RuntimeError: "log2" "_vml_cpu" not implemented for 'Half'

[turian]

🐛 Bug

On CPU, haven't tested on GPU

RuntimeError: "log2" "_vml_cpu" not implemented for 'Half'

RuntimeError: "log" "_vml_cpu" not implemented for 'Half'

RuntimeError: "exp" "_vml_cpu" not implemented for 'Half'

RuntimeError: "pow" not implemented for 'Half'

RuntimeError: "round" "_vml_cpu" not implemented for 'Half'

To Reproduce

Steps to reproduce the behavior:

torch.log2(torch.rand((3, 3), dtype=torch.float16))
torch.log(torch.rand((3, 3), dtype=torch.float16))
torch.exp(torch.rand((3, 3), dtype=torch.float16))
torch.pow(torch.rand((3, 3), dtype=torch.float16))
torch.round(torch.rand((3, 3), dtype=torch.float16))

RuntimeError: "log2" "_vml_cpu" not implemented for 'Half'

RuntimeError: "log" "_vml_cpu" not implemented for 'Half'

RuntimeError: "exp" "_vml_cpu" not implemented for 'Half'

RuntimeError: "pow" not implemented for 'Half'

RuntimeError: "round" "_vml_cpu" not implemented for 'Half'

Expected behavior

log2 and log and exp and pow and round should work for Half. exp2 does work tho.

Environment

• PyTorch Version (e.g., 1.0): 1.8.1
• OS (e.g., Linux): OSX
• How you installed PyTorch (conda, pip, source): pip3
• Build command you used (if compiling from source):
• Python version: 3.8.5
• CUDA/cuDNN version: n/a
• GPU models and configuration: CPU
• Any other relevant information: n/a

Related to #50789

cc @mruberry @rgommers @heitorschueroff

[imaginary-person]

pow would be enabled for Float16 & BFloat16 on CPU via #50999, but log_vml_cpu, log2_vml_cpu, exp_vml_cpu, and round_vml_cpu can't currently be enabled for Float16 due to lack of complete AVX2 vectorization support for Float16 in /aten/src/ATen/cpu/vec256. So even after #50999 would land, pow wouldn't support autograd for Float16 on CPU until more AVX2 support would be added for Float16.

However, these ops are currently supported for BFloat16 on CPU, and pow's support for it (including autograd) would be enabled soon on CPU. exp2 hasn't been enabled for BFloat16 on CPU or CUDA yet, but I opened #54794 for it.

These ops are already supported on CUDA for Float16.

[mruberry]

It'd be nice if we improved these error messages and better documented which dtypes these operations support.