Aspect-Driven Mixed-Precision Tuning Targeting GPUs

Abstract

Writing mixed-precision kernels allows to achieve higher throughput together with outputs whose precision remain within given limits. The recent introduction of native half-precision arithmetic capabilities in several GPUs, such as NVIDIA P100 and AMD Vega 10, contributes to make precision-tuning even more relevant as of late. However, it is not trivial to manually find which variables are to be represented as half-precision instead of single-or double-precision. Although the use of half-precision arithmetic can speed up kernel execution considerably, it can also result in providing non-usable kernel outputs, whenever the wrong variables are declared using the half-precision data-type. In this paper we present an automatic approach for precision tuning. Given an OpenCL kernel with a set of inputs declared by a user (i.e., the person responsible for programming and/or tuning the kernel), our approach is capable of deriving the mixed-precision versions of the kernel that are better improve upon the original with respect to a given metric (e.g., time-to-solution, energy-to-solution). We allow the user to declare and/or select a metric to measure and to filter solutions based on the quality of the output. We implement a proof-of-concept of our approach using an aspect-oriented programming language called LARA. It is capable of generating mixed-precision kernels that result in considerably higher performance when compared with the original single-precision floating-point versions, while generating outputs that can be acceptable in some scenarios.

Feedback-Driven Performance and Precision Tuning for Automatic Fixed Point Exploitation

“…Tools in the state-of-the-art are aimed at automatically producing an optimized version of a given numerical program, that sacrifices computa-tion accuracy to obtain performance gains. Such tools either target the entire program [17,15,19,2], or just computational kernels identified by the user [16,22,20,9]. Performance gains are obtained by using smaller data types, by using fixed point in place of floating point computations, or both.…”

The ANTAREX domain specific language for high performance computing

“…LARA promotes modularity and aspect reuse, and supports embedding JavaScript code, to specify more sophisticated strategies. As shown in [12], we support exploration of mixed precision OpenCL kernels by using half, single, and double precision floating point data types. We additionally support fixed point representations through a custom C++ template-based implementation for HPC systems, which has already been used in [13].…”

Supporting the Scale-Up of High Performance Application to Pre-Exascale Systems: The ANTAREX Approach

“…Error-tolerating applications are increasingly common in the emerging field of real-time HPC. In ANTAREX, we explored both precision tuning of floating point computation on GPGPU accelerators [12] and floating to fixed point conversion, followed by tuning of the fixed point representation in terms of bit width and point position [13,14]. c) Memoization: Memoization has been employed for a long time as a performance optimization technique, albeit primarilyin functional languages.…”

“…It has 209 nodes based on Intel Sandy Bridge CPUs 10 . It also contains 23 GPU accelerated nodes 11 and 4 MIC accelerated nodes 12 .…”