quadric.io SDK

Introduction

The quadric SDK provides tools and libraries which enable the scheduling and profiling of neural networks, and the development of high performance C++ data parallel kernels on the quadric Edge Processing Unit (EPU).

The quadric SDK consists of four major components, which are packaged into a single tool. The components are the following:

• quadric Scheduler: Enables inferencing and profiling of frozen neural network graph files. Exposed via Graph Mode (Neural Networks)

• quadric C++ API: C++ APIs used to develop high performance data parallel kernels. Exposed via Source Mode (C++ Kernels)

• quadric compiler: LLVM with a custom quadric EPU backend which compiles quadric C++ to quadric assembly.

• quadric simulator: Instruction set simulator that emulates various EPU configurations and outputs detailed profiling information.

Introduction

• quadric.io SDK
• EPU Architecture
  • A program through the lens of quadric Hardware
  • Common EPU Configurations
  • Memory Hierarchy
  • Core memory
  • Border array
• Installing the quadric SDK
  • Hardware Requirements
  • Software Requirements (Docker)
  • Installing or Upgrading the quadric SDK

quadric devkit

• Installing Devkit Drivers
  • Requirements
  • Steps

Using the SDK

• Graph Mode (Neural Networks)
  • Intro
  • The Configuration File
  • PyTorch Example
  • TensorFlow Example
• Source Mode (C++ Kernels)
  • Requirements
  • Options
  • Examples
• Profiling Information
  • Overview

quadric C++ API Usage

• Data Types
  • Tensors
  • qVar_t
  • FixedPoint
• Data Access Patterns
  • Data Movement
  • Data Sharing Between Cores
  • Data movement DDR <-> OCM
  • Data movement OCM <-> Array
• Debugging
  • debugPrint
  • debugOCMPrint

quadric C++ API Example Kernels

• Creating a Simple Kernel
  • Requirements
  • Getting Started
• Gaussian Blur
  • Concepts demonstrated
  • Gaussian blur algorithm
  • Writing the kernel routine
  • Initializing and calling the kernel
  • Further Improvements
  • Key concepts

quadric C++ API Reference

• DDR <-> OCM Data Fetching
  • memcpy
• OCM <-> Array Data Fetching
  • Fetching Data from OCM to the Array
  • Writing Data to OCM from the Array
  • Broadcasting Data From OCM to All Cores on the Array
• Math Library
  • qMath Functions
  • GEMM
• Neural Network Blocks
  • Fully Connected Layers
  • Upsampling
  • Convolutions
  • Pooling
  • Activation Functions
  • LSTM
• Data Serialization
  • readTensors
  • writeTensors
  • Example Usage
  • Writing binary tensors in Python (numpy) and reading them in a kernel
  • Parsing binary tensors created by a kernel using Python (numpy)
  • Parsing JSON serialized tensors in Python with the json_tricks library