent of the following, but with a dithering setting:
	# img = img.quantize(palette=palette)

	img.load()
	palette.load()
	im = img.im.convert("P", int(dither), palette.im)
	img = img._new(im).convert("RGB")

	# Put back the alpha channel
	img.putalpha(alpha_channel)

	# Premultiply alpha
	pixels = img.load()
	for y in range(img.size[1]):
		for x in range(img.size[0]):
			r, g, b, a = pixels[x, y]
			if a == 0:
				r, g, b, = 0, 0, 0
			pixels[x, y] = (r, g, b, a)

	return img

def r5g6b5(img, color_count=0, trim_palette=False, palette_base=0,
	force_alpha=None):
	"""
	Convert a PIL.Image.Image into an RGB565 byte stream. If there are
	transparent pixels, chooses a color to implement alpha and replaces them
	with this color.

	When color_count=0, converts the image to 16-bit; returns a bytearray of
	pixel data and the automatically-chosen alpha value (or None).

	* If force_alpha is a pair (value, replacement), then the alpha is forced
	  to be the indicated value, and any natural occurrence of the value is
	  substituted with the replacement.

	When color_count>0, if should be either 16 or 256. The image is encoded
	with a palette of that size. Returns the converted image as a bytearray,
	the palette as a bytearray, and the alpha value (None if there were no
	transparent pixels).

	* If trim_palette is set, the palette bytearray is trimmed so that only
	  used entries are set (this does not include alpha, which is at the end).
	* If palette_base is provided, palette entries will be numbered starting
	  from that value, wrapping around modulo color_count.
	* If force_alpha=True, an index will be reserved for alpha even if no pixel
	  is transparent in the source image.
	"""

	def rgb24to16(rgb24):
		r, g, b = rgb24
		r = (r & 0xff) >> 3
		g = (g & 0xff) >> 2
		b = (b & 0xff) >> 3
		return (r << 11) | (g << 5) | b

	#---
	# Initial image transforms
	# Separate the alpha channel and generate a first palette.
	#---

	# Save the alpha channel and make it 1-bit. If there are transparent
	# pixels, set has_alpha=True and record the alpha channel in alpha_pixels.
	try:
		alpha_channel = img.getchannel("A").convert("1", dither=Image.NONE)
		alpha_levels = { t[1]: t[0] for t in alpha_channel.getcolors() }
		has_alpha = 0 in alpha_levels
		replacement = None

		if has_alpha:
			alpha_pixels = alpha_channel.load()

	except ValueError:
		has_alpha = False

	# Convert the input image to RGB
	img = img.convert("RGB")

	# Transparent pixels also have values on the RGB channels, so they use up a
	# palette entry (in indexed mode) or possible alpha value (in 16-bit mode)
	# even though their color is unused. Replace them with a non-transparent
	# color used elsewhere in the image to avoid that.
	if has_alpha:
		nontransparent_pixels = { (x,y)
			for x in range(img.width) for y in range(img.height)
			if alpha_pixels[x,y] > 0 }

		if nontransparent_pixels:
			x0, y0 = nontransparent_pixels.pop()
			pixels = img.load()

			for y in range(img.height):
				for x in range(img.width):
					if alpha_pixels[x,y] == 0:
						pixels[x,y] = pixels[x0,y0]

	# In indexed mode, generate a specific palette
	if color_count:
		palette_max_size = color_count - int(has_alpha or force_alpha == True)
		img = img.convert("P",
			dither=Image.NONE,
			palette=Image.ADAPTIVE,
			colors=palette_max_size)

		# Format for the first palette is a list of N triplets where N is the
		# number of used opaque colors; obviously N <= palette_max_size
		# Note: sometimes colors are not numbered 0..N-1, so we take the max
		# value rather than len(img.getcolors()); we don't try to remap indices
		pixels = img.load()
		N = 1 + max(pixels[x,y]
		            for y in range(img.height)
		            for x in range(img.width))
		palette1 = img.getpalette()[:3*N]
		palette1 = list(zip(palette1[::3], palette1[1::3], palette1[2::3]))
	else:
		pixels = img.load()

	#---
	# Alpha encoding
	# Find a value to encode transparency and map it into palettes.
	#---

	# RGB565A with fixed alpha value (fi. alpha=0x0001 in libimg)
	if color_count == 0 and force_alpha is not None:
		alpha, replacement = force_alpha

