alpaka::onAcc::SimdAlgo.concurent() does not respect const modifier

[SimeonEhrig]

In the babelstream benchmark kernel, all IMdSpan objects in the SimdForEachKernel kernel are annotate with const. Therefore they should be not writable.

alpaka3/benchmark/babelstream/src/babelStreamMainTest.cpp

Lines 62 to 77 in 66540b9

	struct SimdForEachKernel
	{
	//! \param acc The accelerator to be executed on.
	//! \param func functor applied to each SIMD package.
	//! \param arg0 MdSpan from which the problem size is derived
	//! \param args MdSpan other spans
	ALPAKA_FN_ACC void operator()(
	auto const& acc,
	auto const& func,
	alpaka::concepts::MdSpan auto const& arg0,
	alpaka::concepts::MdSpan auto const&... args) const
	{
	auto simdGrid = onAcc::SimdAlgo{onAcc::worker::threadsInGrid};
	simdGrid.concurrent(acc, arg0.getExtents(), func, arg0, args...);
	}
	};

The kernel is called with the functor SimdCopyOp functor, which writes to a buffer, which should be forbidden.

alpaka3/benchmark/babelstream/src/babelStreamMainTest.cpp

Lines 90 to 96 in 66540b9

	struct SimdCopyOp
	{
	constexpr void operator()(auto const&, auto const a, auto c) const
	{
	c = a.load();
	}
	};

alpaka3/benchmark/babelstream/src/babelStreamMainTest.cpp

Lines 291 to 299 in 66540b9

	measureKernelExec(
	[&]()
	{
	queue.enqueue(
	exec,
	dataBlocking,
	KernelBundle{SimdForEachKernel{}, SimdCopyOp{}, bufAccInputA, bufAccOutputC});
	},
	"CopyKernel");

I didn't check the code, but I think we lost the const information when access to memory pointer to create the simd types.

[psychocoderHPC]

I added to the first unit test in mem/ConstCorrectness.cpp the following test which fails

STATIC_REQUIRE_FALSE(std::constructible_from<MutMdSpan, MutMdSpan const&>);

If I do not oversee something, this is also what happens in the code snippet you posted.

[psychocoderHPC]

@SimeonEhrig Is this not the only case we can not detect, if someone is passing a parameter as const& than the default copy constructor is used. So if a const reference is then passed to a function that takes the argument as copy, we can not avoid the change of constness.

[SimeonEhrig]

I think, we can handle it for our functions, because we use a universal reference. The universal reference deduces a const annotation:

https://github.com/SimeonEhrig/MyCpp/blob/4152da035a4dd509587a0c3adfdbb457a1e76a0c/features/11/type_deduction/main.cpp#L69-L72

Printed the type of c:

alpaka3/benchmark/babelstream/src/babelStreamMainTest.cpp

Lines 90 to 96 in 66540b9

	struct SimdCopyOp
	{
	constexpr void operator()(auto const&, auto const a, auto c) const
	{
	c = a.load();
	}
	};

It is:

alpaka::SimdPtr<alpaka::View<alpaka::api::Host, float, alpaka::Vec<unsigned int, 1>, alpaka::Alignment<16>>, alpaka::Vec<unsigned int, 1>, alpaka::Alignment<>, alpaka::Vec<unsigned int, 1, alpaka::detail::CVec<unsigned int, 4>>> 

I'm not fully sure, if this works like I expected, but I think, if make the value_type of the alpaka::View const because we deduces the const from concurrency(), it is not write-able anymore and we get the expected compiler error.