Transparent Acceleration for Heterogeneous Platforms With Compilation to OpenCL

Abstract

Multi-accelerator platforms combine CPUs and different accelerator architectures within a single compute node. Such systems are capable of processing parallel workloads very efficiently while being more energy efficient than regular systems consisting of CPUs only. However, the architectures of such systems are diverse, forcing developers to port applications to each accelerator using different programming languages, models, tools, and compilers. Developers not only require domain-specific knowledge but also need to understand the low-level accelerator details, leading to an increase in the design effort and costs.To tackle this challenge, we propose a compilation approach and a practical realization called HTrOP that is completely transparent to the user. HTrOP is able to automatically analyze a sequential CPU application, detect computational hotspots, and generate parallel OpenCL host and kernel code. The potential of HTrOP is demonstrated by offloading hotspots to different OpenCL-enabled resources (currently the CPU, the generalpurpose GPU, and the manycore Intel Xeon Phi) for a broad set of benchmark applications. We present an in-depth evaluation of our approach in terms of performance gains and energy savings, taking into account all static and dynamic overheads. We are able to achieve speedups and energy savings of up to two orders of magnitude, if an application has sufficient computational intensity, when compared to a natively compiled application.This article is an extension of a conference paper/poster at PPoPP'18 [49]. Prior work only outlined the tool flow and the parallelization mechanism. Evaluation is entirely new, including energy measurements and trends over several data points per application.

OptCL: A Middleware to Optimise Performance for High Performance Domain-Specific Languages on Heterogeneous Platforms

“…Several works [7,24,26] automatically detect parallelisable loops in sequential representations and translate the loops to OpenCL kernels. The existing approaches targeting OpenCL follow a similar workflow as OptCL: a sequential program is first converted into an IR.…”

Automatic translation of data parallel programs for heterogeneous parallelism through OpenMP offloading

“…Grewe et al [19] translate OMP SMP into OpenCL. HTrOP [26] generates OpenCL applications from the LLVM bitcode. CU2CL [28] and Kim et al [9] translate CUDA into OpenCL to achieve portability.…”

“…However, this kind of existing work has a serious performance portability issue as an application implemented in CUDA, by definition, is not portable to non-NVIDIA systems. Other existing work [19,33,26,34] generates OpenCL code that can run on a wide range of parallel hardware including GPUs, CPUs, and FPGAs. Given that OpenCL remains as a low-level programming language that exposes many hardware details, maintaining the generated code is often too difficult for non-expert programmers.…”