ppe

Platform Performance Engineering tools and studies.

Detect what a machine is, model it, measure against the model, trace what ran, and show the result — across CPU, KPU, and GPU platforms.

Status: early. memory_hierarchy is a real measurement — it resolves the cache hierarchy on the machine it runs on. The remaining executables are placeholders and say so in their own output.

Quick start

cmake --preset ci
cmake --build --preset ci
ctest --preset ci

./build-ci/tools/tool_topology                 # what does the OS claim?
./build-ci/applications/app_memory_hierarchy   # what does it actually do?
./build-ci/applications/app_blocking_study     # GEMM blocking sweep

For measurements you intend to keep, use the host-tuned release preset and pin the process:

cmake --preset release && cmake --build --preset release
taskset -c 4 ./build-release/applications/app_memory_hierarchy \
    --max-mib 128 --csv results.csv

On an i7-12700K pinned to a P-core, the latency sweep is flat at ~1.0 ns to 48 KiB, steps 3x, is flat at ~3.06 ns to ~1 MiB, steps again, and plateaus near 82 ns — matching lscpu's 48 KiB L1d, 1.25 MiB L2, and 25 MiB L3.

See docs/getting-started/ for the full build options, and docs/architecture/ for how the pieces fit together.

Layout

Directory	Contents
include/ppe/	Shared substrate — platform attributes, version, CLI helpers
applications/	Performance studies, one subdirectory per study
benchmarks/	Microbenchmarks, one translation unit per measured thing
tools/	Shippable command-line tools, one subdirectory per tool
docs/	All documentation source
docs-site/	Astro/Starlight site generator — consumes docs/, authors nothing

License

MIT — see LICENSE.