rocm-systems/projects/rocprofiler-compute/tools/config_management

ROCProfiler-Compute Configuration Management

This directory contains the authoritative configuration-management system for ROCProfiler-Compute analysis configurations.

It is designed to guarantee:

• Structural correctness across GPU architectures
• Deterministic deltas relative to a single latest architecture
• Byte-level immutability enforced via hashes
• Safe promotion of a new latest architecture with rollback
• CI enforcement of all invariants

All workflows are orchestrated by a single sequential driver script:

tools/config_management/master_config_workflow_script.py

Repository Layout

rocprofiler-compute/
├── src/rocprof_compute_soc/
│   └── analysis_configs/
│       ├── gfx908/
│       │   ├── 0000_top_stats.yaml
│       │   └── config_delta/
│       │       └── <latest_arch>_diff.yaml
│       ├── gfx90a/
│       ├── gfx940/
│       ├── gfx950/                      # latest_arch
│       └── gfx9_config_template.yaml    # single source of truth
│
├── src/util/
│   ├── hash_checker.py
│   ├── .config_hashes.json
│
└── tools/config_management/
    ├── master_config_workflow_script.py
    ├── parse_config_template.py
    ├── verify_against_config_template.py
    ├── generate_config_deltas.py
    ├── apply_config_deltas.py
    ├── hash_manager.py
    ├── TESTING.md
    └── README.md

Core Concepts

Latest Architecture

• Exactly one architecture is considered latest
• Defined in:

src/rocprof_compute_soc/analysis_configs/gfx9_config_template.yaml

Panel YAMLs

• Live under:

analysis_configs/<arch>/*.yaml

• Must conform strictly to the template schema
• Are edited in-place using ruamel.yaml round-trip mode

Delta YAMLs

• Represent differences from latest → older architecture
• Live under:

analysis_configs/<older_arch>/config_delta/

• Exactly one delta file per arch
• Always named:

<latest_arch>_diff.yaml

Hash Database

• Stored at:

src/utils/.config_hashes.json

• Records:
  • md5 hashes of panel YAMLs per arch
  • md5 hash of the delta YAML (or null for latest)
• Machine-generated only
• Enforced in CI and pytest

Architecture Diagram (End-to-End Flow)

                   ┌──────────────────────────┐
                   │  analysis_configs/       │
                   │  gfx9_config_template    │
                   └───────────┬──────────────┘
                               │
                               ▼
                 ┌───────────────────────────────┐
                 │ verify_against_config_template│
                 │ (structural validation)       │
                 └───────────┬───────────────────┘
                             │
         ┌───────────────────┴───────────────────┐
         │                                       │
         ▼                                       ▼
┌────────────────────┐               ┌──────────────────────┐
│ edit-existing mode │               │ promotion mode       │
│ (local dev only)   │               │ (authoritative path) │
└─────────┬──────────┘               └──────────┬───────────┘
          │                                     │
          ▼                                     ▼
┌────────────────────┐               ┌─────────────────────────────┐
│ generate / apply   │               │ parse_config_template.py    │
│ deltas manually    │               │ (update latest_arch)        │
└────────────────────┘               └──────────┬──────────────────┘
                                                 │
                                                 ▼
                               ┌──────────────────────────────────┐
                               │ generate_config_deltas.py        │
                               │ latest → all older arches        │
                               │ (<latest>_diff.yaml only)        │
                               └──────────┬───────────────────────┘
                                          │
                                          ▼
                               ┌──────────────────────────────────┐
                               │ verify_against_config_template   │
                               │ (post-promotion validation)      │
                               └──────────┬───────────────────────┘
                                          │
                                          ▼
                               ┌──────────────────────────────────┐
                               │ hash_manager.py --compute-all    │
                               │ (new steady state)               │
                               └──────────┬───────────────────────┘
                                          │
                                          ▼
                               ┌──────────────────────────────────┐
                               │ hash_checker.py                  │
                               │ (semantic consistency)           │
                               └──────────────────────────────────┘

Contributor Quick Start

Note

Required Python Dependency This configuration management system requires the ruamel.yaml Python package. It is used to safely modify YAML files while preserving comments, ordering, and formatting. The workflow scripts will not function correctly without it.

Install it via:

pip install ruamel.yaml

1. Validate the current state

Before making any config changes:

python tools/config_management/master_config_workflow_script.py --validate-only

This must pass.

2. Editing an existing architecture (most common)

Edit panel YAMLs directly under:

src/rocprof_compute_soc/analysis_configs/<arch>/

Rules:

• Preserve structure
• Preserve ordering
• Use multiline >- formatting for metric descriptions
• Do not regenerate entire files

After editing:

python tools/config_management/master_config_workflow_script.py --validate-only

3. Generating or applying deltas (advanced / optional)

For local experimentation only:

python tools/config_management/master_config_workflow_script.py --edit-existing

This mode:

• never updates the template
• never updates hashes
• always re-validates after application

4. Promoting a new latest architecture (rare, gated)

Promotion changes global invariants and must use the master script:

python tools/config_management/master_config_workflow_script.py --promote <latest_arch>

The script will:

1. Update latest_arch in the template
2. Regenerate deltas for all older arches
3. Remove stale delta files
4. Re-validate everything
5. Rebuild the hash database
6. Verify semantic consistency

If anything fails:

• all changes are rolled back
• no partial state remains

5. Hash checks (fast local / CI)

python tools/config_management/master_config_workflow_script.py --hash-only

or:

python tools/config_management/master_config_workflow_script.py --ci

This runs semantic hash validation only.

Automated Testing

Pytest Hash Integrity Test

Located at:

tests/test_autogen_config.py

This test:

• parses .config_hashes.json
• verifies byte-for-byte integrity of:
  • panel YAMLs
  • delta YAMLs
• fails on:
  • missing files
  • changed content
  • stale hash DB Semantic correctness is enforced separately by hash_checker.py.

Contributor Rules (Strict)

• Do not edit .config_hashes.json manually
• Do not create multiple delta files per arch
• Do not rename delta files arbitrarily
• Do not regenerate full YAMLs unnecessarily
• Use in-place edits (ruamel round-trip)
• Use the master script for promotions
• Expect CI to reject inconsistent states

Summary

This system guarantees:

• A single source of truth for latest architecture
• Deterministic, reviewable deltas
• Stable diffs for Git review
• Hash-backed immutability
• Safe, transactional promotions
• CI-enforced correctness

All correctness flows through:

master_config_workflow_script.py