	# For palettes, anything works; use the next value
	elif color_count > 0 and (has_alpha or force_alpha == True):
		alpha = N

	# Find an unused RGB565 value and keep the encoding to 16-bit
	elif has_alpha:
		colormap = { rgb24to16(pixels[x, y])
			for x in range(img.width) for y in range(img.height)
			if alpha_pixels[x, y] > 0 }

		available = set(range(65536)) - colormap

		if not available:
			raise FxconvError("image uses all 65536 colors and alpha")
		alpha = available.pop()

	# No transparency in the image
	else:
		alpha = None

	# Function to encode with alpha support in RGB565
	def alpha_encoding(color, a):
		if a > 0: # pixel is not transparent
			return color if color != alpha else replacement
		else:
			return alpha

	# In palette formats, rearrange the palette to account for modulo numbering
	# and trim the palette if needed

	if color_count > 0:
		total = len(palette1) + int(has_alpha or force_alpha == True)
		# The palette_map list associates to indices into palette1, the pixel
		# value in the data array (which starts at palette_base)
		palette_map = [(palette_base + i) % color_count for i in range(total)]

	#---
	# Image encoding
	# Create byte arrays with pixel data and palette data
	#---

	stride = 0

	if not color_count:
		# In RGB16, preserve alignment between rows
		stride = (img.width + 1) // 2 * 2
		size = stride * img.height * 2
	elif color_count == 256:
		# No constraint in P8
		size = img.width * img.height
		stride = img.width
	elif color_count == 16:
		# In P4, pad whole bytes
		stride = (img.width + 1) // 2
		size = stride * img.height

	# Result of encoding
	encoded = bytearray(size)
	# Offset into the array
	offset = 0

	for y in range(img.height):
		for x in range(img.width):
			a = alpha_pixels[x, y] if has_alpha else 255

			if not color_count:
				c = alpha_encoding(rgb24to16(pixels[x, y]), a)
				offset = (stride * y + x) * 2
				encoded[offset:offset+2] = u16(c)

			elif color_count == 16:
				c = palette_map[pixels[x, y] if a > 0 else alpha]

				# Select either the 4 MSb or 4 LSb of the current byte
				if x % 2 == 0:
					encoded[offset] |= (c << 4)
				else:
					encoded[offset] |= c

				offset += (x % 2 == 1) or (x == img.width - 1)

			elif color_count == 256:
				c = palette_map[pixels[x, y] if a > 0 else alpha]
				encoded[offset] = c
				offset += 1

	# Encode the palette as R5G6B5

	if color_count > 0:
		if trim_palette:
			encoded_palette = bytearray(2 * len(palette_map))
		else:
			encoded_palette = bytearray(2 * color_count)

		for i in range(len(palette1)):
			encoded_palette[2*i:2*i+2] = u16(rgb24to16(palette1[i]))

	#---
	# Outro
	#---

	if color_count > 0:
		if alpha is not None:
			alpha = palette_map[alpha]
		return encoded, encoded_palette, alpha
	else:
		return encoded, alpha

def convert(input, params, target, output=None, model=None, custom=None):
	"""
	Convert a data file into an object that exports the following symbols:
	* _<varname>
	* _<varname>_end
	* _<varname>_size
	The variable name is obtained from the parameter dictionary <params>.

	Arguments:
	input   -- Input file path
	params  -- Parameter dictionary
	target  -- String dictionary keys 'toolchain', 'arch' and 'section'
	output  -- Output file name [default: <input> with suffix '.o']
	model   -- 'fx' or 'cg'  (some conversions require this) [default: None]
	custom  -- Custom conversion function

	Produces an output file and returns nothing.
	"""

	if output is None:
		output = os.path.splitext(input)[0] + ".o"

	if "name" not in params:
		raise FxconvError(f"no name specified for conversion '{input}'")

	if target["arch"] is None:
		target["arch"] = model

	if "custom-type" in params:
		t = params["custom-type"]
		# Also copy it in "type" for older converters (this is legacy)
		params["type"] = t
	elif "type" in params:
		t = params["type"]
	else:
		raise FxconvError(f"missing type in conversion '{input}'")

	if t == "binary":
		o = convert_binary(input, params)
	elif t == "bopti-image" and model in [ "fx", None ]:
		o = convert_bopti_fx(input, params)
	elif t == "bopti-image" and model == "cg":
		o = convert_bopti_cg(input, params)
	elif t == "font":
		o = convert_topti(input, params)
	elif t == "libimg-image" and model in [ "fx", None ]:
		o = convert_libimg_fx(input, params)
	elif t == "libimg-image" and model == "cg":
		o = convert_libimg_cg(input, params)
	elif custom is not None:
		for converter in custom:
			if converter(input, output, params, target) == 0:
				return
		raise FxconvError(f'unknown custom resource type \'{t}\'')
	else:
		raise FxconvError(f'unknown resource type \'{t}\'')

