Dynamic Precision Autotuning with TAFFO

Abstract

Many classes of applications, both in the embedded and high performance domains, can trade off the accuracy of the computed results for computation performance. One way to achieve such a trade-off is precision tuning-that is, to modify the data types used for the computation by reducing the bit width, or by changing the representation from floating point to fixed point. We present a methodology for high-accuracy dynamic precision tuning based on the identification of input classes (i.e., classes of input datasets that benefit from similar optimizations). When a new input region is detected, the application kernels are re-compiled on the fly with the appropriate selection of parameters. In this way, we obtain a continuous optimization approach that enables the exploitation of the reduced precision computation while progressively exploring the solution space, thus reducing the time required by compilation overheads. We provide tools to support the automation of the runtime part of the solution, leaving to the user only the task of identifying the input classes. Our approach provides a significant performance boost (up to 320%) on the typical approximate computing benchmarks, without meaningfully affecting the accuracy of the result, since the error remains always below 3%.

Architecture-aware Precision Tuning with Multiple Number Representation Systems

“…This paper provides the following main contributions: 1) we present the Integer Equivalent Bit Width (IEBW), a new metric to compare floating point and fixed point representations; 2) we provide LUIS (LLVM-based precision tuning through ILP for mixed number systems), a new precision tuning methodology which exploits this metric, and describe its implementation based on Integer Linear Programming (ILP) model solving. We implemented our approach into an existing tool, TAFFO 1 [6], [7], as a sequence of LLVM optimization passes.…”

TEXTAROSSA: Towards EXtreme scale Technologies and Accelerators for euROhpc hw/Sw Supercomputing Applications for exascale

“…In TEXTAROSSA, we aim at exploiting and extending the tools for precision tuning developed as part of the H2020 FETHPC ANTAREX project [24] to cover a wider range of target platforms, targeting FPGAs through integration with the TEXTAROSSA High Level Synthesis (HLS) toolchain. These tools, collected in the TAFFO framework [25], [26] are implemented as a set of plugins for the LLVM compiler, and, based on programmer hints expressed as attributes, perform value range analysis, data type and code conversion, and static estimation of the performance impact. We aim at improving the performance estimation by exploiting recent analysis techniques [27] as well as deeper understanding of the target processor pipeline, by expanding the use of the tools to heterogeneous systems with reconfigurable components, and by considering emerging data types such as Posit and Bfloat16.…”

The Italian research on HPC key technologies across EuroHPC

“…TAFFO allows fine-tuning the selection of data types, taking into account the cost of casting. It can be coupled with a code versioning library to dynamically perform the tuning, which has been proven a key point in addressing large scale applications where the input data characteristics can change over time [11]. libVC [10] is the code versioning library adopted for TAFFO.…”