Comparison typo in index_sequence_scan_impl::get

[elfego]

Dear Kokkos developers,

I believe I've found a mistake in the template specialization index_sequence_scan_impl<R, FirstVal> within include/experimental/__p0009_bits/extents.hpp.

In the get member function (lines 162 and 166), the current comparison logic seems wrong and inconsistent with the implementation on line 147.

• Current: Uses r < R or after the static_cast<int64_t> of each variable
• Expected: I believe the correct comparison should be R < r.

This can be check by

// currently: FirstVal = 8, the cummulative sum before element r = 2 (which is past FirstVal) should be 8
constexpr size_t r = 2;
static_assert((detail::index_sequence_scan_impl<0, 8>::get(r) == 8)); // Fails

Could you please take a look to see if this is indeed a bug or if I am misinterpreting the intended logic?

Best regards,

Elfego

[nmm0]

Thanks for posting an issue! The intention is this code is not supposed to be a public API and it is only used for accessing inside the dynamic extents array. The way it is currently written means that access past-the-end will always be in bounds of the dynamic extents array, though admittedly it's not very obvious from the code.

That said past-the-end access isn't intended to be a use case of this internal utility.

[elfego]

Many thanks for the clarification! I appreciate that this is an internal utility, but I’m concerned that the current logic leads to an off-by-one error in the internal calculations.

Specifically, the inequality r < R in lines 162/166 of extents.hpp (compared to R < r in line 147) appears to cause the scan to miss the final element in the sequence. While it may stay "in bounds" of the array, it effectively treats the last value as zero regardless of the actual slice type.

Since this is used by deduce_layout_{left,right}_submapping in submdspan_mapping.hpp, I worry it might miscalculate the count_range for certain submdspan operations.

I’m happy to put together a minimal test case to see if this manifests as a detectable bug in submdspan. Would that be helpful, or would you prefer to just verify the consistency between line 147 and lines 162/166?

[nmm0]

Certainly more tests would be welcome, especially if you can find a case where it causes a problem!

For reference, I wrote a little compile time test:

using scan0 = Kokkos::detail::index_sequence_scan_impl<0, 0, 1, 1, 0, 1>;
MDSPAN_STATIC_TEST(scan0::get(0) == 0);
MDSPAN_STATIC_TEST(scan0::get(1) == 0);
MDSPAN_STATIC_TEST(scan0::get(2) == 1);
MDSPAN_STATIC_TEST(scan0::get(3) == 2);
MDSPAN_STATIC_TEST(scan0::get(4) == 2);
MDSPAN_STATIC_TEST(scan0::get(5) == 2); // out of bounds
MDSPAN_STATIC_TEST(scan0::get(6) == 2); // out of bounds

And basically these correspond to what we want -- (the first template argument to the scan is the starting index), so this reflects the compressed index you would get if you have 1s for every dynamic index. Since the size of the dynamic index array is 3, even if you go past the end it will be in bounds though that's not necessary/required.