ViewBackward

Autograd node for reshape, squeeze, and unsqueeze.

Forward reinterprets the flat element buffer under a new shape. The backend dispatcher's reshape creates a new Storage alias over the same physical buffer; no data movement occurs. Backward reshapes the incoming gradient back to the original input shape recorded in input_shapes_[0]; because reshape is its own inverse on contiguous buffers (element ordering is unchanged), the backward pass is simply another reshape call.

Math

$y_k = x_k \quad \text{for } k \in [0, \mathrm{numel}(x)),$ where indices are taken in row-major flat ordering. The mapping between multi-dim indices changes, but the linear element ordering does not.

Shape

input_shapes_[0] — shape of the input tensor before the reshape. out_shape_ — shape after the reshape (inherited from FuncOp). Both must have identical element counts.

See Also

reshape_op, squeeze_op, unsqueeze_op, contiguous_op.