msopprof Simulator Mode Performance Data

Overview

The msOpProf simulator mode performs instruction-level performance simulation of operators. After the collection is complete, a folder named OPPROF_{timestamp}_XXX is generated in the specified --output directory, containing the following performance data files.

Output directory structure:

OPPROF_{timestamp}_XXX
├── dump                                        // Folder for storing raw simulation dump data.
└── simulator                                   // Folder for storing dump data analysis results.
    ├── core0.veccore0                          // Data file directories for each core.
    │   ├── core0.veccore0_code_exe.csv         // Code line time consumption data file for this core.
    │   ├── core0.veccore0_instr_exe.csv        // Code instruction information file for this core.
    │   └── trace.json                          // Simulation instruction pipeline chart file for this core.
    ├── core0.veccore1
    │   ├── core0.veccore1_code_exe.csv
    │   ├── core0.veccore1_instr_exe.csv
    │   └── trace.json
    ├── ...
    ├── visualize_data.bin                      // Visualization file for simulation pipeline charts and hot spot functions.
    └── trace.json                              // Summary simulation instruction pipeline chart file for all cores.

NOTE:

Directories are organized in the format core*.veccore* (Vector Core) or core*.cubecore* (Cube Core) to store data for each compute unit.

In multi-operator scenarios, CSV file names include a timestamp suffix, for example, core*_code_exe_20240429111143146.csv.

In single-operator scenarios, visualize_data.bin and the summary trace.json are located in the root of the simulator/ directory.

Code Line Time Consumption Data Files

The code line time consumption data files are core*_code_exe.csv, where * represents the core number (0 to n). These files store the execution time consumption information for each code line on each compute unit, helping you quickly locate the most time-consuming parts of the code.

Figure 1 core*_code_exe.csv file

Table 1 Field description

Field	Description
code	Code line. The format is code file path:line number.
call_count	Number of calls to instructions involved in a code line.
cycles	Total number of cycles in which instructions involved in a code line are executed on the AI Vector Core/AI Cube Core.
running_time(us)	Valid execution time of a code line, in μs.

Code Instruction Information Files

The detailed code instruction information files are core*_instr_exe.csv, where * represents the core number (0 to n). These files store detailed execution information for each instruction on each compute unit, helping you identify the most time-consuming single instructions.

Figure 2 core*_instr_exe.csv file

Table 2 Field description

Field	Description
instr	Name of a code instruction.
addr	PC address corresponding to a code instruction.
pipe	PIPE type, including instruction queues (MTE1/MTE2/MTE3) and compute units (VEC/CUBE).
call_count	Number of times that an instruction is called.
cycles	Total number of cycles in which an instruction is executed on the AI Vector Core/AI Cube Core.
running_time(us)	Valid execution time of an instruction, in μs.
detail	Detailed parameters for executing an instruction, such as source/destination address, length, and stride of data transfer.

Visualization Data File (visualize_data.bin)

The visualize_data.bin file is used for visualizing simulation performance data and must be viewed through MindStudio Insight. After import, the following content can be displayed:

Feature	Description
Instruction pipeline chart	Displays timing relationships by instruction and associates with the call stack to quickly locate bottlenecks.
Operator code hot spot map	Displays the mapping between operator source code and instructions, as well as time consumption, helping you identify hot spot code.

Figure 3 Operator code hot spot map example

For details about the MindStudio Insight import operation, see the Importing Profile Data chapter in the MindStudio Insight User Guide.

Instruction Pipeline Chart File (trace.json)

The trace.json files are the raw data files for simulation instruction pipeline charts, including per-core sub-files (located under core*.veccore*/ or core*.cubecore*/ directories) and the summary file for all cores (located in the root of the simulator/ directory).

trace.json can be viewed in the following two ways:

Chrome browser: Enter chrome://tracing in the Chrome address bar and drag the trace.json file into the window for viewing.

Figure 4 Viewing the instruction pipeline chart in the Chrome browser

Use keyboard shortcuts (W: zoom in, S: zoom out, A: move left, D: move right) to navigate.
MindStudio Insight: Import trace.json into MindStudio Insight for visual presentation, including instruction pipeline timing and the execution status of each PIPE.

NOTE:

The per-core trace.json file displays the instruction pipeline for that specific core only.

The summary trace.json file displays the aggregated instruction pipeline for all cores.

If you only need to focus on the performance of specific operators, call the TRACE_START and TRACE_STOP APIs within a single core and add -DASCENDC_TRACE_ON to the compilation configuration file to generate pipeline chart information for the specified range. For details about these APIs, see the Operator Debugging APIs in the Ascend C Operator Development API.