CIL FBP/FDK recon with TIGRE backend

[gfardell]

Note: timings are looping in CIL over the CIL wrapped astra and tigre, so not necessarily representative of the backends when called directly.

• For single calls times are mean of 10 repeats.
• server has 2x V100s and 180GB RAM although it seems that the backends currently only utilise a single card for FBP

FBP Times

1k cube

• 1800 1024x1024 radiographs
• reconstruction volume 1024x1024x1024

Full volume

Implementation	time (s)	speed-up
TIGRE 2.1	166	1
TIGRE (master)	50	3.3
ASTRA	58	N/A
CIL	19	8.7

Processed slice-by-slice

Implementation	time (s)	speed-up
TIGRE 2.1	212	1
TIGRE (master)	103	2
ASTRA	213	N/A
CIL	44	4.8

Processed in chunks of 32 slices

The number of slices per chunk can be set by the user with:

fbp.set_split_processing(slices_per_chunk=32)

Implementation	time (s)
CIL	22

2k cube

• 3600 2048x2048 radiographs
• reconstruction volume 2048x2048x2048

Processed slice-by-slice

Implementation	time (s)	speed-up
TIGRE 2.1	1806	1
TIGRE (master)	941	1.9
ASTRA	1715	N/A
CIL	498	3.6

Processed in chunks of 32 slices

Implementation	time (s)
CIL	275

FDK Times

• For single calls times are mean of 10 repeats.
• server has 2x V100s

1k cube

• 1800 1024x1024 radiographs
• reconstruction volume 1024x1024x1024

Implementation	time (s)	speed-up
TIGRE 2.1	161	1
TIGRE (master)	44	3.7
ASTRA	56	N/A
CIL	11	15

2k cube

• 3600 2048x2048 radiographs

• reconstruction volume 2048x2048x2048

• I'm currently unable to run the TIGRE or ASTRA FDKs wrapped by CIL on this size without running in to memory issues.

Implementation	time (s)
CIL	162