Prepare Analysis Files

PyPTO Toolkit does not collect or generate PyPTO runtime records. It reads files already produced by the framework and visualizes them. More complete input data enables more detailed views.

Runtime files

File	Opens independently	Purpose
chip_swimlane_records.json	Yes	Provides task, core, scheduler, orchestrator, and timing information
serving-strace-swimlane.json	Yes	Provides the Serving task timeline and aggregates task count and duration by WorkerProcess
deps.json	Yes	Provides the task DAG, tensor information, and swimlane dependency lines
name_map*.json	Conditional	Maps Func IDs to function names and generates function performance data with a swimlane file
CPM_static*.json	No	Provides tasks from an existing static critical-path result
CPM_observed*.json	No	Provides tasks from an existing observed critical-path result

Recommended layout:

dfx_outputs/
├── chip_swimlane_records.json
├── deps.json
├── name_map_<timestamp>.json
├── CPM_static.json             # optional
└── CPM_observed.json           # optional

serving-strace-swimlane.json opens independently and does not require the related Chip Swimlane files listed above.

When you open a swimlane file, the extension looks for deps.json and name_map*.json in the same directory:

chip_swimlane_records.json
├── deps.json       → task dependencies, kernel_ids, and Func IDs
└── name_map*.json  → Func ID-to-function-name mapping

Missing related files have the following effects:

• without deps.json, the extension cannot fully recover Func IDs or display data-dependency lines;
• without name_map*.json, some tasks display only IDs instead of function names;
• without CPM files, the corresponding critical-path highlight mode is unavailable;
• with only name_map*.json and no swimlane file, the function performance view cannot calculate durations.

IR pass snapshots

IR Pass Trace requires a directory whose name starts with passes_dump and contains consecutively numbered Python snapshots:

passes_dump_<timestamp>/
├── 00_frontend.py
├── 01_PassA.py
├── 01_PassA.log       # optional; displays warnings for this pass
├── 02_PassB.py
└── 03_PassC.py

The following constraints apply:

• 00_frontend.py must exist;
• subsequent filenames must follow NN_<PassName>.py;
• pass numbering must start at 1 and remain consecutive, with no missing or duplicate numbers;
• a matching .log file is optional.

The same pass dump can also be processed by memory_map for Memory Reuse Analysis. The analysis uses the pass dump to show the address range and lifetime of each MemRef.