1. “MultiRing Almagest with GPU” Gianluca Lamanna (CERN) Mainz Collaboration meeting 8.9.2011 TDAQ WG

2. Gianluca Lamanna – 8.9.2011 Multi Ring @L1 interesting processes multitracks Several interesting processes (for physics or to study the detector performances) we want to collect are multitracks: 3 , Ke4,  0D,  e, … L1 triggers bandwidth limitations Selective L1 triggers are required to guarantee the bandwidth limitations 2 End decay region RICH @L1 1 track 3  The bandwidth in input @L1 is shared by 1 track (10% Ke3, 44%  0 ) and 3  (46%) L0 close RICH The decay vertex of the L0 survived 1 trk events is close to the RICH CHOD RICH ~90% see my talk 6.4.2011 Using the CHOD and the RICH is possible to reconstruct the decay vertex with enough resolution: ~90% reduction on 1 trk events (see my talk 6.4.2011) >60% 3  RICH >60% of the 3  has 2 or 3 trks on the RICH

3. Gianluca Lamanna – 8.9.2011 The Almagest algorithm @L1 single detector @L1 the trigger decision is based on single detector (or on “simply” combined detectors) RICHfast tracklesshigh resolution For the RICH we need a fast, trackless and high resolution ring reconstruction algorithm No suitable solutions in the market 3 Almagest Ptolemy’s theorem New algorithm (Almagest) based on Ptolemy’s theorem: “A quadrilater is cyclic (the vertex lie on a circle) if and only if is valid the relation: AD*BC+AB*DC=AC*BD “ GPUs Easily implemented on GPUs [See my talk 25.5.2011]

4. Gianluca Lamanna – 8.9.2011 Almagest for 2 rings 2 rings version 2 rings version: resolution High resolution Good capability to reconstruct intersecting rings fast Very fast two rings Only two rings Poor Poor immunity to noise 4

5. Gianluca Lamanna – 8.9.2011 Almagest for multiring multiring New version for multiring (up to N) 5 three points Select three points Almagest Almagest procedure: check if the other points are on the ring and refit Remove other rings Remove the used points and search for other rings

6. Gianluca Lamanna – 8.9.2011 Threshold optimization thresholds The thresholds to accept or reject the points are empirically determined: SS_THR SS_THR is for the Ptolemy’s theorem condition RCHK_THR RCHK_THR is for the recollecting procedure ring distributionhits density This optimization depends on ring distribution and hits density: have to be redone in official Montecarlo 6

7. Gianluca Lamanna – 8.9.2011Efficiency efficiency number of rings “physical” constraints The efficiency to found the correct number of rings can be improved using “physical” constraints in the real Montecarlo ~80% 5 or 3 rings~70% 4 rings At the moment is ~80% in case of 5 or 3 rings and ~70% in case of 4 rings 7

8. Gianluca Lamanna – 8.9.2011 Example (1) superimposed fully contained Good results also in case of superimposed or not fully contained rings 8

9. Gianluca Lamanna – 8.9.2011Example(2) external constraintsother strategies Room for improvements using external constraints or other strategies (under study) 9

10. Gianluca Lamanna – 8.9.2011Resolution resolution offline resolution The resolution is compatible with the offline resolution last fitted rings superimposed rings For the last fitted rings the resolution could be worst in case of superimposed rings A correct evaluation will be done on real Montecarlo 10

11. Gianluca Lamanna – 8.9.2011 Computing time 11

12. Gianluca Lamanna – 8.9.2011 Conclusions (1) multirings5 rings Almagest algorithm A procedure to search for multirings (up to 5 rings in this study) using an iterative procedure based on the Almagest algorithm has been implemented 80%) The efficiency (about 80%) can be improved using physical constrains in the real Montecarlo computing time per events maximum number of rings 3 rings~20 us 3 rings ~13 us 2 rings The computing time per events depends on the maximum number of rings allowed: for the 3 rings version ~20 us are need for 3 rings and ~13 us for 2 rings isn’t optimized GPU C1060 The algorithm isn’t optimized on the GPU and the card used is the C1060  room for improvement! 12

13. Gianluca Lamanna – 8.9.2011 Conclusions (2) S.Stamm double core CPU S.Stamm (summer students) made the same exercise using double core CPU [see Weekly meeting1.9.2011]: 0.81.4 ms – from 0.8 to 1.4 ms depending on the number of hits (in the three ring case) GPU The GPU is at least a factor 40 faster, as expected CPU RICH GPU In any case the CPU results are encouraging for the offline RICH reconstruction (…if without GPU…) 13