Provenance

This entry benchmarks the header-only DPF++ library at https://github.com/rh3nry/dpfplusplus, pinned to commit e3f7f0fba75d1736b0da3edb0ce3b22d1b3c73d5 ("ancient, yet functional copy of core dpf++ headers"). The library implements Boyle-Gilboa-Ishai style distributed point functions over __m128i nodes with a fixed-key AES-NI PRG (dpf::prg::aes), packing 128 one-bit leaves per node for dpf::dpf_key<bool>; dpf::dpf_key<uint64_t> supports wider point values. It is the "dpf++" lineage: the pinned repository states that it temporarily holds an old copy of dpf++ and will later be replaced by a redirect to the latest version of libdpf++. Upstream ships no examples or build files, so third_party/dpfplusplus-bench/bench.cc drives the header API (gen, eval, evalfull) directly.

License status and why nothing is vendored

Every upstream header carries a notice that summarizes to: copyright (C) 2019 Anonymous; this is a pre-release version of the DPF++ library distributed anonymously for peer review; a public release under LGPL v2.1 is planned in the near future; please do not redistribute this version of the software. Because redistribution is not permitted, this repository commits zero upstream bytes: the tree is not a Git submodule, and no vendored copy, patched fork, or .patch file with full hunks is stored here. Instead, third_party/dpfplusplus-bench/fetch.sh clones upstream at bench time into the gitignored build directory and applies three scripted text fixes. The fix script quotes only the tiny before/after snippets it needs to locate each edit, which keeps the committed material to interface-level fragments rather than a redistributable copy of the headers.

How the bench acquires the code

bench.py runs bash third_party/dpfplusplus-bench/fetch.sh build/third_party/dpfplusplus during prepare. The script clones the repository into build/third_party/dpfplusplus/src, checks out the pinned commit above, force-resets the tree, and applies the three fixes to src/include/dpf.hpp. Everything lands under the repository's gitignored build/ directory, so a fresh checkout contains no upstream bytes and a clean of build/ removes them entirely. The script is idempotent: re-runs re-fetch, force-reset to the pinned commit, and re-apply the fixes. It runs under set -euo pipefail and each fix demands an exact match count, so upstream drift or a moved commit fails loudly instead of silently benchmarking something else. The CMake wrapper never calls add_subdirectory on the fetched tree; it only adds the fetched src/include directory to the include path, passed as -DDPFPP_INCLUDE_DIR=<build>/src/include.

The three text fixes

The authoritative implementation is the python3 heredoc in third_party/dpfplusplus-bench/fetch.sh; the prose below describes what and why.

The pinned header includes <bsd/stdlib.h> unconditionally for arc4random_buf, which requires libbsd headers and linkage on Linux. The fix guards the include so Apple targets and glibc 2.36 or newer (which provide arc4random_buf in <stdlib.h>) use <stdlib.h>, keeping <bsd/stdlib.h> only for other systems. The bench host (Ubuntu 24.04, glibc 2.39) installs no libbsd headers, so the glibc branch of the guard is required there; the Apple branch covers macOS development hosts.
dpf_key::nodes_per_leaf is declared constexpr but initialized with std::ceil on a double, which is not a constant expression before C++23 and fails to compile under C++17. The fix replaces it with the equivalent integer ceil-div (bits_per_leaf + bits_per_node - 1) / bits_per_node.
Two if constexpr(dpfkey.is_packed) tests, in dpf::eval and dpf::__reeval, read the static member through a function parameter, which is not usable as a constant expression at those sites. The fix names the static class member dpf_key<leaf_t, node_t, prg_t>::is_packed instead, at both occurrences.

Build and benchmark shape

third_party/dpfplusplus-bench/CMakeLists.txt builds the bench executable from bench.cc alone, links Google Benchmark 1.9.5 via FetchContent like the sibling wrappers, and compiles with C++17, -O2, -maes, and -mavx2. The -O2 flag is appended after the build-type flags, so it also pins -O2 when the runner configures CMAKE_BUILD_TYPE=Release. The upstream headers use x86 SSE/AVX/AES intrinsics unconditionally, so the wrapper refuses to configure on non-x86 hosts; the bench machine is x86 Linux. Benchmark IDs are dpfplusplus/CPU/<Op>/<nbits> with state.range(0) carrying nbits: Gen and Eval at 16 and 20 bits, EvalAll at 16, 20, and 24 bits. Gen times dpf_key<bool>::gen producing both parties' keys, including the internal arc4random_buf root draw. Eval times one key.eval point evaluation per iteration and sweeps the input domain between iterations. EvalAll times one party's evalfull over the full domain and reports every domain output as processed work via SetItemsProcessed; following the upstream API, evalfull's internal malloc/free of the traversal hint array is inside the timed region. Each benchmark first runs an untimed self-check: both parties' evalfull shares must XOR to exactly the alpha bit across the whole domain, and a dpf_key<uint64_t> variant is spot-checked by point evaluation (value at the target, zero away from it).

Reproducing

Once bench.py registers the library under the name dpfplusplus, the rows reproduce with:

python3 third_party/bench.py prepare --libraries dpfplusplus --platform cpu
python3 third_party/bench.py build --libraries dpfplusplus --platform cpu --jobs 4
python3 third_party/bench.py run --libraries dpfplusplus --platform cpu --cpu <idle-cpu> --no-build --run-id dpfplusplus-cpu

A manual path without the runner:

bash third_party/dpfplusplus-bench/fetch.sh build/third_party/dpfplusplus
cmake -S third_party/dpfplusplus-bench -B build/third_party/dpfplusplus/cmake \
  -DCMAKE_BUILD_TYPE=Release -DBUILD_TESTING=OFF \
  -DDPFPP_INCLUDE_DIR="$PWD/build/third_party/dpfplusplus/src/include"
cmake --build build/third_party/dpfplusplus/cmake --parallel 16
./build/third_party/dpfplusplus/cmake/bench --benchmark_filter='^dpfplusplus/CPU/' \
  --benchmark_min_time=1s --benchmark_repetitions=5

Numbers from this entry must be measured on the same host as the rest of the comparison bench (cs11369a), under the same runner pinning (taskset, performance governor, single-threaded environment variables). Do not compare smoke-mode timings with full measurements.