def h_analysis(state, data):
"""
Horizontal-only analysis, inverse of h_synthesis (15.4.2).
Returns a tuple (L_data, H_data)
"""
# Bit shift, if required
shift = filter_bit_shift(state)
if shift > 0:
for y in range(0, height(data)):
for x in range(0, width(data)):
data[y][x] = data[y][x] << shift
# Analysis
for y in range(0, height(data)):
oned_analysis(row(data, y), state["wavelet_index_ho"])
# De-interleave the transform data
L_data = new_array(height(data), width(data) // 2)
H_data = new_array(height(data), width(data) // 2)
for y in range(0, (height(data))):
for x in range(0, (width(data) // 2)):
L_data[y][x] = data[y][2 * x]
H_data[y][x] = data[y][(2 * x) + 1]
return (L_data, H_data)
def vh_synthesis(state, LL_data, HL_data, LH_data, HH_data):
"""(15.4.3) Interleaved vertical and horizontal synthesis."""
synth = new_array(2 * height(LL_data), 2 * width(LL_data))
# Interleave transform data (as expected by synthesis routine)
for y in range(height(synth) // 2):
for x in range(width(synth) // 2):
synth[2 * y][2 * x] = LL_data[y][x]
synth[2 * y][2 * x + 1] = HL_data[y][x]
synth[2 * y + 1][2 * x] = LH_data[y][x]
synth[2 * y + 1][2 * x + 1] = HH_data[y][x]
# Synthesis
for x in range(width(synth)):
oned_synthesis(column(synth, x), state["wavelet_index"])
for y in range(height(synth)):
oned_synthesis(row(synth, y), state["wavelet_index_ho"])
# Bit shift, if required
shift = filter_bit_shift(state)
if shift > 0:
for y in range(height(synth)):
for x in range(width(synth)):
synth[y][x] = (synth[y][x] + (1 << (shift - 1))) >> shift
return synth
def two_d_array():
"""A 2D array of numbers for test purposes."""
a = new_array(16, 32)
for y in range(height(a)):
for x in range(width(a)):
a[y][x] = (y * 100) + x
return a
def offset_component(state, comp_data, c):
"""(15.5) Remove picture value offsets from a single component."""
for y in range(height(comp_data)):
for x in range(width(comp_data)):
if c == "Y":
comp_data[y][x] += 2 ** (state["luma_depth"] - 1)
else:
comp_data[y][x] += 2 ** (state["color_diff_depth"] - 1)
def vh_analysis(state, data):
"""
Interleaved vertical and horizontal analysis, inverse of vh_synthesis (15.4.3).
Returns a tuple (LL_data, HL_data, LH_data, HH_data)
"""
# Bit shift, if required
shift = filter_bit_shift(state)
if shift > 0:
for y in range(0, height(data)):
for x in range(0, width(data)):
data[y][x] = data[y][x] << shift
# Analysis
for y in range(0, height(data)):
oned_analysis(row(data, y), state["wavelet_index_ho"])
for x in range(0, width(data)):
oned_analysis(column(data, x), state["wavelet_index"])
# De-interleave the transform data
LL_data = new_array(height(data) // 2, width(data) // 2)
HL_data = new_array(height(data) // 2, width(data) // 2)
LH_data = new_array(height(data) // 2, width(data) // 2)
HH_data = new_array(height(data) // 2, width(data) // 2)
for y in range(0, (height(data) // 2)):
for x in range(0, (width(data) // 2)):
LL_data[y][x] = data[2 * y][2 * x]
HL_data[y][x] = data[2 * y][2 * x + 1]
LH_data[y][x] = data[2 * y + 1][2 * x]
HH_data[y][x] = data[2 * y + 1][2 * x + 1]
return (LL_data, HL_data, LH_data, HH_data)
def h_synthesis(state, L_data, H_data):
"""(15.4.2) Horizontal-only synthesis."""
