When reconstructing > 800 million LORs with OSEM on a decent number of threads, the time taken to
read the LORs from file
fill scattergram
convert HDF5 LORs into MLEM LORs
takes a significant proportion of the total time.
Step 3 should be trivially parallelizable.
2 would require separate scattergrams to be filled on each thread, and joined together at the end, similar to what is being done with images in the MLEM/OSEM iterations.
As for 1, this involves reading from SSD and I have no idea whether this can be parallelized sensibly.
When reconstructing > 800 million LORs with OSEM on a decent number of threads, the time taken to
takes a significant proportion of the total time.
Step 3 should be trivially parallelizable.
2 would require separate scattergrams to be filled on each thread, and joined together at the end, similar to what is being done with images in the MLEM/OSEM iterations.
As for 1, this involves reading from SSD and I have no idea whether this can be parallelized sensibly.