Add missing assertion to joint_of_group tests

[0x12CC]

The joint_of_group tests do not work if input vector has a size of 1. This change adds an assertion to guard against work group sizes too small for the test to run correctly. The new assertion matches the intent of other similar assertions in this header. A correct SYCL implementation that uses a work group size of 1 or 2 will now fail this assertion rather than failing the test itself.

[keryell]

But what about the ones where the checks hold?
Are there some interesting cases where it checks for something not tested otherwise? 🤔

[0x12CC]

But what about the ones where the checks hold? Are there some interesting cases where it checks for something not tested otherwise? 🤔

This test has four predicates:

              auto none_true = [&](T i) { return i == 0; };
              auto one_true = [&](T i) { return i == 1; };
              auto some_true = [&](T i) { return i > size / 2; };
              auto all_true = [&](T i) { return i <= size; };

The predicates for one_true and some_true do not work if the input has exactly one element. This is also the motivation for the existing assertions on line 199 and line 402.

The other two predicates form valid tests even when the workgroup size is 1. I just don't think it's worth running only those tests for that input size since the larger workgroup will also test with the same predicates.

[TApplencourt]

Should we skip and not assert? I mean work_group_size=1 seem to be valid as an implementation choice no?
For me assert should be about conformance precondition. skip should be for tests precondition

[0x12CC]

I used assert since the other tests use assert. My experience with ProtoSYCL was that I had a test failure, because the workgroup size was 2, but the implementation was correct. I thought the assert would make it very obvious to future implementers that the test itself doesn't work in those cases.

Although a workgroup size of 1 is a valid implementation choice, my preference is to only skip tests for practical implementation choices that limit test execution. Making every test work with any strictly conformant implementation with no practical assumptions on the provided capabilities seem like a large development and maintenance effort.

[TApplencourt]

I agree in principle about maintenance and porting effort, but in practice for this tests, it's a 10 characters change no?

[0x12CC]

I agree in principle about maintenance and porting effort, but in practice for this tests, it's a 10 characters change no?

I agree, as long as we change the existing assertions to skip as well. Changing only the one I added wouldn't make sense since a workgroup size of 1 would still hit the other assertions. I don't have a strong opinion about this so I'm happy to make any change the WG agrees on. I just thought leaving the test with no changes at all would be confusing.

[TApplencourt]

. I just thought leaving the test with no changes at all would be confusing.

Totally agree. I see the confusion, sorry my bad. What I mean was to skip the tests if work_group_size not greater than 2.

I would just replace the

  assert(work_group_size > 2 &&
         "Not all test checks hold when the work-group size is not greater "
         "than 2");

with

    if !( work_group_size > 2)
      SKIP(  "Not all test checks hold when the work-group size is not greater "
         "than 2")

About supporting WG size of one, I agree it's maybe not worth the trouble of changing all the logic of this tests.