synth = new_array(height(L_data), 2 * width(L_data))
# Interleave transform data (as expected by synthesis routine)
for y in range(height(synth)):
for x in range(width(synth) // 2):
synth[y][2 * x] = L_data[y][x]
synth[y][(2 * x) + 1] = H_data[y][x]
# Synthesis
for y in range(height(synth)):
oned_synthesis(row(synth, y), state["wavelet_index_ho"])
# Bit shift, if required
shift = filter_bit_shift(state)
if shift > 0:
for y in range(height(synth)):
for x in range(width(synth)):
synth[y][x] = (synth[y][x] + (1 << (shift - 1))) >> shift
return synth
def apply_dc_prediction(band):
"""(13.4) Inverse of dc_prediction."""
for y in reversed(range(0, height(band))):
for x in reversed(range(0, width(band))):
if x > 0 and y > 0:
prediction = mean(band[y][x - 1], band[y - 1][x - 1],
band[y - 1][x])
elif x > 0 and y == 0:
prediction = band[0][x - 1]
elif x == 0 and y > 0:
prediction = band[y - 1][0]
else:
prediction = 0
band[y][x] -= prediction
def clip_component(state, comp_data, c):
"""(15.5)"""
for y in range(height(comp_data)):
for x in range(width(comp_data)):
if c == "Y":
comp_data[y][x] = clip(
comp_data[y][x],
-(2**(state["luma_depth"] - 1)),
2**(state["luma_depth"] - 1) - 1,
)
else:
comp_data[y][x] = clip(
comp_data[y][x],
-(2**(state["color_diff_depth"] - 1)),
2**(state["color_diff_depth"] - 1) - 1,
)
def remove_offset_component(state, comp_data, c):
"""
Inverse of offset_component (15.5). Centers picture values around zero.
Parameters
==========
state : :py:class:`vc2_conformance.pseudocode.state.State`
Where ``luma_depth`` and ``color_diff_depth`` are defined.
current_picture : {comp: [[pixel_value, ...], ...], ...}
Will be mutated in-place.
"""
for y in range(height(comp_data)):
for x in range(width(comp_data)):
if c == "Y":
comp_data[y][x] -= 2 ** (state["luma_depth"] - 1)
else:
comp_data[y][x] -= 2 ** (state["color_diff_depth"] - 1)
def num_array():
a = new_array(3, 5)
for x in range(width(a)):
for y in range(height(a)):
a[y][x] = x + (y * 10)
return a
def test_width():
assert width(new_array(0, 0)) == 0
assert width(new_array(20, 10)) == 10
def make_transform_data_ld_lossy(picture_bytes,
transform_coeffs,
minimum_qindex=0):
"""
Quantize and pack transform coefficients into LD picture slices in a
:py:class:`TransformData`.
Raises :py:exc:`InsufficientLDPictureBytesError` if ``picture_bytes`` is too
small.
Parameters
==========
picture_bytes : int
The total number of bytes the picture slices should take up (i.e.
including qindex and slice_y_length fields). Must be at least 1 byte
per slice.
transform_coeffs : [[:py:class:`SliceCoeffs`, ...], ...]
minimum_qindex : int
If provided, gives the quantization index to start with when trying to
find a suitable quantization index.
