An official website of the United States government
Tensor programs often need to process large tensors (vectors, matrices, or higher order tensors) that require a specialized storage format for their memory layout. Several such layouts have been proposed in the literature, such as the Coordinate Format, the Compressed Sparse Row format, and many others, that were especially designed to optimally store tensors with specific sparsity properties. However, existing tensor processing systems require specialized extensions in order to take advantage of every new storage format. In this paper we describe a system that allows users to define flexible storage formats in a declarative tensor query language, similar to the language used by the tensor program. The programmer only needs to write storage mappings, which describe, in a declarative way, how the tensors are laid out in main memory. Then, we describe a cost-based optimizer that optimizes the tensor program for the specific memory layout. We demonstrate empirically significant performance improvements compared to state-of-the-art tensor processing systems.
more »
« less
UniSparse: An Intermediate Language for General Sparse Format Customization
The ongoing trend of hardware specialization has led to a growing use of custom data formats when processing sparse workloads, which are typically memory-bound. These formats facilitate optimized software/hardware implementations by utilizing sparsity pattern- or target-aware data structures and layouts to enhance memory access latency and bandwidth utilization. However, existing sparse tensor programming models and compilers offer little or no support for productively customizing the sparse formats. Additionally, because these frameworks represent formats using a limited set of per-dimension attributes, they lack the flexibility to accommodate numerous new variations of custom sparse data structures and layouts. To overcome this deficiency, we propose UniSparse, an intermediate language that provides a unified abstraction for representing and customizing sparse formats. Unlike the existing attribute-based frameworks, UniSparse decouples the logical representation of the sparse tensor (i.e., the data structure) from its low-level memory layout, enabling the customization of both. As a result, a rich set of format customizations can be succinctly expressed in a small set of well-defined query, mutation, and layout primitives. We also develop a compiler leveraging the MLIR infrastructure, which supports adaptive customization of formats, and automatic code generation of format conversion and compute operations for heterogeneous architectures. We demonstrate the efficacy of our approach through experiments running commonly-used sparse linear algebra operations with specialized formats on multiple different hardware targets, including an Intel CPU, an NVIDIA GPU, an AMD Xilinx FPGA, and a simulated processing-in-memory (PIM) device.
more »
« less
- PAR ID:
- 10527006
- Publisher / Repository:
- ACM
- Date Published:
- Journal Name:
- Proceedings of the ACM on Programming Languages
- Volume:
- 8
- Issue:
- OOPSLA1
- ISSN:
- 2475-1421
- Page Range / eLocation ID:
- 137 to 165
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Graph convolutional networks (GCNs) are fundamental in various scientific applications, ranging from biomedical protein-protein interactions (PPI) to large-scale recommendation systems. An essential component for modeling graph structures in GCNs is sparse general matrix-matrix multiplication (SpGEMM). As the size of graph data continues to scale up, SpGEMMs are often conducted in an out-of-core fashion due to limited GPU memory space in resource-constrained systems. Albeit recent efforts that aim to alleviate the memory constraints of out-of-core SpGEMM through either GPU feature caching, hybrid CPU-GPU memory layout, or performing the computation in sparse format, current systems suffer from both high I/O latency and GPU under-utilization issues. In this paper, we first identify the problems of existing systems, where sparse format data alignment and memory allocation are the main performance bottlenecks, and propose AIRES, a novel algorithm-system co-design solution to accelerate out-of-core SpGEMM computation for GCNs. Specifically, from the algorithm angle, AIRES proposes to alleviate the data alignment issues on the block level for matrices in sparse formats and develops a tiling algorithm to facilitate row block-wise alignment. On the system level, AIRES employs a three-phase dynamic scheduling that features a dual-way data transfer strategy utilizing a tiered memory system: integrating GPU memory, GPU Direct Storage (GDS), and host memory to reduce I/O latency and improve throughput. Evaluations show that AIRES significantly outperforms the state-of-the-art methods, achieving up to 1.8× lower latency in real-world graph processing benchmarks.more » « less
-
Sparse data structures are ubiquitous in modern computing, and numerous formats have been designed to represent them. These formats may exploit specific sparsity patterns, aiming to achieve higher performance for key numerical computations than more general-purpose formats such as CSR and COO. In this work we present UZP, a new sparse format based on polyhedral sets of integer points. UZP is a fexible format that subsumes CSR, COO, DIA, BCSR, etc., by raising them to a common mathematical abstraction: a union of integer polyhedra, each intersected with an ane lattice. We present a modular approach to building and optimizing UZP: it captures equivalence classes for the sparse structure, enabling the tuning of the representation for target-specifc and application-specifc performance considerations. UZP is built from any input sparse structure using integer coordinates, and is interoperable with existing software using CSR and COO data layout. We provide detailed performance evaluation of UZP on 200+ matrices from SuiteSparse, demonstrating how simple and mostly unoptimized generic executors for UZP can already achieve solid performance by exploiting Z-polyhedra structures.more » « less
-
Processing-in-Memory (PIM) is a promising approach to overcoming the memory-wall bottleneck. However, the PIM community has largely treated its two fundamental data layouts, Bit-Parallel (BP) and Bit-Serial (BS), as if they were interchangeable. These layouts represent different execution models (word-parallel versus bit-serial) that directly determine how workloads utilize parallel compute columns and achieve load balance. This implicit "one-layout-fits-all" assumption, often hard-coded into existing evaluation frameworks, creates a critical gap: architects lack systematic, workload-driven guidelines for choosing the optimal data layout for their target applications. To address this gap, this paper presents the first systematic, workload-driven characterization of BP and BS PIM architectures across both performance and energy dimensions. We develop iso-area, cycle-accurate BP and BS PIM architectural models with detailed energy accounting and conduct a comprehensive evaluation using a diverse set of benchmarks. Our suite includes both fine-grained microworkloads from MIMDRAM to isolate specific operational characteristics, and large-scale applications from the PIMBench suite, such as the VGG network, to represent realistic end-to-end workloads. Our results quantitatively demonstrate that no single layout is universally superior in either performance or energy; the optimal choice is strongly dependent on workload characteristics. BP excels on control-flow-intensive tasks, achieving up to 14x latency and 29% energy advantages for arithmetic operations. Conversely, BS demonstrates energy-latency decoupling for logic- intensive kernels, delivering 6–59% energy savings even when cycle counts equal or exceed BP. Based on this characterization, we distill a set of actionable design guidelines for architects. This work challenges the prevailing one-size-fits-all view on PIM data layouts and provides a principled foundation for designing next-generation, workload-aware, and potentially hybrid PIM systems that optimize both performance and energy.more » « less
-
null (Ed.)Data engineering is becoming an increasingly important part of scientific discoveries with the adoption of deep learning and machine learning. Data engineering deals with a variety of data formats, storage, data extraction, transformation, and data movements. One goal of data engineering is to transform data from original data to vector/matrix/tensor formats accepted by deep learning and machine learning applications. There are many structures such as tables, graphs, and trees to represent data in these data engineering phases. Among them, tables are a versatile and commonly used format to load and process data. In this paper, we present a distributed Python API based on table abstraction for representing and processing data. Unlike existing state-of-the-art data engineering tools written purely in Python, our solution adopts high performance compute kernels in C++, with an in-memory table representation with Cython-based Python bindings. In the core system, we use MPI for distributed memory computations with a data-parallel approach for processing large datasets in HPC clusters.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript
- Journal Article:
- https://doi.org/10.1145/3649816
