Jon Wright

I just hit the same problem with scikit-cuda=0.5.3 from pypi. In cublas.py then pdb shows `_cuda_version=7050` and the machine has `nvcc -V` returning version 6.5. I don't know how to...

So it turns out that if the poni file uses relative paths then it can't be moved around within the same file system. Vadim suggested just copying the entire spline...

Users get a root folder "/data/visitor/ma1234/id11/" where they store their data. For them this is also "~/id11/". When they get home it is an external disk like "E:\ma1234\id11" or gets...

Thanks for taking a look. For the example in your gist: the best time I could see is about 23 ms when running in parallel and this should correspond to...

For the scaling - I was looking for something single threaded to scale via one frame per cpu thread. An [example](https://gist.github.com/jonwright/05334bbd81e9c1559c2b52a9b9cc2349) with concurrent futures seems to be not too bad....

I was confused as to why my csr.dot was 3X slower. It seems that converting to float32 in preproc changes the output type: ``` pye = list(ai.engines.values())[1].engine %timeit result=pye._csr.dot(img.ravel()) print(pye._csr.dot(img.ravel()).dtype)...

So it seems the dynamic mask and 3 integrations might be done in ~45 ms for this case ([singleThread6M-mask_3array.pdf](https://github.com/silx-kit/pyFAI/files/9079071/singleThread6M-mask_3array.pdf). Although it is not really clear how to set this up...

I seem to remember that bitshuffle only scales well for a few cores, so for our data it is another reason to run several smaller jobs...

A good compromise (for me) could be to add a public API for getting a CSR matrix out. Would that be possible? It would need some additions to the docs...

Looks good to me, `ai.setup_CSR()` is returning an object that has `(indptr, indices, data)`. Assuming this will look similar after #1586 then it seems ideal. Thanks! Feel free to close...