Why Test Shapes/Strides in CubeCL Different Rank Tests?

[lucianyao]

Hi CubeCL/Burn community,
I’m puzzled by the test cases in crates/cubecl-core/src/runtime_tests/different_rank.rs. The kernel (kernel_different_rank) performs element-wise addition on tensors (lhs and rhs), but the computation seems unaffected by their shapes or strides. Here’s the kernel:

#[cube(launch)]
pub fn kernel_different_rank<F: Float>(lhs: &Tensor<F>, rhs: &Tensor<F>, output: &mut Tensor<F>) {
    output[ABSOLUTE_POS] = lhs[ABSOLUTE_POS] + rhs[ABSOLUTE_POS];
}

The tests (test_kernel_different_rank_first_biggest and test_kernel_different_rank_last_biggest) use different shapes/strides, e.g.:

lhs: [2, 2, 2] or [2, 4]
rhs: [8]
output: [2, 4] or [2, 2, 2]

However, ABSOLUTE_POS indexes linearly, and all tensors have 8 elements, so shapes/strides don’t impact the result. Additionally, client.create and client.read_one handle raw buffers, ignoring shapes/strides.
What are these tests verifying? Any insights would be appreciated!
Thanks,
Lucian

[laggui]

It seems that this test was added specifically for a previous issue with tensor metadata for CUDA.

The PR tracel-ai/cubecl#124 (linked to issue tracel-ai/cubecl#122) might provide more context.