Asight Compute quick start-ZHENWU PPU Cloud Service(ppu)-阿里云帮助中心

1. Overview

Asight Compute is a kernel profiler for PPU applications. It collects and displays metrics during kernel execution to help you optimize kernel performance. Asight Compute consists of two tools:

The acu command-line tool: Collects kernel metrics and generates reports. It runs on Linux.
The Asight Compute GUI tool: displays reports and supports the Windows and macOS platforms. The GUI tool provides multiple views for displaying metrics, such as a Roofline Chart, Bar Chart, and Memory Table. It also supports baseline comparison and is easy to use.

The typical workflow for Asight Compute is as follows:

Use the acu command-line tool to profile a target application and generate a report.
Open the report in the Asight Compute GUI tool for analysis.

2. Obtain and install

If you use an official image, the tool is pre-installed.

2.1 Configure the acu CLI

Set environment variables
Go to the PPU SDK directory and run the following script to set the required environment variables:
```
source envsetup.sh
```
Verify the environment
To view the acu help documentation, run the following command:
```
acu -h
```
Tip: You can also view the acu help documentation by running acu with no arguments.
To check the version of your installed acu tool, run the following command:
```
acu -v
```

2.2 Obtain the Asight Compute GUI

The Asight Compute GUI tool supports the following operating systems:

Windows 10, Windows 11
macOS 10.15 and later

Click the download link to go to the download page and select your required version. Download the .msi package for Windows or the .dmg package for macOS.

Note

You need an account and password to download the software package. Contact your customer success manager (PDSA) for access.

3. Quick start

In this guide, the device that runs the PPU application is the target machine, and the device used to view the report is the host machine.

The Asight Compute workflow is as follows: Use the acu command-line tool on the target machine to collect a report, and then use the Asight Compute GUI tool on the host machine to open, view, and analyze the report.

3.1 Collect a report with acu

Use the acu command-line tool to launch your target application and collect its performance data. After the application finishes, acu generates a report file.

The basic syntax for acu is: acu [options] <application> [application args]

For example, to collect performance data for all kernel launches in a target application, run the following command:

acu -o test_report -f python test_linear.py

-o test_report specifies the name of the output report. You do not need to specify the file extension.
-f forces an overwrite of the report file if it already exists.
python test_linear.py is the command to run the target application.

After the application finishes, acu generates the corresponding report file. In this example, the file is test_report.acurep.

Important

If acu reports the error "device is not ready for profiling", see Error: device is not ready for profiling.

acu provides options to specify which kernels to profile. Use -c, --launch-count to limit the number of kernel launches to profile, -s, --launch-skip to skip a specified number of kernels before collection begins, and -k, --kernel-name to filter kernels by name using a regular expression. For details, see Filter kernels.

To control the data collected for each kernel launch, use the --set option to specify a metric set. By default, acu collects sections associated with the default set. To view a list of available sets, use --list-sets. To specify an exact section, use the --section option with a *.section file identifier. To view a list of available sections, use --list-sections. You can collect a group of individual metrics with --metrics and query all available metrics with --query-metrics. For details, see Specify the metric sets to collect.

acu also supports multiple replay modes. For details, see Specify a replay mode.

In addition to saving results to a report file, acu can also print results to the console using different pages. These pages correspond to the views available in the GUI report. If you do not specify an output file with the -o option, acu prints the details page by default. To select a different page, or to print a page while also saving to a file, use the --page=<Page> option. The following pages are currently supported: details and raw.

3.2 View a report

Reports collected by acu have an .acurep file extension and can be viewed in the Asight Compute GUI tool. You can open a file in the GUI tool in any of the following ways:

From the main menu: File > Open...
Drag and drop the file into the main Asight Compute GUI window.
From the context menu: Right-click an empty area in the Project Explorer and select Open...
Launch the GUI tool from the command line with the report file as an argument. For example: acu-uireport.acurep

Report files (.acurep) are self-contained and can be shared with other users. However, the version of the Asight Compute suite used to view a report must be the same as or newer than the version used to collect it. The GUI tool displays a warning if you try to open a report that was collected with a newer version of the command-line tool.

For more details about the Asight Compute GUI tool, see Asight Compute GUI tool.