Documentation

Analysis Result Display

MindStudio Advisor analyzes the files prepared in Input Data Description and exports the results of all functions at a time. Therefore, if no file is prepared and saved in the specified directory, the output function results are empty.

Overview Page

Figure 1 Summary page of the analysis result (Model Performance Report)
Table 1 Fields in Model Performance Report

Field

Description

Model Performance Report

Model performance analysis report.

Model Performance

Model performance. The value can be Good or Bad. The performance is calculated based on the total profile data.

Collection Info

Summary information.

Cube Throughput

Cube throughput, in GOps.

Vector Throughput

Vector throughput, in GOps.

Aicore Time

AI Core execution duration, in μs.

Task Duration

Task execution time, in μs.

Avg BlockDim Usage

Average BlockDim usage, that is, the average number of cores used during operator execution. This field reflects the processor usage.

Chip Utilization

Processor usage. Value 80: excellent, green; value less than 80, poor, red. The value is calculated based on the value of Pipeline Bound.

Pipeline Bound

Pipeline usage.

Cube Ratio

Cube usage.

Vector Ratio

Vector usage.

Scalar Ratio

Scalar usage.

MTE1 Bound

MTE1 bottleneck.

MTE2 Bound

MTE2 bottleneck.

MTE3 Bound

MTE3 bottleneck.

Tiling Strategy

Data tiling policy for memory read amount. Value ≥ 80: excellent, green; value < 80, poor, red. The value is calculated based on the value of Memory Redundant.

Memory Redundant

Memory redundancy.

Real Memory Input(GB)

Actual memory read amount, in GB.

Real Memory Output(GB)

Actual memory write amount, in GB.

Theory Memory Input(GB)

Theoretical memory read amount, in GB.

Theory Memory Output(GB)

Theoretical memory write amount, in GB.

Memory Read Redundant

Memory read redundancy coefficient = Actual memory read amount/Theoretical memory read amount

Memory Write Redundant

Memory write redundancy coefficient = Actual memory write amount/Theoretical memory write amount
Figure 2 Summary in the analysis result (Computational Graph Optimization > UB Fusion Recommendation)
Table 2 Fields in Computational Graph Optimization > UB Fusion Recommendation

Field

Description

Computational Graph Optimization

Computational graph optimization: suggestion on operator fusion recommendation of MindStudio Advisor. Operators that can be fused are displayed in different rows. You can click See More... in the upper right corner to view the specific operators that can be fused.

UB fusion operators need to be optimized

Operators for UB fusion.

UB Fusion Recommendation

UB fusion recommendation. Operators that can be fused are displayed in the lower part.

Fusion Type

Type of an operator that can be fused.

Fusion Operator Detail

Details about operators that can be fused. Operator names are separated by commas (,).

Fusion Operator Duration(us)

Execution time of the operators that can be fused. The unit is μs.
Figure 3 Summary in the analysis result (Computational Graph Optimization > AIPP Fusion Recommendation)
Table 3 Fields in Computational Graph Optimization > AIPP Fusion Recommendation

Field

Description

Fuse Cast/TransData with Conv needs to be optimized

Operators that require AIPP input-layer fusion

AIPP Fusion Recommendation

AIPP fusion recommendation. Operators that can be fused are displayed in the lower part.

Fusion Operator Detail

Details about operators that can be fused. Operator names are separated by commas (,).

Fusion Operator Duration(us)

Execution time of the operators that can be fused. The unit is μs.
Figure 4 Summary in the analysis result (Computational Graph Optimization > TransData Fusion Recommendation)
Table 4 Fields in Computational Graph Optimization > TransData Fusion Recommendation

Field

Description

TransData fusion operators need to be optimized.

Total time of TransData task is xx(us), accounted for xx% of the total task.

TransData operators that need to be optimized.

Total duration of TransData operators: xx μs, accounting for xx% of the total task duration

TransData Fusion Recommendation

TransData operators that are recommended to be eliminated. Operators that can be eliminated are displayed in the lower part.

Reshape_Ops_Interrupts_Format

Reshape_Ops_Interrupts_Format tuning suggestions.

Attempt to modify the model to avoid discontinuous operations.

Try to avoid nonconsecutive operations without affecting the accuracy.

Modify the model and use clone and continuous operations to break the combination of multiple non-consecutive operations.

Use clone and contiguous to disconnect multiple nonconsecutive operations.

