Thanks Fred for the time spent on this. It looks like it is a very deep problem !
Concerning the code: This medfilt code never properly worked on POCL CPU-thread version while it works as expected on CUDA (POCL+CUDA) or other drivers. There is a huge amount of shared memory involved which may be the root of the problem (?) But about the interference I mainly see specfile (the C-library) which is our "first" guess inside "silx.io". This legacy code is far from our current standards ... Other point, pyopencl.array.zeros_like() is calling the opencl enfillBuffer function which caused me also issues in (POCL+CUDA). Creating a work-around on the silx side is an option as it has been done many times. I opened a bug on our side: https://github.com/silx-kit/silx/issues/2848 Cheers Begin forwarded message: Date: Thu, 19 Dec 2019 10:08:49 +0000 From: PICCA Frederic-Emmanuel <frederic-emmanuel.pi...@synchrotron-soleil.fr> To: "kief...@esrf.fr" <kief...@esrf.fr> Subject: pocl and silx Salut Jerome, on est en train d'essayer de demerder le problème avec pocl. On a mis les info la https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=946422 est-ce que cela te dis quelque choses. C'est vrai que si l'on import silxs.io avant silxs.opencl dans le petit snipset de test, ca change fondamentalement la done... Amicalement Fred -- Jérôme Kieffer tel +33 476 882 445
