rocm-perf-lab

name: rocm-perf-lab description: Deterministic multi-kernel GPU performance analysis and guarded closed-loop optimization framework for ROCm 7.x using rocprofv3, rocpd, roofline modeling, ATT analysis, and critical-path–weighted kernel optimization.

ROCm Perf Lab Skill

Use this skill when performing structured GPU performance engineering on ROCm-based AMD GPUs.

Assumptions:

• ROCm 7.x
• rocprofv3 available
• rocpd SQLite dispatch database available
• HIP kernels (standalone .cu sources for optimization)
• Applications may contain multiple kernels per execution

Core Capabilities

1. Deterministic GPU Runtime Modeling

• Runtime derived from rocpd kernel dispatch timestamps
• Computed as SUM(dispatch_end - dispatch_start)
• Not based on host wall-clock time
• Works for multi-launch and multi-kernel workloads

The .rocpd_profile database is the authoritative timing source.

2. Multi-Kernel Critical Path Analysis

• Reconstructs dispatch DAG from rocpd trace data
• Supports cross-stream execution
• Computes:
  • critical_path_ns
  • Per-kernel slack
  • Critical-path contribution weighting

Optimization prioritization is based on measured critical-path impact, not isolated kernel time.

3. Architecture-Aware Roofline Modeling

Extracts and computes:

• FP32 FLOPs (CDNA3-aware VALU width scaling)
• MFMA contributions (if present)
• DRAM bytes using RDREQ / WRREQ counters
• Arithmetic intensity
• Achieved GFLOP/s
• Achieved GB/s
• First-order bound classification (memory vs compute)

Validated on gfx942-class GPUs (MI300X / MI325).

4. ATT Deep Analysis

Optional ATT pass provides:

• Wave occupancy
• Stall breakdown
• Instruction mix signals
• Latency indicators

ATT enriches feature extraction but does not define runtime.

5. Rule-Based Bottleneck Classification

Combines roofline position and ATT-derived features to produce deterministic labels such as:

• Memory-bandwidth bound
• Latency bound
• Under-occupied
• Divergence limited
• Compute throughput limited

Classification is deterministic for identical inputs.

6. Guarded Closed-Loop LLM Optimization

Command:

rocm-perf optimize <binary>

Scope (v1):

• Standalone HIP kernel source files (.cu)
• No ABI changes
• Kernel signature must remain identical
• Transformation limited to safe loop unrolling (factor 2–8)

Loop behavior:

1. Profile baseline (.rocpd_profile authoritative)
2. Rank kernels by critical-path contribution
3. Generate loop-unrolling proposal
4. Enforce signature invariance and basic structural checks
5. Compile via hipcc
6. Re-profile
7. Accept only if measured runtime improves
8. Automatically revert on regression

Compilation acts as the primary structural validator.

What This Skill Does Not Do

• No AST-based structural verification
• No cross-file refactoring
• No automatic ABI changes
• No formal numerical equivalence proofs
• No speculative optimization without measurement

All improvements are empirical and hardware-validated.

CLI Overview

Primary commands:

rocm-perf profile <binary>
rocm-perf optimize <binary>

See references/cli.md for additional details.