merged[indices[m][i], ...] = data[m][i, ...]
```

Each `data[i].shape` must start with the corresponding `indices[i].shape`,
and the rest of `data[i].shape` must be constant w.r.t. `i`.  That is, we
must have `data[i].shape = indices[i].shape + constant`.  In terms of this
`constant`, the output shape is

    merged.shape = [max(indices) + 1] + constant

*   Tracing/timeline support for distributed runtime (no GPU profiler yet).
*   C API gives access to inferred shapes with `TF_GraphGetTensorNumDims` and
    `TF_GraphGetTensorShape`.
*   Shape functions for core ops have moved to C++ via
    `REGISTER_OP(...).SetShapeFn(...)`. Python shape inference RegisterShape
    calls use the C++ shape functions with `common_shapes.call_cpp_shape_fn`. A
    future release will remove `RegisterShape` from python.

          type_attr: "dtype"
        }
        attr {
          name: "dtype"
          type: "type"
        }
        attr {
          name: "shape"
          type: "shape"
          default_value {
            shape {
              unknown_rank: true
            }
          }
        }
      }
      op {
        name: "RestoreV2"
        input_arg {

by that little scattering Light which uses to be refracted irregularly
by the Inequalities of the Prism.

Yet instead of the Circular Hole F, 'tis better to substitute an oblong
Hole shaped like a long Parallelogram with its Length parallel to the
Prism ABC. For if this Hole be an Inch or two long, and but a tenth or
twentieth Part of an Inch broad, or narrower; the Light of the Image

2CCF          ; valid                                  # 4.1  COPTIC SMALL LETTER OLD COPTIC HA
2CD0          ; mapped                 ; 2CD1          # 4.1  COPTIC CAPITAL LETTER L-SHAPED HA
2CD1          ; valid                                  # 4.1  COPTIC SMALL LETTER L-SHAPED HA
2CD2          ; mapped                 ; 2CD3          # 4.1  COPTIC CAPITAL LETTER OLD COPTIC HEI