CUDA irfft may be doing unnecessary cloning of input

[ssnl]

Context:

1. In cuFFT,

   Out-of-place complex-to-real FFT will overwrite input buffer if custom strides are set by the user.

2. We therefore had this check in CuFFTPlanCache.h
   

   pytorch/aten/src/ATen/native/cuda/CuFFTPlanCache.h

   Lines 177 to 192 in 899a075

   	// Note that this is before the actual cloning. This is intentional so we can
   	// check for advanced data layout with complex-to-real transform. cuFFT
   	// out-of-place complex-to-real transforms with advanced layout may overwrite
   	// input, and we need to clone the input.
   	//
   	// This just needs contiguity in cases except for twosided real-to-complex
   	// transform where we won't have simple data layout as output is two sided.
   	//
   	// See NOTE [ cuFFT Embedded Strides ] in native/cuda/SpectralOps.cu.
   	bool simple_layout = !(!complex_input && complex_output && !onesided) && // not twosided R2C
   	(clone_input || input.is_contiguous()); // contiguous
   	if (!simple_layout && complex_input && !complex_output) {
   	clone_input = true;
   	simple_layout = true;
   	}

3. Users still reported irfft input being modified on T4 Multidimensional CUDA irfft modifies input #34551
4. An unconditional input cloning for irfft seemed to have fixed it Fix input overwriting in irfft #35219

We should figure out why the previous check doesn't detect all the cases, whether it is a bug in our check or in cuFFT. I don't have access to a T4 so I write this issue to document the situation in case anyone wants to take a look.

cc @ngimel @mruberry @peterbell10 @VitalyFedyunin

[peterbell10]

1. In cuFFT,

Out-of-place complex-to-real FFT will overwrite input buffer if custom strides are set by the user.

Today, that same documentation says:

Out-of-place complex-to-real FFT will always overwrite input buffer.

So, it seems this was just an error in the CuFFT documentation.

[ngimel]

Thanks for checking, @peterbell10 ! Closing the issue, the clone looks to be necessary.