Skip to content


Repository files navigation

Floating Point correctness benchmark repository.

Build Status: Build Status

This repository contains a system for organising and building various benchmarks.

In this system each benchmark is placed in its own folder containings

  • Source files
  • A benchmark specification file (spec.yml)

This system of organisation is fork of

Imperial College and RWTH Aachen benchmarks

The benchmarks for our ASE 2017 paper are not in this repository. The reason for this is that the benchmarks were developed independently. To include these benchmarks in a build do

cd benchmarks/c
git clone imperial
git clone aachen

before configuring the build. The build system will automatically detect the presense of these directories and build the benchmarks contained within them.

Benchmark specification file

This YAML file contains the relevant information for verifiying and compiling the benchmark.

The schema for these benchmark specification files is written using json-schema and be found at svcb/svcb/schema.yml.

Please note that the schema currently isn't finalised.

Having such a format is useful because

  • It is clear how benchmarks should be compiled.
  • It is clear what properties should be verified/checked for.
  • It is easy to gather statistics on the benchmarks.

A key concept to understand with this format is that a spec.yml can declare multiple benchmarks by declaring multiple variants. If no variants are specified or if there is only one variant the spec.yml file only declares a single benchmark.

Having variants is useful for having benchmarks where small changes guarded by macro definitions create different versions of a benchmark from the same source files.


Before building the benchmarks you will need the following installed:

Note we provide a requirements.txt file for the Python dependencies so you can install these via pip.

pip install -r requirements.txt

If you don't want to install these dependencies globally create a Python virtual environment by using virtualenv.

Building benchmarks as native binaries

Here are the basic steps for building the benchmarks.

$ mkdir build
$ cd build
$ make

where <KLEE_RUNTIME_LIB_DIR> is the directory containing and <KLEE_RUNTIME_INCLUDE_DIR> is the directory containing klee/klee.h. You can get and build KLEE from GitHub.

Note by default it is enforced that the KLEE runtime be available because the majority of the benchmarks will require it. However you can change this by passing -DKLEE_NATIVE_RUNTIME_REQUIRED=FALSE when invoking CMake. Doing this will cause only the benchmarks that don't need the KLEE runtime.

Building benchmarks as LLVM bitcode

KLEE can't run native binaries. Instead it runs LLVM bitcode. To build the benchmarks as LLVM bitcode you should follow the steps above but use Whole Program LLVM as the compiler rather than your normal host compiler.

The invocation to CMake will probably look something like this:

$ export WLLVM_COMPILER=clang
$ mkdir build_llvm
$ cd build_llvm
$ make

If the WLLVM_RUN_EXTRACT_BC CMake option is set to TRUE and CMake detects that wllvm is being used as the compiler then it will automatically run the extract-bc tool on each binary and will output the LLVM bitcode in the same directory as the binary with a .bc suffix. Note extract-bc must be in your PATH.

If the WLLVM_RUN_EXTRACT_BC CMake option is set to FALSE you will need to run the extract-bc tool manually.

Building with run-time profiling


Augmented benchmark specification files

These files can be emitted by the build system (see the EMIT_AUGMENTED_BENCHMARK_SPECIFICATION_FILES option) when building the benchmarks (there is a file per target). These files comply to the benchmark specification format but differ from regular spec.yml files in the following ways:

  • All variants are removed. The file describes only a single benchmark.
  • The misc property is added that contains additional properties. These are: exe_path (Name of the corresponding binary), llvm_bc_path (name of the corresponding LLVM bitcode if built using wllvm), and original_spec (absolute path to the spec.yml file that the file was generated from).

Running schema tests

make check-svcb

Displaying the number of benchmarks by category

make show-categories

Displaying the number of verification tasks

make show-tasks

Displaying a summary of benchmark correctness

make show-correctness-summary

Generate list of augmented spec files

make create-augmented-spec-file-list

Benchmark tools

You can find various tools in svcb/tools/.

This tool will recursively traverse a specified directory parsing all found spec.yml files and reporting the found categories and how many benchmarks are in each category.

This tool will recursively traverse a specified directory parsing all found spec.yml files and reporting all the found verification tasks.

Filter a list of augented spec files by some criteria.

A tool for internal use that when given a spec.yml file will declare all the targets (i.e. the benchmarks) to be built for the CMake build system.

This tool when given a spec.yml file will parse it and display all the benchmarks declared by the file. Note there will only be multiple benchmarks in a spec.yml file is multiple variants are declared in it.

This tool when given a file containing of a list of augmented spec files will generate an invocation info file suitable for use by the klee-runner framework.

Exporting an invocation info file for the klee-runner framework

The klee-runner framework consumes an invocation info file to know how KLEE should run a benchmark. The following steps can be performed to generate this file after successfully building the benchmarks.

# Create `augmented_spec_files.txt`
make create-augmented-spec-file-list

# Filter the list using `` (optional)
/path/to/fp-bench/svcb/tools/ --categories examples -- augmented_spec_files.txt > examples.txt

# Generate invocation info file
/path/to/fp-bench/svcb/tools/ examples.txt > examples_invocation_info.yml


No description, website, or topics provided.







No releases published