From: Andreas K. <li...@in...> - 2018-02-08 21:22:59
|
Jeff Hammond <jef...@gm...> writes: >> I've been treating POCL as useful tool for validation -- i.e., it's easy >> to install via conda, and unlike some other OpenCL runtimes (i.e., Intel... >> not that I mean to offend Jeff), you guys are very responsive to >> acknowledging / fixing bugs, or at least explaining why I shouldn't be >> doing something that breaks POCL. I'll add my 'me too' to that in that I'm in a similar situation to Timo and Nick. I reported an issue [1] similar in flavor to Timo's a while back, and any progress on making outer-loop vectorization in POCL for AVX* a reliable and predictable affair [2] has the potential for considerable impact in the computational science community. (I made the conda-forge packages for llvm and pocl, that's how much I believe that's true.) [1] https://github.com/pocl/pocl/issues/340 [2] (say, as predictable and reliable as ISPC, for the sake of argument) > I don't take offense to the suggestion that Intel OpenCL is nontrivial to > install (at the very least, it requires sudo) Heh. If you know exactly how to butcher the tarballs and RPMs, you *can* get by without root. Anyway, moving along... :) Andreas |