Documentation

Obtaining Input Data

Table 1 lists the input data required by MindStudio Advisor. You can provide one or more types (as a combination) of data. For details about the data required by each bottleneck identification model, see Mapping Between Analysis Functions and Inputs.

Table 1 Input data of MindStudio Advisor

Input Data

Storage Directory

Description

*.om

Offline OM model file

${data_path}/data/project
  • It can be specified in MindStudio Advisor Entry.
  • This file is required for timeline-based tuning of AI CPU operators, operator fusion recommendation, Roofline model-based operator bottleneck identification and tuning suggestion, and TransData operator recognition.
  • Ensure that the size of the offline OM model file is less than 1 GB. Otherwise, analysis by MindStudio Advisor cannot be performed.

*.cce

CCE code

${data_path}/data/project
  • It can be specified in MindStudio Advisor Entry.
  • It is used for Roofline model-based operator bottleneck identification and tuning suggestion.
  • Ensure that the size of a single CCE file is less than 100 MB and the total size of CCE files is less than 1 GB. Otherwise, the Advisor-based analysis cannot be performed.

task_time_*.json

Profiling Task Scheduler data

${data_path}/data/profiling/PROF_*/device_0
  • After MindStudio Advisor is started, the Profiling is automatically started to obtain data. No manual configuration is required.
  • It is used for timeline-based tuning of AI CPU operators and Roofline model-based operator bottleneck identification and tuning suggestion. For details about the function configurations, see MindStudio Advisor Entry.
  • Ensure that the size of the Profiling Task Scheduler data file is less than 100 MB. Otherwise, analysis by Advisor cannot be performed.

*.csv

Profiling summary file

${data_path}/data/profiling/PROF_*/device_0
  • After MindStudio Advisor is started, the Profiling is automatically started to obtain data. No manual configuration is required.
  • This file is required for the L2 fusion and TransData operator recognition functions of operator fusion recommendation. For details about the function configurations, see MindStudio Advisor Entry.
  • Ensure that the size of a single Profiling summary file is less than 100 MB and the total size of op_Summary files is less than 1 GB. Otherwise, the Advisor-based analysis cannot be performed.

info.json.0

Basic profiling information file.

${data_path}/data/profiling/PROF_*/device_0
  • After MindStudio Advisor is started, the Profiling is automatically started to obtain data. No manual configuration is required.
  • This file is required for the TransData operator recognition function. For details about the function configurations, see MindStudio Advisor Entry.
  • Ensure that the size of the basic profiling information file is less than 100 MB. Otherwise, analysis by Advisor cannot be performed.

core{id}_instr_popped_log.dump

core{id}_instr_log.dump

Operator simulation file. TMMdoel simulation is supported. {id} indicates the core ID.

{project_location}/out/model/tm/{opname}/{casename}
  • The path in the operator project is automatically identified and does not need to be manually specified.
  • It is used for operator tuning analysis. For details about the function configurations, see Operator Project Entry.
  • Ensure that the size of the operator simulation file is less than 100 MB. Otherwise, analysis by MindStudio Advisor cannot be performed.

Executable File

Executable file for executing the target project.
  • C++: {project_location}/out
  • Python: {project_location}/src
  • It can be specified in MindStudio Advisor Entry.
  • This file is used for precise analysis of models in user applications. It is required when the Inference App mode needs to be specified.
  • For details about how to obtain these software packages, see In the Linux Scenario.

Note: ${data_path} indicates the root path of the data directory.
Parent topic: Input Data Description