Comparing actual tensor against empty reference tensor passes

[hjjvandam]

By accident I discovered that this test:

#include <catch2/catch.hpp>
#include <simde/simde.hpp>

TEST_CASE("allclose") {
    using tensor_t = simde::type::tensor;
    tensor_t t1{{1.0, 1.0}, {1.0, 1.0}};
    auto l = [&](const auto& lo, const auto& hi, auto* tile) {
    };
    using field_t   = typename simde::type::tensor::field_type;
    using shape_t   = typename simde::type::tensor::shape_type;
    using extents_t = typename shape_t::extents_type;
    extents_t extents{2,2};
    tensor_t t2(l,
                std::make_unique<shape_t>(extents),
                tensorwrapper::tensor::default_allocator<field_t>());
    SECTION("reference tensor empty, actual non-empty") {
        REQUIRE_FALSE(tensorwrapper::tensor::allclose(t1,t2));
    }
    SECTION("actual tensor empty, reference non-empty") {
        REQUIRE_FALSE(tensorwrapper::tensor::allclose(t2,t1));
    }
} 

fails in the instance where the reference tensor is empty. I.e. adding this test to SCF/tests I got

$ ./test_scf "allclose"
Filters: allclose

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
test_scf is a Catch v2.13.9 host application.
Run with -? for options

-------------------------------------------------------------------------------
allclose
  reference tensor empty, actual non-empty
-------------------------------------------------------------------------------
/workspaces/NWChemEx-Project.nosync/SCF/tests/allclose.cpp:16
...............................................................................

/workspaces/NWChemEx-Project.nosync/SCF/tests/allclose.cpp:17: FAILED:
  REQUIRE_FALSE( tensorwrapper::tensor::allclose(t1,t2) )
with expansion:
  !true

===============================================================================
test cases: 1 | 1 failed
assertions: 2 | 1 passed | 1 failed

I would have thought that both tests should have detected that the tensors are different. Am I just mistaken?

[ryanmrichard]

@hjjvandam If I'm reading this correctly you are not mistaken. allclose should be symmetrical.

[jwaldrop107]

I've looked into this some now. The way that the lambda constructor works deep down is that a TA tile is initialized and set to zero, then the lambda is used to assign the correct values to the tile. So, this construct results in a (sparse) tensor that is only zeros. That also explains the apparent asymmetry. One test is effectively "(1 - 0) == 1" and the other is "(0 - 1) == 1."

[ryanmrichard]

I thought all close was basically (t1 - t2) == 0 (not ==1 as you put)?

[jwaldrop107]

Yeah, you're right. That last part of my comment wasn't thought all the way out. I'm still working through the asymmetry.

[jwaldrop107]

@ryanmrichard, is this the same as this note to this issue

[ryanmrichard]

@jwaldrop107 probably.

[hjjvandam]

The test close to the note Ryan found is exactly the test I found to fail. So this issue is the same.