vector_reduce_add

Vector Reduce Add:

A single AIE compute tile performs a simple reduction: it sums an N-element int32 input vector into a 1-element int32 output. Input data is brought to the local memory of the Compute tile via a Shim tile (default N = 1024), the reduction is performed in one kernel invocation, and the single output value is copied back to the Shim tile.

The design body is a single aie.iron.algorithms.reduce(reduce_add_vector, in_ty, out_ty) call; the algorithms library handles the ObjectFifo / Worker / Runtime plumbing for the reduce shape (whole-input single-kernel-call).

Source Files Overview

1. vector_reduce_add.py: An @iron.jit-decorated design that delegates its dataflow body to aie.iron.algorithms.reduce. Supports standalone (python3 vector_reduce_add.py) and compile-only (--xclbin-path / --insts-path, used by the Makefile) modes.

2. reduce_add.cc: A C++ implementation of a vectorized add reduction for AIE cores. The kernel uses the AIE API, documented here. Source: here.

3. test.cpp: C++ testbench. Loads the compiled XCLBIN, supplies input, runs on the NPU, and verifies the result.

Ryzen™ AI Usage

Standalone

python3 vector_reduce_add.py

-d npu2 for Strix; -n to override the input length.

Makefile + C++ testbench

make
make run

For NPU2 (Strix): make devicename=npu2 && make run devicename=npu2.