Table 2
Computational benchmarking.
Pre-compute mode | |||
---|---|---|---|
Bandlimit | Forward | Gradient | JIT compilation |
256 | 15 ms | 30 ms | 20 s |
512 | 100 ms | 200 ms | 25 s |
Recursive mode | |||
Bandlimit | Forward | Gradient | JIT compilation |
256 | 120 ms | 300 ms | 90 s |
1024 | 5s | 10s | 3m |
2048 | 20 s | 50 s | 6m |
Notes. Results of the SC transform provided by s2scat. These results were recovered on a single NVIDIA A100 40GB GPU, although it is possible to run across multiple GPUs. In our analysis we generate spherical images through 400 iterations to be conservative. In practice, however, we find that ∼100 iterations is typically sufficient, in which case an image at L = 256 can be generated in ~4s. Furthermore, batched generation can dramatically decrease per sample compute time. For example, 20 images at L = 256 can be generated in ~12s, corresponding to ∼0.5s per sample.
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.