Search

The heterogeneous system architecture (HSA), announced by the HSA Foundation, is an approach to integrate central processing unit (CPU) and graphics processing unit (GPU) architectures. The open computing language (OpenCL) is a programming framework that can help utilize heterogeneous architectures. The well-known OpenCL framework, currently in version 1.2, provides programming models for heterogeneous computing. The proposed specifications of OpenCL 2.0 can help utilize HSA features, such as shared virtual memory (SVM). In previous work, we helped enable Portable Computing Language (PoCL)-based OpenCL 1.2 runtime frameworks on the HSA. In this paper, we further extend the PoCL-based runtime on the HSA to support OpenCL 2.0 features. In addition, this is the first work, to our best knowledge, to support PoCL-based OpenCL 2.0 features on HSA. Compared with the widely used OpenCL 1.2, OpenCL 2.0 will support SVM, nested parallelism, pipes, and atomic operations. It can further support parallel design patterns such as tree searches, pointer-based programming and nested parallelism models. Note that PoCL is a widely used open source implementation of OpenCL. Our design flow can help academics to enable OpenCL 2.0 flow on the HSA and benefit further from advanced academic research. The experimental results indicate that our framework provides adequate features to support advanced research. [ABSTRACT FROM AUTHOR]

In recent years, the parallel computing performance on GPUs (graphics processing units) has grown rapidly. As a result, GPUs have been widely applied in computationally intensive applications such as image processing, deep learning, and artificial intelligence. Because these applications can be modeled by multiple GPU kernels, some of which might even be dependent, it is essential to identify an efficient method for scheduling dependent kernels on GPU cores. Simply observing kernel dependencies by executing them in sequence results in performance degradation. Furthermore, dependent kernels generally need to share data. Consequently, without properly scheduling dependent kernels, unnecessary memory accesses and copies will be generated. Neural network model environments include many operators that are suitable for both parallel and dependent kernels. This paper proposes an efficient and clear method for creating a framework to analyze the ONNX (open neural network exchange) model and find the pattern of dependent kernels, which can then be used for scheduling with the GPU architecture. The preliminary experimental results show that this technique improves the overall performance by 8% and reduces the cache miss rate by 14% on average by combining neural network operators and appropriate memory policies. [ABSTRACT FROM AUTHOR]