def _update_scaled_data(self, label):
# If the data slice hasn't been set yet, we should stop here
if self._data_bounds is None:
return
index = self.volumes[label]['index']
clim = self.volumes[label].get('clim', None)
data = self.volumes[label]['data']
# With certain graphics cards, sending the data in one chunk to OpenGL
# causes artifacts in the rendering - see e.g.
# https://github.com/vispy/vispy/issues/1412
# To avoid this, we process the data in chunks. Since we need to do
# this, we can also do the copy and renormalization on the chunk to
# avoid excessive memory usage.
# To start off we need to tell the texture about the new shape
self.shared_program['u_volumetex_{0:d}'.format(index)].resize(
data.shape)
# Determine the chunk shape - the value of 128 as the minimum value
# is arbitrary but appears to work nicely. We can reduce that in future
# if needed.
sliced_data = data.compute_fixed_resolution_buffer(self._data_bounds)
chunk_shape = [min(x, 128, self.resolution) for x in sliced_data.shape]
# FIXME: shouldn't be needed!
zeros = np.zeros(self._vol_shape, dtype=np.float32)
self.shared_program['u_volumetex_{0:d}'.format(index)].set_data(zeros)
# Now loop over chunks
for view in iterate_chunks(sliced_data.shape, chunk_shape=chunk_shape):
chunk = sliced_data[view]
chunk = chunk.astype(np.float32)
if clim is not None:
chunk -= clim[0]
chunk *= 1 / (clim[1] - clim[0])
# PERF: nan_to_num doesn't actually help memory usage as it runs
# isnan internally, and it's slower, so we just use the following
# methind. In future we could do this directly with a C extension.
chunk[np.isnan(chunk)] = 0.
offset = tuple([s.start for s in view])
if chunk.size == 0:
continue
self.shared_program['u_volumetex_{0:d}'.format(index)].set_data(
chunk, offset=offset)
def contains3d(self, x, y, z):
"""
Test whether the projected coordinates are contained in the 2d ROI.
"""
if not self.defined():
raise UndefinedROI
x = np.asarray(x)
y = np.asarray(y)
z = np.asarray(z)
# Since the projection can significantly increase the memory usage, we
# do the following operation in chunks. In future we could likely use
# e.g. vaex, dask, or other multi-threaded/fast libraries to speed this
# and other ROI code up.
mask = np.zeros(x.shape, dtype=bool)
for slices in iterate_chunks(x.shape, n_max=1000000):
# Work in homogeneous coordinates so we can support perspective
# projections as well
x_sub, y_sub, z_sub = x[slices], y[slices], z[slices]
vertices = np.array([x_sub, y_sub, z_sub, np.ones(x_sub.shape)])
# The following returns homogeneous screen coordinates
screen_h = np.tensordot(self.projection_matrix,
vertices,
axes=(1, 0))
# Convert to screen coordinates, as we don't care about z
screen_x, screen_y = screen_h[:2] / screen_h[3]
mask[slices] = self.roi_2d.contains(screen_x, screen_y)
return mask
def contains3d(self, x, y, z):
"""
Test whether the projected coordinates are contained in the 2d ROI.
"""
if not self.defined():
raise UndefinedROI
x = np.asarray(x)
y = np.asarray(y)
z = np.asarray(z)
# Since the projection can significantly increase the memory usage, we
# do the following operation in chunks. In future we could likely use
# e.g. vaex, dask, or other multi-threaded/fast libraries to speed this
# and other ROI code up.
mask = np.zeros(x.shape, dtype=bool)
for slices in iterate_chunks(x.shape, n_max=1000000):
# Work in homogeneous coordinates so we can support perspective
# projections as well
x_sub, y_sub, z_sub = x[slices], y[slices], z[slices]
vertices = np.array([x_sub, y_sub, z_sub, np.ones(x_sub.shape)])
# The following returns homogeneous screen coordinates
screen_h = np.tensordot(self.projection_matrix,
vertices, axes=(1, 0))
# Convert to screen coordinates, as we don't care about z
screen_x, screen_y = screen_h[:2] / screen_h[3]
mask[slices] = self.roi_2d.contains(screen_x, screen_y)
return mask
