Checkpoint WIP for easier comparison to dhgr branch:

- naive version of NTSC artifacting, it uses a sliding 4-bit window to
  assign a nominal (D)HGR colour to each dot position.  A more
  sophisticated/correct implementation would model the YIQ signal
  directly.

- Switch DHGRBitmap implementation to use a 34-bit representation of
  the 4-byte tuple, comprised of a 3-bit header and footer, plus
  4*7=28-bit body.  The headers/footers account for the influence on
  neighbouring tuples from the 4-bit NTSC window.

- With this model each screen byte influences 13 pixels, so we need to
  precompute 2^26 edit distances for all possible (source, target)
  13-bit sequences.

- Checkpointing not-yet-working HGR implementation.

- Add new unit tests but not yet all passing due to refactoring
This commit is contained in:
kris
2019-07-02 22:40:50 +01:00
parent e2a8bd9b4d
commit 666272a8fc
9 changed files with 1268 additions and 367 deletions

View File

@@ -3,10 +3,11 @@
import bz2
import functools
import pickle
from typing import Union, List
from typing import Union, List, Optional
import numpy as np
import palette
import palette as pal
# Type annotation for cases where we may process either an int or a numpy array.
IntOrArray = Union[int, np.ndarray]
@@ -124,55 +125,334 @@ class MemoryMap:
self.page_offset[page - self._page_start][offset] = val
class DHGRBitmap:
BYTE_MASK32 = [
# 3333333222222211111110000000 <- byte 0.3
#
# 33222222222211111111110000000000 <- bit pos in uint32
# 10987654321098765432109876543210
# 0000GGGGFFFFEEEEDDDDCCCCBBBBAAAA <- pixel A..G
# 3210321032103210321032103210 <- bit pos in A..G pixel
0b00000000000000000000000011111111, # byte 0 influences A,B
0b00000000000000001111111111110000, # byte 1 influences B,C,D
0b00000000111111111111000000000000, # byte 2 influences D,E,F
0b00001111111100000000000000000000, # byte 3 influences F,G
]
@functools.lru_cache(None)
def _edit_distances(name: str, palette_id: pal.Palette) -> List[np.ndarray]:
"""Load edit distance matrices for masked, shifted byte values.
# How much to right-shift bits after masking to bring into int8/int12 range
BYTE_SHIFTS = [0, 4, 12, 20]
This is defined at module level to be a singleton.
"""
data = "transcoder/data/%s_palette_%d_edit_distance.pickle.bz2" % (
name,
palette_id.value
)
with bz2.open(data, "rb") as ed:
return pickle.load(ed) # type: List[np.ndarray]
@staticmethod
@functools.lru_cache(None)
def edit_distances(palette_id: palette.Palette) -> List[np.ndarray]:
"""Load edit distance matrices for masked, shifted byte 0..3 values."""
data = "transcoder/data/palette_%d_edit_distance.pickle.bz2" % (
palette_id.value
)
with bz2.open(data, "rb") as ed:
return pickle.load(ed) # type: List[np.ndarray]
def __init__(self, main_memory: MemoryMap, aux_memory: MemoryMap):
self.main_memory = main_memory
self.aux_memory = aux_memory
class Bitmap:
"""Packed 28-bit bitmap representation of (D)HGR screen memory.
self.packed = np.empty(shape=(32, 128), dtype=np.uint32)
The memory layout is still page-oriented, not linear y-x buffer but the
bit map is such that 20 consecutive entries linearly encode the 28*20 =
560-bit monochrome dot positions that underlie both Mono and Colour (
D)HGR screens.
For Colour display the (nominal) colours are encoded as 4-bit pixels.
