Transfer observations

[mtazzari]

Transfer observations

[mtazzari]

That's a very good point! I have already thought to do that, but I was unsure where/how we could store and then access the data from different processes on GPU (->Unified Memory?).

This feature could be released in a future release.

Within a single fit, single-wavelength this feature could be beneificial, but I see a potential issue when we run a multi-wavelength fit. Then, for each likelihood computation, we have several calls (one for each wavelength) to galario.sample() or galario.chi2() with different urot, vrot, fobs_re, fobs_im.
The currently implemented approach requires MemcpyH2D() these arrays all the times, but it is self-contained: each sample() or chi2() call has all the data it needs and frees the space immediately afterwards. With the proposed feature, it is possible that the space required to store these multi-wavelength static arrays is too large. Anyway, I guess it would depend on:

• the total GPU RAM available
• the number of processes using the GPU.
• the size of the observational data

[mtazzari]

We will not support this because we don't want to handle runtime memory management on the GPU.