Some devices (e.g. CUDA, HIP) need to change some per-thread state, which must be later restored Most device calls must be between enter() and leave()

Converts accumulator to output data type, applies bias and activation

Original ISPC reference version; this always rounds ties up.

Issues all previously submitted commands (does not block)

Because formatting using c++ streams is stateful, drop down to C is required Alternatively we could use stringstream, but its performance is... not go

Returns the function handle for the given symbol

Returns the work-group size for the kernel, given the global size For some architectures, the global size may need to be adjusted

Returns the number of supported physical devices.

IContext

Support must be checked before getting the scratch size or submission

r""" interface of ms-ssim Args: X (torch.Tensor): a batch of images, (N,C,H,W) Y (torch.Tensor): a batch of images, (N,C,H,W)