"""
def __init__(
self,
palette: pal.Palette,
main_memory: MemoryMap,
aux_memory: Optional[MemoryMap]
):
self.palette = palette # type: pal.Palette
self.main_memory = main_memory # type: MemoryMap
self.aux_memory = aux_memory # type: Optional[MemoryMap]
self.packed = np.empty(
shape=(32, 128), dtype=np.uint64) # type: np.ndarray
self._pack()
def _pack(self) -> None:
"""Interleave and pack aux and main memory into 28-bit uint32 array"""
"""Pack MemoryMap into 34-bit representation."""
raise NotImplementedError
NAME = None
@functools.lru_cache(None)
def edit_distances(self, palette_id: pal.Palette) -> List[np.ndarray]:
"""Load edit distance matrices for masked, shifted byte values."""
return _edit_distances(self.NAME, palette_id)
def apply(
self,
page: int,
offset: np.uint8,
is_aux: bool,
value: np.uint8) -> None:
raise NotImplementedError
@functools.lru_cache(None)
def byte_pair_difference(
self,
byte_offset: int,
old_packed: int,
content: int
) -> int:
raise NotImplementedError
def diff_weights(
self,
other: "DHGRBitmap",
is_aux: bool
) -> np.ndarray:
raise NotImplementedError
def compute_delta(
self,
content: int,
old: np.ndarray,
is_aux: bool
) -> np.ndarray:
raise NotImplementedError
class HGRBitmap(Bitmap):
BYTE_MASK16 = [
# 11111110000000 <- byte 0, 1
# 1111110000000000
# 5432109876543210
# 00GGFFEEDDCCBBAA <- pixel A..G
0b0000000011111111,
0b0011111111000000
]
# Representation
#
# 1111110000000000
# 5432109876543210
# PGGFFEEDPDCCBBAA
#
# Where palette bit influences all of the pixels in the byte
#
# Map to 3-bit pixels, i.e. 21-bit quantity
#
# 222211111111110000000000
# 321098765432109876543210
# 000PGGPFFPEEPDDPCCPBBPAA
BYTE_MASK32 = [
0b000000000000111111111111,
0b000111111111111000000000
]
# XXX 3-bit pixel isn't quite correct, e.g. the case of conflicting
# palette bits across byte boundary
# Also hard to interleave the palette bit in multiple places - could use
# a mapping array but maybe don't need to, can just use 8-bit values as is?
# But need contiguous representation for edit distance tables
# P
# (0)00 --> 0.0.
# (0)01 --> 0.1.
#
# (1)01 --> .0.1
# (1)11 --> .1.1
# etc
#
BYTE_SHIFTS = [0, 9]
NAME = 'HGR'
def __init__(self, palette: pal.Palette, main_memory: MemoryMap):
super(HGRBitmap, self).__init__(palette, main_memory, None)
def _pack(self) -> None:
"""Pack main memory into (28+3)-bit uint64 array"""
# 00000000001111111111222222222233
# 01234567890123456789012345678901
# AAAABBBBCCCCDDd
# AAAABBBBCCCCDd
# DDEEEEFFFFGGGGg
# dDDEEEEFFFFGGGg
# Even, P0: store unshifted (0..14)
# Even, P1: store shifted << 1 (1..15) (only need 1..14)
# Odd, P0: store shifted << 14 (14 .. 28) - set bit 14 as bit 0 of next
# byte
# Odd, p1: store shifted << 15 (15 .. 29) (only need 15 .. 28) - set
# bit 13 as bit 0 of next byte
# Odd overflow only matters for even, P1
# - bit 0 is either bit 14 if odd, P0 or bit 13 if odd, P1
# - but these both come from the undoubled bit 6.
main = self.main_memory.page_offset.astype(np.uint64)
