.. _concat:

CONCAT
==============
Overview
----------
This mega block performs concatenation of up to four input layers in a channel-wise manner.

At its core, concat is a memory read–write operation without computation. Multiple input feature maps are simply arranged contiguously in the channel dimension.

Q-linear Concat
###############

Q-linear Concat, or Quantized-linear Concat, is a quantized version of the concat operator that additionally requires scaling each input tensor with its respective scale.



.. image:: _static/concat_block_diag.svg
   :width: 1000px
   :align: center

Architecture
##############

To implement the above-mentioned operator in the GATI infrastructure, two operations are required:

1.Scale multiplication

2.Channel-wise concatenation

1. Scale Multiplication
~~~~~~~~~~~~~~~~~~~~~~~~


Multiplication is an operation that demands a high number of resources in terms of DSPs and LUTs. Therefore, the design focuses on **reusing existing multiplication-capable operators**.

One such operator that performs scale multiplication natively is **ELTWISE**. Hence, the ELTWISE (addition) operator is reused for scale multiplication. Since the native behavior of ELTWISE addition is to add two scaled inputs, the second input is scaled with **zero**, which results in the required scaled input at the output.

.. image:: _static/concat_eltwise.svg
   :width: 600px
   :align: center

Thus, for every input feature map in a concat operation, there is a corresponding ELTWISE operation prior to it.  
For example, if there are **three input feature maps**, then **three ELTWISE addition layers** are required.

2. Channel-wise Concat
~~~~~~~~~~~~~~~~~~~~~~

The concat operation itself is straightforward. Conceptually, the process is as follows:

- Read data from DRAM for the first input feature map  
- Write it back to DRAM  
- Read the second input feature map  
- Write it to a **continuous address range following the first input**

.. image:: _static/concat_memory.svg
   :width: 600px
   :align: center

Implementation Details
----------------------

The concat operator consists of the following modules:

1. Concat Controller  
2. Concat Length Switcher  
3. Concat Address Switcher  
4. Top Concat  
5. Request Controller (Concat)

Like every other mega block, Concat has a **request controller** and a **memory read controller** for reading input data from DDR. However, a new feature is introduced here: the addresses to be read are **not continuous**.

To handle this, the request controller must read data from **multiple non-contiguous addresses**.

To manage this address switching, the **Concat Address Switcher** is designed. This module monitors the request address controller and, once the current feature map has been fully read, it updates the address for the next read request.

Stop Address Calculation
------------------------

The most critical computation in concat is the **STOP address calculation**.

GATICC provides the following instructions for each of the four feature maps:

1. **Start Address** – Start address for the current input  
2. **KN** – Number of kernels / channels  
3. **IH** – Input height / width  
   (It is assumed that inputs are square. Rectangular inputs will be supported in the future and will require ISA changes.)  
4. **IN-NUM** – Number of concats (out of the maximum four inputs)

Using these parameters, the offset and stop address are calculated as follows::

    offset_1 = CONCAT_KN_1 * CONCAT_IH_1 * CONCAT_IW_1;
    CONCAT_StopAdd_1 = offset_1 + CONCAT_StartAdd_1;

Concat Controller
-----------------

The Concat Controller performs a simple task: it reads **256-bit packets** from the DRAM FIFO and writes them to the DRAM output data aligner FIFO.

While doing so, it also calculates the **total number of elements written per layer**. This allows the system to detect when the current input feature map has finished and when the length should be switched to the next input feature map.

There is **no computation** in the concat mega block—only memory read and write operations, as described above.

Limitations and Assumptions
===========================

The number of input channels for each feature map **must be a multiple of NSA**.  
If this condition is not met, the concat operation will fail.