	# Standard conversions: save to ELF in the natural way
	elf(o, output, "_" + params["name"], **target)

def elf(data, output, symbol, toolchain=None, arch=None, section=None,
	assembly=None):
	"""
	Call objcopy to create an object file from the specified data. The object
	file will export three symbols:
	* <symbol>
	* <symbol>_end
	* <symbol>_size

	The symbol name must have a leading underscore if it is to be declared and
	used from a C program.

    The toolchain can be any target triplet for which the compiler is
    available. The architecture is deduced from some typical triplets;
    otherwise it can be set, usually as "sh3" or "sh4-nofpu". This affects the
	--binary-architecture flag of objcopy. If arch is set to "fx" or "cg", this
	function tries to be smart and:

	* Uses the name of the compiler if it contains a full architecture name
	  such as "sh3", "sh4" or "sh4-nofpu";
	* Uses "sh3" for fx9860g and "sh4-nofpu" for fxcg50 if the toolchain is
	  "sh-elf", which is a custom set;
	* Fails otherwise.

	The section name can be specified, along with its flags. A typical example
	would be section=".rodata,contents,alloc,load,readonly,data", which is the
	default.

	If assembly is set to a non-empty assembly program, this function also
	generates a temporary ELF file by assembling this piece of code, and merges
	it into the original one.

	Arguments:
	data       -- A bytes-like or ObjectData object to embed into the output
	output     -- Name of output file
	symbol     -- Chosen symbol name
	toolchain  -- Target triplet [default: "sh3eb-elf"]
	arch       -- Target architecture [default: try to guess]
	section    -- Target section [default: above variation of .rodata]
	assembly   -- Additional assembly code [default: None]

	Produces an output file and returns nothing.
	"""

	# Unfold ObjectData into data and assembly
	if isinstance(data, ObjectData):
		asm =   ".section .rodata\n"
		asm += f".global {symbol}\n"
		asm += f"{symbol}:\n"
		asm += data.link(symbol)[0]
		asm += (assembly or "")

		data = None
		assembly = asm

	if data is None and assembly is None:
		raise FxconvError("elf() but no data and no assembly")

	# Toolchain parameters

	if toolchain is None:
		toolchain = "sh3eb-elf"
	if section is None:
		section = ".rodata,contents,alloc,load,readonly,data"

	if arch in ["fx", "cg", None] and toolchain in ["sh3eb-elf", "sh4eb-elf",
		"sh4eb-nofpu-elf"]:
		arch = toolchain.replace("eb-", "-")[:-4]

	elif arch == "fx" and toolchain == "sh-elf":
		arch = "sh3"
	elif arch == "cg" and toolchain == "sh-elf":
		arch = "sh4-nofpu"

	elif arch in ["fx", "cg", None]:
		raise FxconvError(f"non-trivial architecture for {toolchain} must be "+
			"specified")

	# Generate data - in <output> directly if there is no assembly

	if data:
		fp_obj = tempfile.NamedTemporaryFile()
		fp_obj.write(data)
		fp_obj.flush()

		sybl = "_binary_" + fp_obj.name.replace("/", "_")
		objcopy_args = [
			f"{toolchain}-objcopy", "-I", "binary", "-O", "elf32-sh",
			"--binary-architecture", arch, "--file-alignment", "4",
			"--rename-section", f".data={section}",
			"--redefine-sym", f"{sybl}_start={symbol}",
			"--redefine-sym", f"{sybl}_end={symbol}_end",
			"--redefine-sym", f"{sybl}_size={symbol}_size",
			fp_obj.name, output if not assembly else fp_obj.name + ".o" ]

		proc = subprocess.run(objcopy_args)
		if proc.returncode != 0:
			raise FxconvError(f"objcopy returned {proc.returncode}")

