debug multi-GPU SYCL nstream

[jeffhammond]

There may be a bug in multi-GPU SYCL nstream as reported in AdaptiveCpp/AdaptiveCpp#399

@illuhad

[illuhad]

Yes, details here:

the value of ls as defined here:

Kernels/Cxx11/nstream-multigpu-dpcpp.cc

Line 162 in eff76d7

ls[g] = elements_per_gpu;

is used as start offsets of the local buffers, e.g.

Kernels/Cxx11/nstream-multigpu-dpcpp.cc

Line 184 in eff76d7

const size_t start = (g>0) ? ls[g-1] : 0;

this does not look correct to me as I believe you would need to sum all the local lengths of all preceding GPUs to get the correct start offset.