This is the software we used to generate the benchmarks in our paper.

The following build instructions are written for users of Ubuntu and other Debian-derived Linux systems; adjust as needed for your own platform.

1. Install Python version 3.6 or higher, Meson, and Ninja:

   sudo apt-get install python3 meson ninja-build
   

2. Clone this repository and cd into the benchmark directory.

Running the benchmarks on your host machine is convenient for development and a good preliminary test, but the results it produces aren't currently very meaningful since most algorithms don't have x86-optimized implementations included in the benchmark suite yet.

1. Set up the build directory:

   meson build/host
   

2. Build the benchmark tool:

   ninja -C build/host
   

3. Run the benchmark tool:

   ./build/host/cipherbench
   

This will test the 32-bit ARM assembly code and the generic code.

1. Download the Android NDK.

2. Connect an Android device and get adb access.

3. If the device is rooted, run adb root to restart adb with root privileges. This isn't required, but it will give more accurate results.

4. Set up the build directory, providing the path to your NDK directory:

   ./cross-tools/setup-build --build-type=android-arm --ndk-dir=/path/to/ndk/dir
   

5. Build the benchmark tool:

   ninja -C build/android-arm
   

6. Run the benchmark tool:

   cross-tools/adb-exe-wrapper adb ./build/android-arm/cipherbench
   

Some algorithms have aarch64 (64-bit ARM) assembly code. To build and run the benchmark tool on Android on aarch64, follow the directions for arm above, but replace all occurrences of "android-arm" with "android-aarch64".

By default, the benchmarks are run using 4096-byte messages and are repeated 5 times for each algorithm, with the fastest speed being chosen. These parameters can be configured via the --bufsize and --ntries options.

To prevent CPU frequency scaling from causing inconsistent results, the benchmark tool tries to temporarily set all CPUs to their maximum frequency. The code which does this assumes a Linux-based system (e.g. Android) and requires root privileges. On other systems, or as a non-root user, you'll see warnings about being unable to set the CPU frequency. You can ignore these if you don't need precise results.

Instead of manually running the tool, you may instead pass one of the predefined run targets to the ninja command, e.g. ninja -C build/host output4096.

By default, most implementations in this benchmark suite optimize solely for speed. However, in some cases the Linux kernel patches for Adiantum make slightly different tradeoffs, considering concerns such as code size and power consumption, and the additional overhead to using SIMD instructions in kernel-mode when compared to userspace. To measure speed in a way more representative of the Linux patches, set up the build with the "kernelish" option, for example:

./cross-tools/setup-build --build-type=android-arm \
    --ndk-dir=/path/to/ndk/dir -- -Dkernelish=true

• src/: C sources for ciphers and benchmark driver
• src/arm/: ARM assembly
• src/aarch64/: ARM64 assembly
• src/x86_64/: x86_64 assembly
• testvectors/: Test vectors for Adiantum and HPolyC as C header files
• ../third_party/: dependencies under the GPLv2 license, not MIT.
• meson_options.txt, meson.build: Meson build control files.
• cross-tools/: Cross compilation support files
• convert_testvecs.py: converts test vectors from JSON to C header form