# Double 7-bit pixel data from a into 14-bit fat pixels, and extend MSB
# into 15-bits tohandle case when subsequent byte has palette bit set,
# i.e. is right-shifted by 1 dot. This only matters for even bytes
# with P=0 that are followed by odd bytes with P=1; in other cases
# this extra bit will be overwritten.
double = (
# Bit pos 6
((main & 0x40) << 8) + ((main & 0x40) << 7) + (
(main & 0x40) << 6)) + (
# Bit pos 5
((main & 0x20) << 6) + ((main & 0x20) << 5)) + (
# Bit pos 4
((main & 0x10) << 5) + ((main & 0x10) << 4)) + (
# Bit pos 3
((main & 0x08) << 4) + ((main & 0x08) << 3)) + (
# Bit pos 2
((main & 0x04) << 3) + ((main & 0x04) << 2)) + (
# Bit pos 1
((main & 0x02) << 2) + ((main & 0x02) << 1)) + (
# Bit pos 0
((main & 0x01) << 1) + (main & 0x01))
a_even = main[:, ::2]
a_odd = main[:, 1::2]
double_even = double[:, ::2]
double_odd = double[:, 1::2]
# Place even offsets at bits 1..15 (P=1) or 0..14 (P=0)
packed = np.where(a_even & 0x80, double_even << 1, double_even)
# Place off offsets at bits 15..27 (P=1) or 14..27 (P=0)
packed = np.where(
a_odd & 0x80,
np.bitwise_xor(
np.bitwise_and(packed, (2 ** 15 - 1)),
double_odd << 15
),
np.bitwise_xor(
np.bitwise_and(packed, (2 ** 14 - 1)),
double_odd << 14
)
)
# Patch up even offsets with P=1 with extended bit from previous odd
# column
previous_odd = np.roll(a_odd, 1, axis=1).astype(np.uint64)
packed = np.where(
a_even & 0x80,
# Truncate to 28-bits and set bit 0 from bit 6 of previous byte
np.bitwise_xor(
np.bitwise_and(packed, (2 ** 28 - 2)),
(previous_odd & (1 << 6)) >> 6
),
# Truncate to 28-bits
np.bitwise_and(packed, (2 ** 28 - 1))
)
# Append first 3 bits of next even byte so we can correctly
# decode the effective colours at the end of the 28-bit tuple
trailing = np.roll(packed, -1, axis=1).astype(np.uint64)
packed = np.bitwise_xor(
packed,
(trailing & 0b111) << 28
)
self.packed = packed
@staticmethod
@functools.lru_cache(None)
def byte_offset(x_byte: int) -> int:
"""Returns 0..1 offset in ByteTuple for a given x_byte,"""
is_odd = x_byte % 2 == 1
return 1 if is_odd else 0
@staticmethod
def masked_update(
byte_offset: int,
old_value: IntOrArray,
new_value: int) -> IntOrArray:
raise NotImplementedError
def apply(self, page: int, offset: int, is_aux: bool, value: int) -> None:
"""Update packed representation of changing main/aux memory."""
assert not is_aux
# XXX fix
byte_offset = self.byte_offset(offset)
packed_offset = offset // 2
self.packed[page, packed_offset] = self.masked_update(
byte_offset, self.packed[page, packed_offset], value)
# XXXX Generic?
def mask_and_shift_data(
self,
data: IntOrArray,
byte_offset: int) -> IntOrArray:
"""Masks and shifts data into the 8 or 12-bit range."""
return (data & self.BYTE_MASK32[byte_offset]) >> (
self.BYTE_SHIFTS[byte_offset])
class DHGRBitmap(Bitmap):
# NOTE: See https://github.com/numpy/numpy/issues/2524 and related issues
# for why we have to cast things explicitly to np.uint64 - type promotion
# to uint64 is broken in numpy :(
# 3-bit header + 28-bit body + 3-bit trailer
BYTE_MASK34 = [
# 3333333222222211111110000000 <- byte 0.3
#
# 3333222222222211111111110000000000 <- bit pos in uint64
# 3210987654321098765432109876543210
# tttGGGGFFFFEEEEDDDDCCCCBBBBAAAAhhh <- pixel A..G
# 3210321032103210321032103210 <- bit pos in A..G pixel
np.uint64(0b0000000000000000000001111111111111), # byte 0 int13 mask
np.uint64(0b0000000000000011111111111110000000), # byte 1 int13 mask
np.uint64(0b0000000111111111111100000000000000), # byte 2 int13 mask
np.uint64(0b1111111111111000000000000000000000), # byte 3 int13 mask
]
# How much to right-shift bits after masking to bring into int13 range
BYTE_SHIFTS = [np.uint64(0), np.uint64(7), np.uint64(14), np.uint64(21)]
NAME = 'DHGR'
def _pack(self) -> None:
"""Interleave and pack aux and main memory into 34-bit uint64 array"""
# Palette bit is unused for DHGR so mask it out
aux = (self.aux_memory.page_offset & 0x7f).astype(np.uint32)
main = (self.main_memory.page_offset & 0x7f).astype(np.uint32)
aux = (self.aux_memory.page_offset & 0x7f).astype(np.uint64)
main = (self.main_memory.page_offset & 0x7f).astype(np.uint64)
