Action #3405: Implement vectorised response computation for energy dispersion - GammaLib - CTA IRAP Project Gateway

Action #3405

Implement vectorised response computation for energy dispersion

Added by Knödlseder Jürgen about 4 years ago. Updated almost 3 years ago.

Status:

In Progress

Start date:

10/18/2020

Priority:

Normal

Due date:

Assigned To:

Knödlseder Jürgen

% Done:

20%

Category:

Target version:

Duration:

Description

For the moment the vectorised response computation brings only a speed-up without energy dispersion. The code needs to be restructured to achieve also a speed-up with energy dispersion enabled.

Recurrence

No recurrence.

Related issues

History

#1 Updated by Knödlseder Jürgen about 4 years ago

One possibility to do this at high level would be to add methods that GEventCube that return the number of energy bins and that allow iterating over all event bins for a given energy bin. This will allow for instruments where data are binned in energy to process energy bin after energy bin, using high-level energy dispersion information.

What would be needed is something like this:

GEventBin* GEventCube::first_event_bin(const int& iebin);
GEventBin* GEventCube::next_event_bin(const int& iebin);

Note that a

GEbounds   GEvents::ebounds(void) const;

method exists already at the level of the GEvents class, which allows accessing how many energy bins exist.
The methods can then be used as follows to loop over all events in an energy bin:

for (GEventBin* bin = first_event_bin(iebin); bin != NULL; bin = next_event_bin(iebin)) {
    ...
}

#2 Updated by Knödlseder Jürgen about 4 years ago

I’m not sure that such methods are actually useful, since what counts is the minimisation of the spatial transformations.

#3 Updated by Knödlseder Jürgen about 4 years ago

Here is the current performance of the code using energy dispersion. The values were taken from issue #3203.

Code	CPU	Iterations	logL
Reference	11385.7 s	2	122531.429
Using vector response, still event-by-event evaluation	10871.2 s	2	122531.429

#4 Updated by Knödlseder Jürgen about 4 years ago

Status changed from New to In Progress
Assigned To set to Knödlseder Jürgen
% Done changed from 0 to 10

Before starting the vectorisation I wrote down the current code status in the TN0003 and restructured a bit the GResponse class so that four virtual methods are now dealing with event-wise and vectorised event probability computation for the two cases of neglecting or using the energy dispersion. This allows overloading of the various computations by instrument-specific implementations.

virtual double  eval_prob_no_edisp(const GModelSky&    model,
                                   const GEvent&       event,
                                   const GEnergy&      srcEng,
                                   const GTime&        srcTime,
                                   const GObservation& obs,
                                   const bool&         grad) const;
virtual double  eval_prob_edisp(const GModelSky&    model,
                                const GEvent&       event,
                                const GTime&        srcTime,
                                const GObservation& obs,
                                const bool&         grad) const;
virtual GVector eval_probs_no_edisp(const GModelSky&    model,
                                    const GObservation& obs,
                                    GMatrixSparse*      gradients) const;
virtual GVector eval_probs_edisp(const GModelSky&    model,
                                 const GObservation& obs,
                                 GMatrixSparse*      gradients) const;

#5 Updated by Knödlseder Jürgen about 4 years ago

% Done changed from 10 to 20

As a first modification, the vector computation of analytical spatial model gradients was added. So far there is no support for the event wise spatial model gradients in the CTA interface, but at least the code structure makes such a support possible, and support can be implemented at CTA level.

Before doing so, we should make sure that the code works still as expected for all spatial models. Note that the make check is successful.

#6 Updated by Knödlseder Jürgen about 4 years ago

Here the test results, where _b refers for former results obtained with the old code and _a refers to new results obtained with the new code.

Model	CPU_b	Iter_b	logL_b	CPU_a	Iter_a	logL_a	Comments
Disk	14.81	2	156240.662	14.80	2	156240.662	identical result
Gaussian	16.55	2	118106.615	16.25	2	118106.615	identical result
Ring	257.40	31	125762.471	263.59	31	125762.471	identical result
Shell	48.47	4	127541.952	49.38	4	127541.952	identical result
Profile	97.40	2	118004.032	105.36	2	118004.032	identical result
Gaussian (edisp)	10871.2	2	122531.429	10848	2	122531.429	identical result
Gaussian (unbinned)	3313.09	2	-11118399.291	3261.26	2	-11118399.291	identical result
Gaussian (binned)	1974.85	2	118107.706	1938.22	2	118107.706	identical result

#7 Updated by Knödlseder Jürgen about 4 years ago

Related to Action #3390: Implement vectorised computation of energy dispersion added

#8 Updated by Knödlseder Jürgen almost 3 years ago

Target version deleted (~~2.0.0~~)

Also available in: Atom PDF

GammaLib

Issues

Tags