Fast Hough Transform¶
The ARTISTS, are developing their own fast hough transform algorithm:
They perform the hough transforms, but not the peak finding or parameter extraction. We can certainly show you our serial and parallel efforts in that regard.
They are actually using FD simulation from Gavin, with 2:1 signal hits to noise hits, approximately: about 90K hits per 5ms window in a 1:2 XZ - YZ ratio.
Some detailed parameters:
time bin width: 4 ticks (62.5ns)
but the time cluster may contain more than 1 effective bin (# of hits > certain threshold)
The attached presentation (slides 9-13) shows their performance measurements.
Step 1: shorter time window (done)¶
Compared both current time slice algorithm and Artists' time clustering, the time slicer algorithm we are using right now makes more physics sense
and by resizing the window, it is actually faster by saving the combining time bins process in time clustering.
The "vpro" has been turned from a voting par to a config par, since we are dealing with fast particles in the first step.
Now the time per NDOS event (5ms) cost is 237ms, for just finding the hough peaks without the later production and trigger decision process (I think that is as far as the Artists get right now).