# Interleave aux and main memory columns and pack 7-bit masked values
# into a 28-bit value. This sequentially encodes 7 4-bit DHGR pixels.
# into a 28-bit value, with 3-bit header and trailer. This
# sequentially encodes 7 4-bit DHGR pixels, together with the
# neighbouring 3 bits that are necessary to decode artifact colours.
#
# See make_data_tables.py for more discussion about this representation.
self.packed = (
aux[:, 0::2] +
(main[:, 0::2] << 7) +
(aux[:, 1::2] << 14) +
(main[:, 1::2] << 21)
packed = (
(aux[:, 0::2] << 3) +
(main[:, 0::2] << 10) +
(aux[:, 1::2] << 17) +
(main[:, 1::2] << 24)
)
# Prepend last 3 bits of previous main odd byte so we can correctly
# decode the effective colours at the beginning of the 28-bit
# tuple
prevcol = np.roll(packed, 1, axis=1).astype(np.uint64)
# Append first 3 bits of next aux even byte so we can correctly
# decode the effective colours at the end of the 28-bit tuple
nextcol = np.roll(packed, -1, axis=1).astype(np.uint64)
self.packed = np.bitwise_xor(
np.bitwise_xor(
packed,
# Prepend last 3 bits of 28-bit body from previous column
(prevcol & (0b111 << 28)) >> 28
),
# Append first 3 bits of 28-bit body from next column
(nextcol & (0b111 << 3)) << 28
)
@staticmethod
@@ -190,31 +470,207 @@ class DHGRBitmap:
else:
return 1
# XXX test
@staticmethod
def masked_update(
def masked_update_scalar(
byte_offset: int,
old_value: IntOrArray,
new_value: int) -> IntOrArray:
old_value: np.uint64,
new_value: np.uint8) -> np.uint64:
# Mask out 7-bit value where update will go
masked_value = old_value & ~(0x7f << (7 * byte_offset))
masked_value = old_value & (
~np.uint64(0x7f << (7 * byte_offset + 3)))
update = (new_value & 0x7f) << (7 * byte_offset)
update = (new_value & np.uint64(0x7f)) << np.uint64(
7 * byte_offset + 3)
return masked_value ^ update
new = masked_value ^ update
return new
def apply(self, page: int, offset: int, is_aux: bool, value: int) -> None:
# XXX test
@staticmethod
def masked_update_array(
byte_offset: int,
old_value: np.ndarray,
new_value: int) -> np.ndarray:
# Mask out 7-bit value where update will go
masked_value = old_value & (
~np.uint64(0x7f << (7 * byte_offset + 3)))
update = (new_value & np.uint64(0x7f)) << np.uint64(7 * byte_offset + 3)
new = masked_value ^ update
# TODO: don't leak headers across screen rows.
if byte_offset == 0:
# Need to also update the 3-bit trailer of the preceding column
shifted = np.roll(new, -1, axis=1)
new &= np.uint64(2 ** 31 - 1)
new ^= (shifted & np.uint64(0b111 << 3)) << np.uint64(28)
elif byte_offset == 3:
# Need to also update the 3-bit header of the next column
shifted = np.roll(new, 1, axis=1)
new &= np.uint64((2 ** 31 - 1) << 3)
new ^= (shifted & np.uint64(0b111 << 28)) >> np.uint64(28)
return new
# XXX test
def apply(
self,
page: int,
offset: int,
is_aux: bool,
value: np.uint8) -> None:
"""Update packed representation of changing main/aux memory."""
byte_offset = self.interleaved_byte_offset(offset, is_aux)
packed_offset = offset // 2
self.packed[page, packed_offset] = self.masked_update(
self.packed[page, packed_offset] = self.masked_update_scalar(
byte_offset, self.packed[page, packed_offset], value)
# TODO: don't leak headers/trailers across screen rows.
if byte_offset == 0 and packed_offset > 0:
# Need to also update the 3-bit trailer of the preceding column
self.packed[page, packed_offset - 1] &= np.uint64(2 ** 31 - 1)
self.packed[page, packed_offset - 1] ^= (
(self.packed[page, packed_offset] & np.uint64(0b111 << 3))
<< np.uint64(28)
)
elif byte_offset == 3 and packed_offset < 127:
# Need to also update the 3-bit header of the next column
self.packed[page, packed_offset + 1] &= np.uint64(
(2 ** 31 - 1) << 3)
self.packed[page, packed_offset + 1] ^= (
(self.packed[page, packed_offset] & np.uint64(0b111 << 28))
>> np.uint64(28)
)
def mask_and_shift_data(
self,
data: IntOrArray,
byte_offset: int) -> IntOrArray:
"""Masks and shifts data into the 8 or 12-bit range."""
