Remove assertions in minmax_element binary operator

[julianmi]

The current minmax_element implementation requires in-order processing since its binary operator assumes that the index of the first input is lower than that of the second input. This is a performance improvement since we can skip the case when both input values are equal. The in-order processing is guaranteed by following the non-commutative reduce pattern. We currently check the index order with assertions with lead to high overheads when compiled without NDEBUG. This may lead to test timeouts or user code looking to be hanging. I think the assertions are not needed since we enforce them in the in-order reduce pattern and we don't use them in the similar min_element algorithm.

[dmitriy-sobolev]

@julianmi, could you remove the assertions in include\oneapi\dpl\pstl\hetero\algorithm_ranges_impl_hetero.h?

Meanwhile, I found no mention if assert can be used in SYCL device code reading its specification.
Given that there is a corresponding extension in oneAPI DPC++ Compiler, I assume that we should avoid using asserts in kernels, or at least use them conditionally.

[julianmi]

@julianmi, could you remove the assertions in include\oneapi\dpl\pstl\hetero\algorithm_ranges_impl_hetero.h?

Done.

Meanwhile, I found no mention if assert can be used in SYCL device code reading its specification. Given that there is a corresponding extension in oneAPI DPC++ Compiler, I assume that we should avoid using asserts in kernels, or at least use them conditionally.

I agree with your statement. I looks like assert seems to work for now in kernel code though. Anyways, I think our tests should capture such semantic issue in our kernels (and they do for this specific case).

[dmitriy-sobolev]

LGTM

[danhoeflinger]

LGTM