Concat

API reference: C, C++

The concat primitive concatenates \(N\) tensors over concat_axis (here designated as \(C\) axis) and defined as:

\[ dst(\overline{ou}, c, \overline{in}) = src_i(\overline{ou}, c', \overline{in}), \]

where \(c = C_1 + .. + C_{i-1} {}_{} + c'\).

The concat primitive doesn't have a notion of forward or backward propagations. The backward propagation for the concatenation operation is simply an identity operation.

Implementation Details

General Notes

1. The \(dst\) memory format can be either specified by a user or derived by the primitive. The recommended way is to allow the primitive to choose the most appropriate format.
2. The concat primitive requires all source and destination tensors to have the same shape except for the concat_axis. The destination dimension for the concat_axis must be equal to the sum of the concat_axis dimensions of the sources (i.e. \(C = \sum_i C_i\)). Implicit broadcasting is not supported.

Data Types Support

The concat primitive supports arbitrary data types for source and destination tensors according to the Data Types page. However, it is required that all source tensors are of the same data type (but not necessarily matching the data type of the destination tensor).

Data Representation

The concat primitive works with arbitrary data tensors. There is no special meaning associated with any logical dimensions.

Post-ops and Attributes

The concat primitive doesn't support any post-ops or attributes.

Implementation Limitations

1. The primitive works with blocked memory formats, such as plain formats dnnl_nchw, dnnl_nhwc, and blocked formats dnnl_nChw16c, dnnl_nCdhw8c that appear in convolutions. The primitive does not support non-blocked formats that are typically used in prepacked weights, such as:
   • Winograd format dnnl_format_kind_wino,
   • RNN format dnnl_format_kind_rnn_packed, or
   • Blocked format with attached compensation (dnnl_memory_extra_flag_compensation_conv_s8s8), that is used in s8s8 convolutions (see Int8 Computation Aspects).
2. Refer to Data Types for limitations related to data types support.

Performance Tips

1. Whenever possible, avoid specifying the destination memory format so that the primitive is able to choose the most appropriate one.
2. The concat primitive is highly optimized for the cases in which all source tensors have same memory format and data type matches the destination tensor data type. For other cases, more general but slower code is working. Consider reordering sources to the same data format before using the concat primitive.