def test01_construct(variant_scalar_rgb):
from mitsuba.core.xml import load_string
from mitsuba.render import ImageBlock
im = ImageBlock([33, 12], 4)
assert im is not None
assert ek.all(im.offset() == 0)
im.set_offset([10, 20])
assert ek.all(im.offset() == [10, 20])
assert ek.all(im.size() == [33, 12])
assert im.warn_invalid()
assert im.border_size() == 0 # Since there's no reconstruction filter
assert im.channel_count() == 4
assert im.data() is not None
rfilter = load_string("""<rfilter version="2.0.0" type="gaussian">
<float name="stddev" value="15"/>
</rfilter>""")
im = ImageBlock([10, 11], 2, filter=rfilter, warn_invalid=False)
assert im.border_size() == rfilter.border_size()
assert im.channel_count() == 2
assert not im.warn_invalid()
im = ImageBlock([10, 11], 6)
assert im is not None
im.channel_count() == 6
def test03_put_values_basic(variant_scalar_rgb):
from mitsuba.core import srgb_to_xyz
from mitsuba.core.xml import load_string
from mitsuba.render import ImageBlock
# Recall that we must pass a reconstruction filter to use the `put` methods.
rfilter = load_string("""<rfilter version="2.0.0" type="box">
<float name="radius" value="0.4"/>
</rfilter>""")
im = ImageBlock([10, 8], 5, filter=rfilter)
im.clear()
# From a spectrum & alpha value
border = im.border_size()
ref = np.zeros(shape=(im.height() + 2 * border, im.width() + 2 * border,
3 + 1 + 1))
for i in range(border, im.height() + border):
for j in range(border, im.width() + border):
spectrum = np.random.uniform(size=(3, ))
ref[i, j, :3] = srgb_to_xyz(spectrum)
ref[i, j, 3] = 1 # Alpha
ref[i, j, 4] = 1 # Weight
# To avoid the effects of the reconstruction filter (simpler test),
# we'll just add one sample right in the center of each pixel.
im.put([j + 0.5, i + 0.5], [], spectrum, alpha=1.0)
check_value(im, ref, atol=1e-6)
def test04_put_packets_basic(variant_scalar_rgb):
from mitsuba.core import srgb_to_xyz
try:
mitsuba.set_variant("packet_rgb")
from mitsuba.core.xml import load_string
from mitsuba.render import ImageBlock
except ImportError:
pytest.skip("packet_rgb mode not enabled")
# Recall that we must pass a reconstruction filter to use the `put` methods.
rfilter = load_string("""<rfilter version="2.0.0" type="box">
<float name="radius" value="0.4"/>
</rfilter>""")
im = ImageBlock([10, 8], 5, filter=rfilter)
im.clear()
n = 29
positions = np.random.uniform(size=(n, 2))
positions[:,
0] = np.floor(positions[:, 0] * (im.width() - 1)).astype(np.int)
positions[:,
1] = np.floor(positions[:, 1] * (im.height() - 1)).astype(np.int)
# Repeat some of the coordinates to make sure that we test the case where
# the same pixel receives several values
positions[-3:, :] = positions[:3, :]
spectra = np.arange(n * 3).reshape((n, 3))
alphas = np.ones(shape=(n, ))
border = im.border_size()
ref = np.zeros(shape=(im.height() + 2 * border, im.width() + 2 * border,
3 + 1 + 1))
for i in range(n):
(x, y) = positions[i, :] + border
ref[int(y), int(x), :3] += srgb_to_xyz(spectra[i, :])
ref[int(y), int(x), 3] += 1 # Alpha
ref[int(y), int(x), 4] += 1 # Weight
# Vectorized `put`
im.put(positions + 0.5, [], spectra, alphas)
check_value(im, ref, atol=1e-6)
def test05_put_with_filter(variant_scalar_rgb):
from mitsuba.core import srgb_to_xyz
from mitsuba.core.xml import load_string
from mitsuba.render import ImageBlock
"""The previous tests used a very simple box filter, parametrized so that
it essentially had no effect. In this test, we use a more realistic
Gaussian reconstruction filter, with non-zero radius."""
try:
mitsuba.set_variant("packet_rgb")
from mitsuba.core.xml import load_string as load_string_packet
from mitsuba.render import ImageBlock as ImageBlockP
except ImportError:
pytest.skip("packet_rgb mode not enabled")
rfilter = load_string("""<rfilter version="2.0.0" type="gaussian">
<float name="stddev" value="0.5"/>
</rfilter>""")
rfilter_p = load_string_packet("""<rfilter version="2.0.0" type="gaussian">
<float name="stddev" value="0.5"/>
</rfilter>""")
size = [12, 12]
im = ImageBlockP(size, 5, filter=rfilter_p)
im.clear()
im2 = ImageBlock(size, 5, filter=rfilter)
im2.clear()
positions = np.array([[5, 6], [0, 1], [5, 6], [1, 11], [11, 11], [0, 1],
[2, 5], [4, 1], [0, 11], [5, 4]],
dtype=np.float)
n = positions.shape[0]
positions += np.random.uniform(size=positions.shape, low=0, high=0.95)
spectra = np.arange(n * 3).reshape((n, 3))
alphas = np.ones(shape=(n, ))
radius = int(math.ceil(rfilter.radius()))
border = im.border_size()
ref = np.zeros(shape=(im.height() + 2 * border, im.width() + 2 * border,
3 + 1 + 1))
for i in range(n):
# -- Scalar `put`
im2.put(positions[i, :], [], spectra[i, :], alpha=1.0)
# Fractional part of the position
offset = positions[i, :] - positions[i, :].astype(np.int)
# -- Reference
# Reconstruction window around the pixel position
pos = positions[i, :] - 0.5 + border
lo = np.ceil(pos - radius).astype(np.int)
hi = np.floor(pos + radius).astype(np.int)
for dy in range(lo[1], hi[1] + 1):
for dx in range(lo[0], hi[0] + 1):
r_pos = np.array([dx, dy])
w_pos = r_pos - pos
if (np.any(r_pos < 0) or np.any(r_pos >= ref.shape[:2])):
continue
weight = rfilter.eval_discretized(
w_pos[0]) * rfilter.eval_discretized(w_pos[1])
xyz = srgb_to_xyz(spectra[i, :])
ref[r_pos[1], r_pos[0], :3] += weight * xyz
ref[r_pos[1], r_pos[0], 3] += weight * 1 # Alpha
ref[r_pos[1], r_pos[0], 4] += weight * 1 # Weight
# -- Vectorized `put`
im.put(positions, [], spectra, alphas)
check_value(im, ref, atol=1e-6)
check_value(im2, ref, atol=1e-6)
