GammaLib

I implemented a vectorised response computation in GCOMResponse::irf_ptsrc() making use of the characteristics of the COMPTEL data space. Specifically, phigeo is now only computed once for the Chi/Psi pixel.

Initial testing using a five energy-band Crab fit reduced the ctlike execution time from 23.23 seconds to 12.48 seconds, hence leading to a speed-up of a factor of 2. More testing is needed to verify for example that the code also works for TS map computation. Also, a valgrind analysis may be done to further speed-up the analysis.

There is basically no speed-up in the spectral analysis, which is normal since the spatial parameters are fixed in the spectral analysis. Consequently, the cached values will be used.

The same holds for the TS map computation. No speed-up because the spatial parameters are fixed.

Here is a valgrind analysis of the 5 standard band ctlike run on viewing period 1. Obviously a significant fraction of time is spent in the matrix and vector handling.

Specifically, the GMatrixSparse::row method is called many times in GObservation::likelihood_poisson_binned which uses quite some computing time. The reason for this is that the matrices are stored column-wise, hence row access needs quite some computations. It should be checked whether a transposition of the matrix and a column access residuals in actually faster code.

Transposing the matrix in GObservation::likelihood_poisson_binned and accessing it through the GMatrixSparse::column method leads to faster code, using 10.21 instead of 12.50 seconds for the 5-band ctlike fit of viewing period 1. For the TS map, the computation needed 3671.03 seconds instead of 5017.22 seconds. Overall, a speed-up of about 20-25% is achieved.

I also checked the speed-up for fitting the Crab to simulated CTA data (fixed position, only prefactor and spectral index are free). Below the computation times and speed-up for the various analysis methods:

In all cases the code is faster. I will therefore implement the matrix transformation.

Below the valgrind analysis for the transposed matrix. As expected, matrix access is no longer a bottleneck (the GMatrixSparse::column method uses less than 1% of the time).

Merged into devel.