Returns
=======
transform_data : :py:class:`vc2_conformance.bitstream.TransformData`
"""
# Used by slice_bytes
state = State(
slices_x=width(transform_coeffs),
slices_y=height(transform_coeffs),
slice_bytes_numerator=picture_bytes,
slice_bytes_denominator=width(transform_coeffs) *
height(transform_coeffs),
)
transform_data = TransformData(ld_slices=[])
for sy, transform_coeffs_row in enumerate(transform_coeffs):
for sx, transform_coeffs_slice in enumerate(transform_coeffs_row):
target_size = 8 * slice_bytes(state, sx, sy)
target_size -= 7 # qindex field
target_size -= intlog2(target_size) # slice_y_length field
if target_size < 0:
raise InsufficientLDPictureBytesError()
# Interleave color components
y_coeffs = transform_coeffs_slice.Y
c_coeffs = ComponentCoeffs(
coeff_values=interleave(
transform_coeffs_slice.C1.coeff_values,
transform_coeffs_slice.C2.coeff_values,
),
quant_matrix_values=interleave(
transform_coeffs_slice.C1.quant_matrix_values,
transform_coeffs_slice.C2.quant_matrix_values,
),
)
# Quantize each slice to fit
qindex, (y_transform, c_transform) = quantize_to_fit(
target_size,
[y_coeffs, c_coeffs],
minimum_qindex=minimum_qindex,
)
transform_data["ld_slices"].append(
make_ld_slice(
y_transform,
c_transform,
qindex,
))
return transform_data
def make_transform_data_hq_lossy(picture_bytes,
transform_coeffs,
minimum_qindex=0,
minimum_slice_size_scaler=1):
"""
Quantize and pack transform coefficients into HQ picture slices in a
:py:class:`TransformData`.
Raises :py:exc:`InsufficientHQPictureBytesError` if ``picture_bytes`` is too
small.
Parameters
==========
picture_bytes : int
The total number of bytes the picture slices should take up (i.e.
including ``qindex`` and ``slice_{y,c1,c2}_length`` fields). Must allow
at least 4 bytes per slice. When slice sizes are large enough to
require a slice_size_scaler larger than 1, ``picture_bytes -
4*num_slices`` must be a multiple of ``slice_size_scaler``, otherwise
the total picture size will deviate by up to ``slice_size_scaler``
bytes from ``picture_bytes``.
transform_coeffs : [[:py:class:`SliceCoeffs`, ...], ...]
minimum_qindex : int
If provided, gives the quantization index to start with when trying to
find a suitable quantization index.
minimum_slice_size_scaler : int
Specifies the minimum slice_size_scaler to be used for high quality
pictures. Ignored in low delay mode.
Returns
=======
slice_size_scaler : int
transform_data : :py:class:`vc2_conformance.bitstream.TransformData`
"""
slices_x = width(transform_coeffs)
slices_y = height(transform_coeffs)
# Determine the largest size a slice could be
num_slices = slices_x * slices_y
# Find the smallest slice size scaler which lets this size fit into 8 bits
slice_size_scaler = max(
get_safe_lossy_hq_slice_size_scaler(picture_bytes, num_slices),
minimum_slice_size_scaler,
)
# Work out the total bitstream space available after slice overheads are
# accounted for (NB: 4 bytes overhead per slice due to qindex and
# slice_{y,c1,c2}_length fields).
total_coeff_bytes = picture_bytes - (num_slices * 4)
if total_coeff_bytes < 0:
raise InsufficientHQPictureBytesError()
# We'll repurpose slice_bytes (13.5.3.2) to compute the number of
# slice_size_scaler bytes available for transform coefficients in each
# picture slice. (i.e. it won't compute the size of the slice in bytes but
# the number of slice_size_scaler bytes available for transform
# coefficients in each slice).
state = State(
slices_x=slices_x,
slices_y=slices_y,
slice_bytes_numerator=total_coeff_bytes,
slice_bytes_denominator=num_slices * slice_size_scaler,
)
transform_data = TransformData(hq_slices=[])
for sy, transform_coeffs_row in enumerate(transform_coeffs):
for sx, transform_coeffs_slice in enumerate(transform_coeffs_row):
# NB: Actually calculates multiples of slice_size_scaler bytes
# after all length/qindex fields accounted for. See comment above
# "state = State(".
total_length = slice_bytes(state, sx, sy)
target_size = 8 * slice_size_scaler * total_length
# Quantize each slice to fit
qindex, (y_transform, c1_transform,
c2_transform) = quantize_to_fit(
target_size,
transform_coeffs_slice,
8 * slice_size_scaler,
minimum_qindex,
)
transform_data["hq_slices"].append(
make_hq_slice(
y_transform,
c1_transform,
c2_transform,
total_length,
qindex,
slice_size_scaler,
))
return slice_size_scaler, transform_data