Abstract SYCL is an open standard for targeting heterogeneous hardware from C++. In this work, we evaluate a SYCL implementation for a discontinuous Galerkin discretization of the 2D shallow water equations targeting CPUs, GPUs, and also FPGAs. The discretization uses polynomial orders zero to two on unstructured triangular meshes. Separating memory accesses from the numerical code allow us to optimize data accesses for the target architecture. A performance analysis shows good portability across x86 and ARM CPUs, GPUs from different vendors, and even two variants of Intel Stratix 10 FPGAs. Measuring the energy to solution shows that GPUs yield an up to 10x higher energy efficiency in terms of degrees of freedom per joule compared to CPUs. With custom designed caches, FPGAs offer a meaningful complement to the other architectures with particularly good computational performance on smaller meshes. FPGAs with High Bandwidth Memory are less affected by bandwidth issues and have similar energy efficiency as latest generation CPUs.

Although High-Level Synthesis has dramatically improved the usability of FPGAs in High-Performance Computing applications, it still requires much experience to create well-performing FPGA designs. We introduce the stencil simulation framework StencilStream to separate the concerns of performance engineers and domain scientists. We have formulated our performance-oriented stencil simulation designs for FPGAs along with their management code as C++ templates using SYCL language features. Application developers instantiate these templates with their concrete stencil code and obtain a complete application with little management overhead. In this work, we describe the most interesting API features and the two backend architectures of StencilStream, along with their corresponding performance models. We then evaluate the performance of the architectures using three manually implemented sample applications and highlight the real-world usability of StencilStream further with shallow water simulations obtained from a code-generation flow. Compiled with Intel oneAPI for an Intel Stratix 10 GX 2800 target FPGA, the best-performing sample design exceeds a throughput of 1 TFLOPS.

Despite the recent progress, little is known about the features captured by state-of-the-art neural relation extraction (RE) models.Common methods encode the source sentence, conditioned on the entity mentions, before classifying the relation.However, the complexity of the task makes it difficult to understand how encoder architecture and supporting linguistic knowledge affect the features learned by the encoder.We introduce 14 probing tasks targeting linguistic properties relevant to RE, and we use them to study representations learned by more than 40 different encoder architecture and linguistic feature combinations trained on two datasets, TACRED and SemEval 2010 Task 8. We find that the bias induced by the architecture and the inclusion of linguistic features are clearly expressed in the probing task performance.For example, adding contextualized word representations greatly increases performance on probing tasks with a focus on named entity and part-of-speech information, and yields better results in RE.In contrast, entity masking improves RE, but considerably lowers performance on entity type related probing tasks.

TACRED (Zhang et al., 2017) is one of the largest, most widely used crowdsourced datasets in Relation Extraction (RE).But, even with recent advances in unsupervised pretraining and knowledge enhanced neural RE, models still show a high error rate.In this paper, we investigate the questions: Have we reached a performance ceiling or is there still room for improvement?And how do crowd annotations, dataset, and models contribute to this error rate?To answer these questions, we first validate the most challenging 5K examples in the development and test sets using trained annotators.We find that label errors account for 8% absolute F1 test error, and that more than 50% of the examples need to be relabeled.On the relabeled test set the average F1 score of a large baseline model set improves from 62.1 to 70.1.After validation, we analyze misclassifications on the challenging instances, categorize them into linguistically motivated error groups, and verify the resulting error hypotheses on three state-of-the-art RE models.We show that two groups of ambiguous relations are responsible for most of the remaining errors and that models may adopt shallow heuristics on the dataset when entities are not masked.

Distantly supervised relation extraction is widely used to extract relational facts from text, but suffers from noisy labels.Current relation extraction methods try to alleviate the noise by multi-instance learning and by providing supporting linguistic and contextual information to more efficiently guide the relation classification.While achieving state-of-the-art results, we observed these models to be biased towards recognizing a limited set of relations with high precision, while ignoring those in the long tail.To address this gap, we utilize a pre-trained language model, the OpenAI Generative Pre-trained Transformer (GPT) (Radford et al., 2018).The GPT and similar models have been shown to capture semantic and syntactic features, and also a notable amount of "common-sense" knowledge, which we hypothesize are important features for recognizing a more diverse set of relations.By extending the GPT to the distantly supervised setting, and fine-tuning it on the NYT10 dataset, we show that it predicts a larger set of distinct relation types with high confidence.Manual and automated evaluation of our model shows that it achieves a state-of-the-art AUC score of 0.422 on the NYT10 dataset, and performs especially well at higher recall levels.