Timing versus DRR size

Along with tips for rendering DRRs that don’t fit in memory 
import numpy as np
import torch

from diffdrr.data import load_example_ct
from diffdrr.drr import DRR
from diffdrr.visualization import plot_drr
# Read in the volume
volume, spacing = load_example_ct()
device = "cuda" if torch.cuda.is_available() else "cpu"

# Get parameters for the detector
bx, by, bz = np.array(volume.shape) * np.array(spacing) / 2
translations = torch.tensor([[bx, by, bz]]).to(device)
rotations = torch.tensor([[np.pi, 0, np.pi / 2]]).to(device)
height = 100

drr = DRR(volume, spacing, sdr=300.0, height=height, delx=4.0).to(device)

del drr
9.28 ms ± 375 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)
height = 200

drr = DRR(volume, spacing, sdr=300.0, height=height, delx=4.0).to(device)

del drr
33.3 ms ± 14.3 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
height = 300

drr = DRR(volume, spacing, sdr=300.0, height=height, delx=4.0).to(device)

del drr
72.1 ms ± 30.2 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
height = 400

drr = DRR(volume, spacing, sdr=300.0, height=height, delx=4.0).to(device)

del drr
123 ms ± 35.1 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
height = 500

drr = DRR(volume, spacing, sdr=300.0, height=height, delx=4.0).to(device)

del drr
187 ms ± 96.1 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

Memory constraints

Up until this point, we could compute every ray in the DRR in one go on the GPU. However, as the DRRs get bigger, we will quickly run out of memory. For example, on a 12 GB GPU, computing a 600 by 600 DRR will raise a CUDA memory error.

Tip

To render DRRs whose computation won’t fit in memory, we can compute patches of the DRR at a time. Pass patch_size to the DRR module to specify the size of the patch. Note the patch size must evenly tile (height, width).

height = 600
patch_size = 150

drr = DRR(
    volume, spacing, sdr=300.0, height=height, delx=4.0, patch_size=patch_size
).to(device)

del drr
183 ms ± 214 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
height = 750
patch_size = 150

drr = DRR(
    volume, spacing, sdr=300.0, height=height, delx=4.0, patch_size=patch_size
).to(device)

del drr
259 ms ± 92.7 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)
height = 1000
patch_size = 250

drr = DRR(
    volume, spacing, sdr=300.0, height=height, delx=4.0, patch_size=patch_size
).to(device)

del drr
417 ms ± 234 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)
height = 1500
patch_size = 250

drr = DRR(
    volume, spacing, sdr=300.0, height=height, delx=4.0, patch_size=patch_size
).to(device)

del drr
826 ms ± 268 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)

With patch_size, the only limitation is DRR storage, not computation.