[DFT][cuFFT] Fix "Invalid strides" for shapes with a unit last dimension (#631)

[zjin-lcf]

Summary

Fixes #631.

oneapi::math::dft on the cuFFT backend throws dft/backends/cufft/commit: Invalid strides for multi-dimensional transforms whose last dimension has extent 1 (e.g. {3,1}, {2,1}, {5,1}, {2,3,1}). The same descriptors work on the Intel (MKL) backends, and the equivalent transform works when calling cuFFT directly.

Root cause

In src/dft/backends/cufft/commit.cpp, the backend uses std::min_element to find the dimension with the smallest stride (which becomes the cuFFT istride/ostride, i.e. the innermost dimension).

When a dimension has extent 1, its default stride ties with its neighbour's stride. std::min_element returns the first element among ties, so an outer dimension is mis-identified as the innermost one. This triggers the dimension-swap heuristic, producing a dimension/stride ordering that the check_stride_validity() pre-check then rejects — computing inner_n (e.g. 3) <= inner_stride (1) as false for both directions and throwing Invalid strides. The underlying cufftPlanMany call that would have been issued is actually valid.

There is also latent undefined behaviour: the swap logic computes the selected index from a_min after vec_a.erase(a_min) has invalidated that iterator. This is why the failure is compiler-dependent (it reproduces with icpx, but a recent open-source DPC++/clang build happens to take a different, passing path).

Fix

• Search for the smallest stride in reverse, so ties resolve to the innermost (highest-index) dimension. This avoids the spurious dimension swap and lets valid transforms through.
• Capture the selected stride index before erase(), removing the invalidated-iterator UB.

No behavioural change for shapes without unit dimensions: for strictly decreasing default strides the reverse search selects the same (last) element as before.

Test plan

Validated end-to-end on an NVIDIA sm_103 device, CUDA 13.2 / cuFFT 12.2, using an open-source DPC++ toolchain (intel/llvm nightly with the CUDA backend). Built oneMath with -DENABLE_CUFFT_BACKEND=ON -DTARGET_DOMAINS=dft.

With the fix, the reproducer from #631 and related shapes all pass:

shape {3,1}   : OK -> [(3,0) (0,0) (0,0)]
shape {2,1}   : OK -> [(2,0) (0,0)]
shape {5,1}   : OK -> [(5,0) (0,0) (0,0) (0,0) (0,0)]
shape {2,3,1} : OK -> [(6,0) ...]
shape {3,4}   : OK   (regression check)
shape {2,3,4} : OK   (regression check)
shape {8}     : OK   (regression check)

• Confirmed direct cuFFT (cufftPlan2d(3,1) and the exact cufftPlanMany params oneMath generates) succeed on the B300 — the failure is in the pre-check, not cuFFT.
• Standalone comparison of the stride algorithm: original logic REJECTS {3,1}, {2,1}, {5,1}, {2,3,1}; fixed logic ACCEPTS them; regression shapes unchanged.
• End-to-end build + run on B300: all shapes above pass.
• CI: DFT cuFFT functional tests.

Suggestion for maintainers: consider adding a DFT test case covering shapes with a unit last dimension to guard against regressions.