FPBench Standards

Standards for floating-point research

FPBench develops standards for describing floating point benchmarks and for measuring their accuracy. All standards are listed on here.

FPBench 2.0

FPBench 2.0 adds support for tensors.

FPCore 2.0: syntax and semantics for benchmarks.
Examples 2.0: FPCore example inputs
Metadata 2.0: descriptive properties for benchmarks
Measures 2.0: standard error measurements
Notes for Implementors: how to use the standards

Changes:

Support for structured data as N-dimensional arrays, or tensors
Bounded iteration with for loops
Composable FPCores
Syntactic sugar for integer rounding contexts

For more information, see the design document.

FPBench 1.2

FPBench 1.2 improves support for imperative programs.

FPCore 1.2: syntax and semantics for benchmarks.
Examples 1.2: FPCore example inputs
Metadata 1.2: descriptive properties for benchmarks
Measures 1.2: standard error measurements
Notes for Implementors: how to use the standards

Changes:

Sequentially binding let* and while*

FPBench 1.1

FPBench 1.1 adds support for mixed-precision computations and clarifies the semantics of rounded operations.

FPCore 1.1: syntax and semantics for benchmarks.
Examples 1.1: FPCore example inputs
Metadata 1.1: descriptive properties for benchmarks
Measures 1.1: standard error measurements
Notes for Implementors: how to use the standards

Changes:

Support for multiple and mixed-precision computations
Precision can be specified per-operation using the rounding context
Semantics of numerical operations clarified as rounded real operations
New number formats like exact rationals and hexadecimal floating-point
New metadata properties :round and :spec

FPBench 1.0

FPBench 1.0 introduced a common format for floating point computations and standard measures for their accuracy.

FPCore 1.0: syntax and semantics for benchmarks.
Metadata 1.0: descriptive properties for benchmarks
Measures 1.0: standard error measurements