Other_Transform

Other_Transform tuning suggestions.

It is a reasonable scenario that transdata operation exists, for example, 4D to 5D before Conv2D.

This is a reasonable TransData scenario. For example, before running the Conv2D operator, change the format from 4D to 5D.

Op Name

Operator name.

Task Duration(us)

Operator execution duration.

Input Formats

Input formats.

Output Formats

Output formats.
Figure 5 Summary in the analysis result (Computational Graph Optimization > L2Cache Fusion Recommendation)
Table 5 Fields in Computational Graph Optimization > L2Cache Fusion Recommendation

Field

Description

L2 fusion operators need to be optimized

Operators for L2 fusion

L2Cache Fusion Recommendation

L2 cache fusion recommendation. Operators that can be fused are displayed in the lower part.

Fusion Operator Detail

Details about operators that can be fused. Operator names are separated by commas (,).

Fusion Operator Duration(us)

Execution time of the operators that can be fused. The unit is μs.
Figure 6 Summary in the analysis result (Roofline)
Table 6 Roofline field description

Field

Description

Roofline

Information about top 3 operators for Roofline model-based operator bottleneck identification and tuning suggestion. You can click See More... in the upper right corner to view the detailed result.

Top Ops

The first three operators. The basic information about the first three operators that can be tuned of the Roofline model is displayed in the lower part.

Op Name

Operator name.

Aicore Time(us)

AI Core running duration (μs).

Bottleneck pathway

Bottleneck channel, that is, the shortest channel from the working point to the Roofline.

Bottleneck Rate

Bottleneck rate, that is, the percentage of the working point to the Roofline upper limit.

Bottleneck Pipeline

Pipeline with the highest proportion.

Pipeline Rate

Highest pipeline rate.

Bound Type

Bottleneck type.

Task Duration Ratio(%)

Task duration percentage.
Figure 7 Summary in the analysis result (Model Graph Optimization)
Table 7 Fields in Model Graph Optimization

Field

Description

Model Graph Optimization

Model tuning suggestions.

Top AICPU Ops

Operator list (sorted by time consumption in descending order)

Operator name

Operator name.

Task Start Time

Start time of a task.

Task Duration

Task duration.

Task Duration Ratio

Task duration percentage.

Recommendations of aicpu operations optimization

Tuning suggestions for AI CPU operators.
Figure 8 Summary in the analysis result (Operating Environment)
Table 8 Fields in Operating Environment

Field

Description

Operating Environment

Operating environment.

Host Operating System

Host OS.

Host Computer Name

Name of the computer on the host.

CPU Name

CPU name.

CPU Name Type

CPU name type.

Control CPU Type

Ctrl CPU type.

Control CPU Number

Number of Ctrl CPUs.

TS CPU Number

Number of TS CPUs

AI CPU Number

Number of AI CPUs

Computational Graph Optimization Page (Output of Operator Fusion Recommendation)

The output on the Computational Graph Optimization page contains two functions: operator fusion recommendation and TransData operator recognition. This section describes the operator fusion recommendation function.

Figure 9 Analysis result on Computational Graph Optimization (operator fusion recommendation function)

The preceding picture is described as follows:

  • Area 1 is the model visualization page, which displays all operators in the model and highlights the operators that can be fused.
  • Area 2 displays the operator details when you click an operator on the visualization page.
  • Area 3 is the operator search area, which displays all operators in the model by name. You can search for a specific operator name. You can also click the operator name in the search area to view its detailed information.
  • Area 4 displays analysis result. For details about the fields, see Table 2, Table 3, and Table 5.
    • The tab pages display the analysis results of the UB, AIPP, and L2Cache fusion functions of operator fusion recommendation. You can click a function tab to view the corresponding analysis result.
    • The list displays operators that can be fused. Each row displays a piece of fusion information, in which operator names are separated by commas (,). If the content is too long, hover the cursor over the corresponding row to display the detailed operators. After clicking the corresponding row, you will be redirected to the corresponding operators in the visualization view.

    The suggestion "For more case references, please visit here." in Figure 9 indicates that you can click Analysis Example of UB Operator Fusion Recommendation to access the analysis result output by the operator fusion recommendation function. To click the hyperlink on the result page, install the Firefox browser on the Linux server in advance.

