I’ve recently started learning SYCL and I’m using CodePlay’s ComputeCpp CE 2.3 to build some sample proglets (that I cannot link since new users can only put two links in a post; links to the proglets can be found in the LLVM and pocl issue linked below). I’ve successfully managed to run
sample (and the equivalent
sample-select on both CPU and GPU devices without issues.
sample-reduce implements a simple parallel reduction using local memory, and it works correctly on the CPU and on the Intel iGP, but when I try to run it on a CUDA GPU through pocl I get a segmentation fault in LLVM. Details about the crash have also been reported in the LLVM issue tracker and in the pocl issue tracker.
As detailed in the LLVM issue tracker in particular, the issue seems specifically related to the declaration of the local memory for the reduction kernel.