Evaluation Mappings of Spatial Accelerator Based On Data Placement
The scheduling strategies of workloads are critical to fully exploiting the performance of spatial accelerators, accurate performance models are required to evaluate the mapping of workloads.Recent works proposed various cost-model to describe the dataflow of the spatial accelerator. However, they are less expressive about customized memory hierarchies and thus lead to inaccurate performance models. In this paper, we propose, PolyAcc, a framework for evaluating the mappings of workload on spatial accelerator based on data placement. The Data placement relation describes the temporal-spatial relation of data at different memory levels, which can accurately capture the runtime behavior of hardware units. Based on data placement relations, polyAcc accurately analyzes the data volume for different reuse patterns and estimate metrics, including data reuse, latency, and energy. Overall, polyAcc closely matches the ideal execution time and PE utilization for GEMM and Conv workloads, respectively achieves 0.82%, 18.8% improvements for execution time and energy consumption estimates in validation against Eyeriss architecture compared to the state-of-the-art framework.
Code (0)
등록된 구현이 없습니다.
Tasks
RelationSchedulingSimilar Papers 제목 키워드 기반
Evaluating Spatial Accelerator Architectures with Tiled Matrix-Matrix Multiplication
There is a growing interest in custom spatial accelerators for machine learning applications. These accelerators employ a spatial array of processing elements (PEs) interacting via custom buffer hierarchies and networks-…
SchedulingExploring Quantization and Mapping Synergy in Hardware-Aware Deep Neural Network Accelerators
Energy efficiency and memory footprint of a convolutional neural network (CNN) implemented on a CNN inference accelerator depend on many factors, including a weight quantization strategy (i.e., data types and bit-widths)…
QuantizationSchedulingSynthesizing Optimal Parallelism Placement and Reduction Strategies on Hierarchical Systems for Deep Learning
We present a novel characterization of the mapping of multiple parallelism forms (e.g. data and model parallelism) onto hierarchical accelerator systems that is hierarchy-aware and greatly reduces the space of software-t…
GPUProgram SynthesisMarvel: A Data-centric Compiler for DNN Operators on Spatial Accelerators
The efficiency of a spatial DNN accelerator depends heavily on the compiler and its cost model ability to generate optimized mappings for various operators of DNN models on to the accelerator's compute and memory resourc…
Union: A Unified HW-SW Co-Design Ecosystem in MLIR for Evaluating Tensor Operations on Spatial Accelerators
To meet the extreme compute demands for deep learning across commercial and scientific applications, dataflow accelerators are becoming increasingly popular. While these "domain-specific" accelerators are not fully progr…