Cross-architecture performance prediction (XAPP) using CPU code to predict GPU performance

Abstract

GPUs have become prevalent and more general purpose, but GPU programming remains challenging and time consuming for the majority of programmers. In addition, it is not always clear which codes will benefit from getting ported to GPU. Therefore, having a tool to estimate GPU performance for a piece of code before writing a GPU implementation is highly desirable. To this end, we propose Cross-Architecture Performance Prediction (XAPP), a machine-learning based technique that uses only single-threaded CPU implementation to predict GPU performance. Our paper is built on the two following insights: i) Execution time on GPU is a function of program properties and hardware characteristics. ii) By examining a vast array of previously implemented GPU codes along with their CPU counterparts, we can use established machine learning techniques to learn this correlation between program properties, hardware characteristics and GPU execution time. We use an adaptive two-level machine learning solution. Our results show that our tool is robust and accurate: we achieve 26.9% average error on a set of 24 real-world kernels. We also discuss practical usage scenarios for XAPP.

Comparing DSP Software Performance Prediction Models at Source Code Level — From Analytical to Statistical

“…Reference [7] uses both static attributes such as numbers of different types of instructions and dynamic attributes such as number of cache misses for regression. Reference [2] proposes cross-architecture performance prediction. It is a machine training based technique using both static and dynamic attributes from many programs from some/different benchmarks.…”

A survey and measurement study of GPU DVFS on energy conservation

“…Ardalani et al also used machine learning to train GPU performance models [7]. Their modeling included two techniques: the forward feature selection stepwise regression and the bootstrap aggregating.…”

GPU Performance Estimation using Software Rasterization and Machine Learning

“…Each slice is decomposed into three subgroups: (1) the Slice Common (Figure 8) which provides additional fixed function architectural units; (2) the Sub-Slice (Figure 9) which contains 24 Execution Units (EUs) and supporting execution hardware; and (3) an L3 cache. RastSim models only the portions of the Slice Common and Sub-Slice that are needed to provide functionally correct rendering.…”