IntelPython / dpnp / 28230601268

78%

push

github

Optimize FFT logic for complex strided input to avoid oversized memory allocation (#2939)

This PR fixes an issue discovered while implementing #2927 where several
FFT tests started failing after adding validation for stride
configurations with oversized memory footprints.

The problem was that FFT could allocate output arrays using the same
strided layout as a non-contiguous input. For inputs such as `a[::2`,
this resulted in oversized allocations followed by an additional copy to
a contiguous array.

This fix checks whether the memory footprint implied by the input
strides exceeds the number of elements in the array. If so, the input is
copied to a contiguous layout before configuring the FFT descriptor
allowing oneMKL FFT to produce a contiguous output directly

By avoiding oversized allocations and the extra copy this significantly
improves the performance of all FFT operations in dpnp with complex
strided inputs

Additionally, the cuFFT C-contiguous requirement is now enforced for all
code paths. Previously it was only applied within `batch_fft` branch
which was a latent bug for 1D non-batch inputs on CUDA.

1532 of 2854 branches covered (53.68%)

Branch coverage included in aggregate %.

5 of 5 new or added lines in 1 file covered. (100.0%)

25977 of 32382 relevant lines covered (80.22%)

8020.04 hits per line