return (data & self.BYTE_MASK32[byte_offset]) >> (
"""Masks and shifts data into the 13-bit range."""
res = (data & self.BYTE_MASK34[byte_offset]) >> (
self.BYTE_SHIFTS[byte_offset])
assert np.all(res <= 2 ** 13)
return res
@functools.lru_cache(None)
def byte_pair_difference(
self,
byte_offset: int,
old_packed: np.uint64,
content: np.uint8
) -> int:
old_pixels = self.mask_and_shift_data(
old_packed, byte_offset)
new_pixels = self.mask_and_shift_data(
self.masked_update_scalar(
byte_offset, old_packed, content), byte_offset)
pair = (old_pixels << np.uint64(13)) + new_pixels
return self.edit_distances(self.palette)[byte_offset][pair]
def diff_weights(
self,
source: "DHGRBitmap",
is_aux: bool
) -> np.ndarray:
return self._diff_weights(source.packed, is_aux)
def _diff_weights(
self,
source_packed: np.ndarray,
is_aux: bool
) -> np.ndarray:
"""Computes diff from source_packed to self.packed"""
diff = np.ndarray((32, 256), dtype=np.int)
if is_aux:
offsets = [0, 2]
else:
offsets = [1, 3]
dists = []
for o in offsets:
# Pixels influenced by byte offset o
source_pixels = self.mask_and_shift_data(source_packed, o)
target_pixels = self.mask_and_shift_data(self.packed, o)
# Concatenate 13-bit source and target into 26-bit values
pair = (source_pixels << np.uint64(13)) + target_pixels
dist = self.edit_distances(self.palette)[o][pair].reshape(
pair.shape)
dists.append(dist)
diff[:, 0::2] = dists[0]
diff[:, 1::2] = dists[1]
return diff
def compute_delta(
self,
content: int,
old: np.ndarray,
is_aux: bool
) -> np.ndarray:
# TODO: use error edit distance
# XXX reuse code
diff = np.ndarray((32, 256), dtype=np.int)
if is_aux:
# Pixels influenced by byte offset 0
source_pixels0 = self.mask_and_shift_data(
self.masked_update_array(0, self.packed, content), 0)
target_pixels0 = self.mask_and_shift_data(self.packed, 0)
# Concatenate 13-bit source and target into 26-bit values
pair0 = (source_pixels0 << np.uint64(13)) + target_pixels0
dist0 = self.edit_distances(self.palette)[0][pair0].reshape(
pair0.shape)
# Pixels influenced by byte offset 2
source_pixels2 = self.mask_and_shift_data(
self.masked_update_array(2, self.packed, content), 2)
target_pixels2 = self.mask_and_shift_data(self.packed, 2)
# Concatenate 13-bit source and target into 26-bit values
pair2 = (source_pixels2 << np.uint64(13)) + target_pixels2
dist2 = self.edit_distances(self.palette)[2][pair2].reshape(
pair2.shape)
diff[:, 0::2] = dist0
diff[:, 1::2] = dist2
else:
# Pixels influenced by byte offset 1
source_pixels1 = self.mask_and_shift_data(
self.masked_update_array(1, self.packed, content), 1)
target_pixels1 = self.mask_and_shift_data(self.packed, 1)
pair1 = (source_pixels1 << np.uint64(13)) + target_pixels1
dist1 = self.edit_distances(self.palette)[1][pair1].reshape(
pair1.shape)
# Pixels influenced by byte offset 3
source_pixels3 = self.mask_and_shift_data(
self.masked_update_array(3, self.packed, content), 3)
target_pixels3 = self.mask_and_shift_data(self.packed, 3)
pair3 = (source_pixels3 << np.uint64(13)) + target_pixels3
dist3 = self.edit_distances(self.palette)[3][pair3].reshape(
pair3.shape)
diff[:, 0::2] = dist1
diff[:, 1::2] = dist3
# TODO: try different weightings
return (diff * 5) - old