Overview

Matrix runs above the operating system (OS) and below service applications. It provides unified standard application programming interfaces (APIs) for applications, shielding differences between OSs. It also provides multi-node scheduling and multi-process management. In addition, Matrix can set up and run service processes based on configuration files and summarize statistics.

The Matrix logic consists of three parts: Matrix Agent (running on the host side), Matrix Daemon (running on the device side), and Matrix Service (running on the device side).

• The host side refers to the x86 server, ARM server, or Windows PC connected to the device side. The host side uses the neural-network (NN) computing capability provided by the device side to implement services.
• The device side refers to an Ascend 310–powered hardware device that provides the host side with the NN computing capability over the PCIe interface.
• The digital vision pre-processing (DVPP) implements video encoding and decoding, JPEG encoding and decoding, PNG decoding, and vision pre-processing.
• Framework is a deep learning framework that can convert models under open source frameworks such as Caffe into models supported by Mind Studio.
• The cube-based computing engine (CCE) acceleration library accelerates upper-layer applications (frameworks or applications provided for machine learning) through APIs.
• Runtime runs in the application process space and provides Ascend 310–powered devices with various functions, such as memory management, device management, stream management, event management, and kernel execution.

Matrix is a general-purpose service process execution engine that consists of the following parts:

• Matrix Agent: It runs on the host side and provides the following functions:
  • Interacts with the host application for controlling commands and processing data.
  • Implements interprocess communication (IPC) with the device side.
• Matrix Daemon: It runs on the device side and provides the following functions:
  • Set ups a service process based on the configuration file.
  • Destroys the service process and reclaims resources according to commands.
  • Serves as the Daemon process, which is used to start the Matrix process.
• Matrix Service: It runs on the device side and provides the following functions:
  • Implements media preprocessing by using the DVPP APIs called by engines.
  • Implements model inference by using the AIModelManger APIs called by engines.

You can view the API definition files in ddk/include/inc/hiaiengine under the installation directory of the device development kit (DDK). If you install both Mind Studio and the DDK by following the installation wizard, you can log in to the Mind Studio server as the Mind Studio installation user and view the API definition file in ~/tools/che/ddk/ddk/include/inc/hiaiengine. For details about the mapping between the definition files and the APIs, see the API description.