What is the performance difference you are talking about? What you posted is a difference of ~2.27% which is very small (in my books anything <3% is insignificant and well withing the margin of statistical error but lets not argue about that).

That said, hot/cold caches could be at play here, can you retry the same measurements but run the opencv code first and sycl second? Other than that i suppose we need to compare the algorithms/code.

final note: i do not trust handwritten benchmarking code like this, i prefer to use google benchmark to help with benchmarking boilerplate (experiment setup/teardown and statistical significance.