AVX512 support machine cannot resize uint8 image with BILINEAR interpolation as it is

[vinnamkim]

🐛 Describe the bug

Please see this line on the main branch.

vision/torchvision/transforms/v2/functional/_geometry.py

Line 200 in 5181a85

return "AVX2" in torch.backends.cpu.get_cpu_capability()

It decides whether the image resize can be done for uint8 datatype with native CPU.

However, the return type of torch.backends.cpu.get_cpu_capability() is not a kind of set or sequence of strs. It returns str, https://pytorch.org/docs/2.2/backends.html#torch.backends.cpu.get_cpu_capability. Therefore, on a AVX512 support machine, torch.backends.cpu.get_cpu_capability() returns "AVX512" and it leads to float32 casting at

vision/torchvision/transforms/v2/functional/_geometry.py

Lines 256 to 258 in 5181a85

	need_cast = dtype not in acceptable_dtypes
	if need_cast:
	image = image.to(dtype=torch.float32)

although this machine can support AVX2.

Versions

torch>=2.1 and torchvision>=0.16.1