Jerome Kieffer

On Sat, 03 Nov 2018 01:34:06 -0700 picca wrote: > Hello guyes, here the error reported by the Debian ci during the as installed unittest. > > https://ci.debian.net/data/autopkgtest/unstable/amd64/p/pyfai/1241085/log.gz thanks for...

On Sat, 03 Nov 2018 03:31:22 -0700 picca wrote: > Maybe this is because I am using this command line in order to run the tests. > > just tell...

PyFAI does not perform 1 or 2 histograms per integration, but 4 and soon 5 to be able to handle uncertainties (needed for sigma-clipping). On my Epyc processor (i.e. @...

So you are using the CSR matrix, transposed, to perform histograms ? then it is no more parallelizable. The CSR format was chosen because a single integration can be parallelized....

I just checked, see [profiling.pdf](https://github.com/silx-kit/pyFAI/files/9077832/profiling.pdf), and indeed: * scipy.sparse is faster than the cython code from pyFAI when working in single-threaded (157 vs 176ms) * The CSC matrix format is...

You code raises a comment: accumulator should not be float32, while coefficient of the matrix, diffraction intensity and normalization are all OK with 32 bits, you need at least 64...

The performance of preprocessing is terrible on my computer:  The best would probably to disable OpenMP for this function ! A lot of time (~50%) is lost in...

Vincent highlighted that the issue of "preproc" being slow does not show off on GPU. There, the processing takes 0.15s.Thus optimizing pyFAI for single threaded CPU code with the idea...

what about this ? https://github.com/silx-kit/pyFAI/blob/master/pyFAI/azimuthalIntegrator.py?plain=1#L447-L500 There are typos in the docstring but any of those returned object has a "csr" attribute which allows you to redefine your own integrator

I'll keep it open because several interesting benchmarks are referenced. You convinced me that having optimized single-threaded version could make sense in specific use-cases despite it is not ESRF's priority...