	# Generate assembly - in <output> directly if there is no data

	if assembly:
		fp_asm = tempfile.NamedTemporaryFile()
		fp_asm.write(assembly.encode('utf-8'))
		fp_asm.flush()

		as_args = [
			f"{toolchain}-as", "-c", fp_asm.name, "-o",
			output if not data else fp_asm.name + ".o" ]

		proc = subprocess.run(as_args)
		if proc.returncode != 0:
			raise FxconvError(f"as returned {proc.returncode}")

	# If both data and assembly are specified, merge everyone

	if data and assembly:
		ld_args = [
			f"{toolchain}-ld", "-r", fp_obj.name + ".o", fp_asm.name + ".o",
			"-o", output ]

		proc = subprocess.run(ld_args)
		if proc.returncode != 0:
			raise FxconvError(f"ld returned {proc.returncode}")

		os.remove(fp_obj.name + ".o")
		os.remove(fp_asm.name + ".o")

	if data:
		fp_obj.close()
	if assembly:
		fp_asm.close()

#
# Meta API
#

def parse_parameters(params):
	"""Parse parameters of the form "NAME:VALUE" into a dictionary."""
	d = dict()

	def insert(d, path, value):
		if len(path) == 1:
			d[path[0]] = value
		else:
			if not path[0] in d:
				d[path[0]] = dict()
			insert(d[path[0]], path[1:], value)

	for decl in params:
		if ":" not in decl:
			raise FxconvError(f"invalid parameter {decl}, ignoring")
		else:
			name, value = decl.split(":", 1)
			value = value.strip()
			if name == "name_regex":
				value = value.split(" ", 1)
			insert(d, name.split("."), value)

	return d

def _parse_metadata(contents):
	"""
	Parse the contents of an fxconv-metadata.txt file. Comments start with '#'
	anywhere on a line and extend to the end of the line.

	The file is divided in blocks that start with a "<wildcard>:" pattern at
	the first column of a line (no leading spaces) followed by zero or more
	properties declared as "key: value" (with at least one leading space).

	The key can contain dots (eg. "category.field"), in which case the value
	for the main component ("category") is itself a dictionary.
	"""

	RE_COMMENT = re.compile(r'#.*$', re.MULTILINE)
	contents = re.sub(RE_COMMENT, "", contents)

	RE_WILDCARD = re.compile(r'^(\S(?:[^:\s]|\\:|\\ )*)\s*:\s*$', re.MULTILINE)
	lead, *elements = [ s.strip() for s in re.split(RE_WILDCARD, contents) ]

	if lead:
		raise FxconvError(f"invalid metadata: {lead} appears before wildcard")

	# Group elements by pairs (left: wildcard, right: list of properties)
	elements = list(zip(elements[::2], elements[1::2]))

	metadata = []
	for (wildcard, params) in elements:
		params = [ s.strip() for s in params.split("\n") if s.strip() ]
		metadata.append((wildcard, parse_parameters(params)))

	return metadata

class Metadata:
	def __init__(self, path=None, text=None):
		"""
		Load either an fxconv-metadata.txt file (if path is not None) or the
		contents of such a file (if text is not None).
		"""

		if (path is not None) == (text is not None):
			raise ValueError("Metadata must have exactly one of path and text")

		if path is not None:
			self._path = path
			with open(path, "r") as fp:
				self._rules = _parse_metadata(fp.read())
		elif text is not None:
			self._path = None
			self._rules = _parse_metadata(text)

	def path(self):
		"""
		Returns the path of the file from which the metadata was parsed, or
		None if the metadata was parsed from string.
		"""
		return self._path

	def rules(self):
		"""
		Returns a list of pairs (wildcard, rules) where the wildcard is a
		string and the rules are a nested dictionary.
		"""
		return self._rules

	def rules_for(self, path):
		"""
		Returns the parameters that apply to the specified path, or None if no
		wildcard matches it. The special key "name_regex" is also resolved into
		the regular key "name".
		"""

		basename = os.path.basename(path)
		params = dict()
		matched = False

		for (wildcard, p) in self._rules:
			if fnmatch.fnmatchcase(basename, wildcard):
				params.update(**p)
				matched = True

		if not matched:
			return None

		if "name_regex" in params and not "name" in params:
			params["name"] = re.sub(*params["name_regex"], basename)
		return params