stochastic failure in `test++`

[evaleev]

seemingly stochastic failures, in the same spot:
https://github.com/TAPPorg/reference-implementation/actions/runs/17947746998/job/51038316258
https://github.com/TAPPorg/reference-implementation/actions/runs/17947746998/job/51038316169
https://github.com/TAPPorg/reference-implementation/actions/runs/17947746998/job/51038316169

[evaleev]

@SpaceyLake seems that test_negative_strides in test++ fails, here: 6176ae3

[SpaceyLake]

I don't really understand why this happens. TBLIS seems to abort because of internal::are_compatible(oldlengths, oldstrides[i], lengths, *strides[i]), when TBLIS makes a fold with tensor B and D(C in TBLIS). I guess it deems indices, extents, and/or strides incompatible. It only seems to occur with negative strides on tensor D(C in TBLIS).

@devinamatthews do you have more insight into what might happen, what TBLIS might dislike about the input?

Here is the information Valgrind and stderr gives me about the termination:

internal::are_compatible(oldlengths, oldstrides[i], lengths, strides[i])
==46787==
==46787== Process terminating with default action of signal 6 (SIGABRT)
==46787== at 0x4FC4B2C: __pthread_kill_implementation (pthread_kill.c:44)
==46787== by 0x4FC4B2C: __pthread_kill_internal (pthread_kill.c:78)
==46787== by 0x4FC4B2C: pthread_kill@@GLIBC_2.34 (pthread_kill.c:89)
==46787== by 0x4F6B27D: raise (raise.c:26)
==46787== by 0x4F4E8FE: abort (abort.c:79)
==46787== by 0x4950034: tblis_abort_with_message(char const, ...) (basic_types.h:66)
==46787== by 0x4950073: tblis_check_assert(char const*, bool) (basic_types.h:72)
==46787== by 0x495E5BA: void tblis::fold<MArray::short_vector<long, 8ul, std::allocator > >(MArray::short_vector<long, 8ul, std::allocator >&, std::__cxx11::basic_string<char, std::char_traits, std::allocator >&, MArray::short_vector<long, 8ul, std::allocator >&, MArray::short_vector<long, 8ul, std::allocator >&) (tensor.hpp:278)
==46787== by 0x4970B02: tblis_tensor_mult (mult.cxx:123)
==46787== by 0x111C94: run_tblis_mult_s(int, long*, long*, float*, int, long*, int, long*, long*, float*, int, long*, int, long*, long*, float*, int, long*, int, long*, long*, float*, int, long*, float, float) (test.cpp:80)
==46787== by 0x126059: test_negative_strides() (test.cpp:3254)
==46787== by 0x110BF2: main (test.cpp:25)

Here is an example of tensor-structures that causes TBLIS to abort, to me they look correct:

Tensor A:
idx: 97
ndim: 1
extents: 1
strides: -1

Tensor B:
idx: 98 99 97
ndim: 3
extents: 3 2 1
strides: -1 -3 -6

Tensor C:
idx: 98 99
ndim: 2
extents: 3 2
strides: -1 -3

Tensor D:
idx: 98 99
ndim: 2
extents: 3 2
strides: -1 -3

[devinamatthews]

Please make an issue on TBLIS.

[SpaceyLake]

So I have looked into it a bit more. What I have found is that running this first set of tensor structures in TBLIS tensor mult gives the "internal::are_compatible(oldlengths, oldstrides[i], lengths, *strides[i])" abort. I don't see why it wouldn't work, but I don't rule out that I am doing something wrong. The structures are written idx, extent, stride

A:
a 1 -1

B:
b 3 1
c 2 3
a 1 6

C:
b 3 -1
c 2 -3

It works if either the strides for C are positive or the strides for B and C are negative. The sign of the strides for A doesn't seem to be relevant. This problem only seems to occur for some structures.
I'll create an issue on TBLIS for it.