torch.fft.irfft(torch.fft.rrft) is not identity on MPS

[cjolif]

🐛 Describe the bug

On CPU, call torch.fft.irfft(torch.fft.rrft) gives back the initial signal (as expected) it is not on MPS.

t = torch.linspace(0, 1, 5, device="cpu")
T = torch.fft.rfft(t)
roundtrip = torch.fft.irfft(T, t.numel())
print(t, roundtrip)
torch.testing.assert_close(roundtrip, t, check_stride=False)

t = torch.linspace(0, 1, 5, device="mps")
T = torch.fft.rfft(t)
roundtrip = torch.fft.irfft(T, t.numel())
print(t, roundtrip)
torch.testing.assert_close(roundtrip, t, check_stride=False)

gives

tensor([0.0000, 0.2500, 0.5000, 0.7500, 1.0000]) tensor([0.0000, 0.2500, 0.5000, 0.7500, 1.0000])
tensor([0.0000, 0.2500, 0.5000, 0.7500, 1.0000], device='mps:0') tensor([0.0000, 1.0000, 0.7500, 0.5000, 0.2500], device='mps:0')
Traceback (most recent call last):
  File "/Users/cjolif/Personnal/Playground/torch_playground.py", line 48, in <module>
    torch.testing.assert_close(roundtrip, t, check_stride=False)
  File "/Users/cjolif/Personnal/Playground/.venv_main/lib/python3.10/site-packages/torch/testing/_comparison.py", line 1523, in assert_close
    raise error_metas[0].to_error(msg)
AssertionError: Tensor-likes are not close!

Mismatched elements: 4 / 5 (80.0%)
Greatest absolute difference: 0.75 at index (1,) (up to 1e-05 allowed)
Greatest relative difference: 3.0 at index (1,) (up to 1.3e-06 allowed)

This was tested on both 2.3.1 and 2.4.0-rc

Versions

Collecting environment information...
PyTorch version: 2.4.0.dev20240519
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A

OS: macOS 14.5 (arm64)
GCC version: Could not collect
Clang version: 15.0.0 (clang-1500.3.9.4)
CMake version: Could not collect
Libc version: N/A

Python version: 3.10.14 (main, May 18 2024, 21:07:07) [Clang 15.0.0 (clang-1500.3.9.4)] (64-bit runtime)
Python platform: macOS-14.5-arm64-arm-64bit
Is CUDA available: False
CUDA runtime version: No CUDA
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Apple M3

Versions of relevant libraries:
[pip3] numpy==1.26.4
[pip3] torch==2.4.0.dev20240519
[pip3] torchaudio==2.2.0.dev20240519
[pip3] torchvision==0.19.0.dev20240519
[conda] Could not collect

cc @kulinseth @albanD @malfet @DenisVieriu97 @jhavukainen