[TOC]
This document also explains in a high-level manner how different components are organized, the principles behind them and the conventions we followed. The document assume that you have basic understanding of asynchronous digital circuits at the behavioral level abstraction.
Handshake/dataflow IR is describes independent, unsynchronized processes communicating data through First-in First-out (FIFO) communication channels. This can be implemented in many ways, such as using synchronous logic, or with processors.
MLIR is a common infrastructure to build your own specific IR to target different architectures and needs. We use MLIR because of its extensibility. We can apply the various transformations and optimization of MLIR on this IR. We can also lower the std MLIR produced by different frontends to Handshake IR.
TensorFlow LLVM Pytorch
| | |
|-----------------------------------|
| MLIR |
| ----------------- |
| | opt/transform | |
| ----------------- |
| |
| ----------------- |
| | opt/transform | |
| ----------------- |
| |
|-----------------------------------|
| | | |
GPU LLVM Affine **Dataflow**
Simple Handshake IR snippet for add function look like this -
handshake.func @simple_addi(%arg0: index, %arg1: index, %arg2: none, ...) -> (index, none) {
%0 = addi %arg0, %arg1 : index
handshake.return %0, %arg2 : index, none
}
It accepts two input streams (modeled as MLIR operands) and produces one output stream (modeled as an MLIR result).
The Handshake dialect adopts the following conventions for IR:
- The prefix for all Handshake types and operations are
handshake..
MLIR Handshake Dialect-slides by Stephen Neuendorffer (Xilinx) + Lana Josipović (EPFL)
[include "Dialects/HandshakeOps.md"]