Computational Graph Optimization Page (Output of TransData Operator Recognition)

The output on the Computational Graph Optimization page contains two functions: operator fusion recommendation and TransData operator recognition. This section describes the TransData operator recognition function.

Figure 10 Analysis result on Computational Graph Optimization (TransData operator recognition function)

The preceding picture is described as follows:

  • Area 1 is the model visualization page, which displays all operators in the model and highlights the operators that can be eliminated.
  • Area 2 displays the operator details when you click an operator on the visualization page.
  • Area 3 is the operator search area, which displays all operators in the model by name. You can search for a specific operator name. You can also click the operator name in the search area to view its detailed information.
  • Area 4 displays analysis result. For details about the fields, see Table 4.
    • On the tab pages, you can click TransData Fusion Recommendation to view the analysis result of TransData operator recognition.
    • The drop-down list displays tuning suggestions in different dimensions.
    • The list displays operators that can be eliminated. Each row displays a piece of fusion information. After clicking the corresponding row, you will be redirected to the corresponding operators in the visualization view.

Roofline Page (Output of Roofline Model-based Operator Bottleneck Identification and Tuning Suggestion)

Figure 11 Roofline display of the analysis result

The preceding picture is described as follows:

  • Area 1 displays the channels of the Roofline model in the analysis result of MindStudio Advisor.
    • Each item in area 1 corresponds to a working point in area 4. If an item is selected, the corresponding working point is displayed in area 4. By default, all working points are selected.
    • The mapping between the channel colors in area 1 and the working points and line colors in area 4 is as follows:
      • Cube: MTE1-blue; MTE2-green; MTE3-red
      • Vector: MTE2-green; MTE3-red; PIPE_V-yellow.
  • Area 2 displays the top N operator information. You can choose to display top 3, top 5, or top 10 operators. You can also search for an operator. For details about the fields, see Table 6. You can click an operator name to view the suggestions provided by MindStudio Advisor in area 3, and the working point and bottleneck information of the operator in area 4.
  • Area 3 displays the suggestions on a specific bottleneck operator. You need to click the operator name in area 2.

    The fourth suggestion in Figure 11 "For more case references, please visit here.", indicating that you can click Analysis Example of Roofline Model Tuning to access the analysis result output by the Roofline model-based operator bottleneck identification and tuning suggestion function. To click the hyperlink on the result page, install the Firefox browser on the Linux server in advance.

  • Area 4 displays the analysis result of the Roofline model in a coordinate axis. The computing power, theoretical computing power, and bandwidth of operators in the Cube and Vector compute units are displayed.
    • In the coordinate axis, the unit of the horizontal coordinate is Ops/Byte, indicating the computing intensity. The value indicates the number of operations for each 1-byte data movement. A larger value indicates a higher memory movement utilization rate.
    • The unit of vertical coordinates is Tops/s, indicating the operation speed. A larger value indicates faster operations.
    • The line turning part divides the Roofline model into two parts. The slashes indicate Memory Bound (memory limit), and the horizontal lines indicate Compute Bound (computing limit). The closer an actual working point (color point in the figure) to the slash of the corresponding color, the more serious the Bound is, which is the main bottleneck.
    • The display effect is determined by the selection in area 1. The slash of Multiple roofline overlap indicates that the lines overlap because the output bandwidths of multiple channels are the same. You can deselect the channels with the same output bandwidth in area 1 to determine the channel represented by the slash.
    • The information about a working point is displayed when the cursor is placed on it.
    • Operators with bottlenecks are ranked from top 1 to top 3 in descending order of working points.

Model Graph Optimization Page (Output of Timeline-based AI CPU Operator Tuning)

Figure 12 Model Graph Optimization display of the analysis result

The preceding picture is described as follows:

  • Area 1 displays the execution time and tuning suggestions of the top 3 AI CPU operators after serial waiting based on the analysis result of Timeline-based AI CPU operator tuning of MindStudio Advisor.
  • Area 2 displays AI CPU operator tuning suggestions.
  • Area 3 displays the Timeline view of Profiling.

The fifth suggestion in area 2 in Figure 12 is "For more case references, please visit here.", indicating that you can click Analysis Sample of Timeline-Based Tuning of AI CPU Operators to access the analysis result output by the timeline-based tuning of AI CPU operators function. To click the hyperlink on the result page, install the Firefox browser on the Linux server in advance.

Parent topic: MindStudio Advisor Entry