What is ROCm Systems Profiler?

ROCm Systems Profiler is designed for the high-level profiling and comprehensive tracing of applications running on the CPU or the CPU and GPU. It supports dynamic binary instrumentation, call-stack sampling, and various other features for determining which function and line number are currently executing.

ROCm Systems Profiler can emit results in several output formats, selected with the --output-format command-line argument. The detailed trace formats capture the full timeline of events, while the aggregated formats summarize high-level results:

• ROCm Profiling Data (rocpd): A detailed trace stored as a SQLite3 database (.db), selected with --output-format rocpd. This will soon be the default output format. You can view and analyze rocpd files with the ROCm Optiq tool.

• Perfetto: A detailed trace stored as a protocol buffer (.proto), selected with --output-format proto. Upload the Perfetto output files at ui.perfetto.dev to visualize the results in any modern web browser.

• Text: Aggregated high-level results as human-readable text files, selected with --output-format text (txt is an alias).

• JSON: Aggregated high-level results as JSON files for programmatic analysis, selected with --output-format json. The JSON output files are compatible with the hatchet Python package, which converts the performance data into pandas data frames and facilitates multi-run comparisons, filtering, and visualization in Jupyter notebooks.

Tokens are space- or comma-separated, so multiple formats can be requested at once, for example --output-format proto rocpd. For details on all formats, see Understanding the Systems Profiler output.

To use ROCm Systems Profiler for instrumentation, follow these two configuration steps:

1. Indicate the functions and modules to instrument in the target binaries, including the executable and any libraries

2. Specify the instrumentation parameters to use when the instrumented binaries are launched