README.md

Kerncap Examples

Examples demonstrating how to use kerncap to extract, replay, and validate GPU kernels.

Examples

extract_and_replay.py

Full kerncap pipeline on a multi-kernel HIP application.

Run:

python examples/extract_and_replay.py

What it does:

• Compiles mini_pipeline.hip (five GPU kernels in a single file)
• Profiles the application to rank kernels by GPU time
• Extracts the target kernel into a standalone reproducer
• Replays the captured kernel in isolation and reports timing
• Validates the reproducer for correctness

Options:

# Extract a different kernel (default: vector_scale)
python examples/extract_and_replay.py --kernel histogram_atomic

# Benchmark with more iterations
python examples/extract_and_replay.py --iterations 50

# Save the reproducer to a specific directory
python examples/extract_and_replay.py --output ./my_reproducer

mini_pipeline.hip

A standalone HIP application with five kernels exercising common GPU patterns:

Kernel	Pattern
vector_add	Elementwise addition
vector_scale	Scalar multiplication
vector_bias_relu	Fused bias + ReLU activation
vector_shift	Elementwise shift
histogram_atomic	Atomic histogram (different grid size)

Compile and run directly:

hipcc -O2 -o mini_pipeline examples/mini_pipeline.hip
./mini_pipeline

Prerequisites

• ROCm installed (hipcc, rocprofv3 on PATH)
• AMD GPU (MI300+ recommended)
• kerncap installed: pip install -e kerncap/