static_op_mem (Static Graph Operator Memory)

The static graph operator memory does not contain the timeline information. The summary information is summarized in the static_op_mem_*.csv file.

Applicability

Product

Supported

Atlas 350 Accelerator Card

Atlas A3 training product / Atlas A3 inference product

Atlas A2 training product / Atlas A2 inference product

Atlas 200I/500 A2 inference product

Atlas inference product

Atlas training product

Atlas 350 Accelerator Card

Data Description of the static_op_mem_*.csv File

The following shows the file content format.

Figure 1 static_op_mem_*.csv

In single-operator scenarios, the ACL_PROF_TASK_MEMORY data collection function is enabled by a call to aclprofCreateConfig. The data is reported only in the model compilation phase. You can use this file to view the memory allocation of operators in each subgraph in static graph scenarios.

In the static graph scenario, the graph ID is used to distinguish different computational graphs. In the dynamic subgraph scenario, the model name (root node name) is used to distinguish different subgraphs.

Table 1 Field description

Field

Description

Device_id

Device ID.

Op Name

Operator name. The last line TOTAL shows the total allocated memory.

Model Name

Name of the root node of a static submap. If the value is 0, the graph is a static graph and the static subgraph does not exist. If a static submap exists, the root node name is displayed.

Graph ID

Graph ID. Each graph ID corresponds to a computational graph.

Node Index Start

Logical time of operator memory allocation.

Node Index End

Logical time of operator memory deallocation. When displayed as 4294967295, it indicates the maximum value of the time for operator memory allocation, that is, the operator memory deallocation time is the end time of the computational graph's lifetime.

Size(KB)

Size of the allocated memory, in KB.