General Matrix-matrix Multiplication Research Articles

Predicting optimal sparse general matrix-matrix multiplication algorithm on GPUs

Sparse General Matrix-Matrix Multiplication (SpGEMM) has played an important role in a number of applications. So far, many efficient algorithms have been proposed to improve the performance of SpGEMM on GPUs. However, the performance of each algorithm for matrices of different structures varies a lot. There is no algorithm that can achieve the optimal performance of SpGEMM computation on all matrices. In this article, we design a machine learning based approach for predicting the optimal SpGEMM algorithm on input matrices. By extracting features from input matrices, we utilize LightGBM and XGBoost to train different lightweight models. The models are capable of predicting the best performing algorithm with low inference overhead and high accuracy for the given input matrices. We also investigate the impact of tree depth on model accuracy and inference overhead. Our evaluation shows that an increase in tree depth leads to a corresponding increase in prediction accuracy, reaching a maximum of approximately 85%, while resulting in increased inference overhead of approximately 10 µs. Compared with the state-of-the-art algorithms on three GPU platforms, our method achieves better overall performance.

An Electro-Photonic System for Accelerating Deep Neural Networks

The number of parameters in deep neural networks (DNNs) is scaling at about 5× the rate of Moore’s Law. To sustain this growth, photonic computing is a promising avenue, as it enables higher throughput in dominant general matrix-matrix multiplication (GEMM) operations in DNNs than their electrical counterpart. However, purely photonic systems face several challenges including lack of photonic memory and accumulation of noise. In this article, we present an electro-photonic accelerator, ADEPT, which leverages a photonic computing unit for performing GEMM operations, a vectorized digital electronic application-specific integrated circuits for performing non-GEMM operations, and SRAM arrays for storing DNN parameters and activations. In contrast to prior works in photonic DNN accelerators, we adopt a system-level perspective and show that the gains while large are tempered relative to prior expectations. Our goal is to encourage architects to explore photonic technology in a more pragmatic way considering the system as a whole to understand its general applicability in accelerating today’s DNNs. Our evaluation shows that ADEPT can provide, on average, 5.73× higher throughput per watt compared to the traditional systolic arrays in a full-system, and at least 6.8× and 2.5× better throughput per watt, compared to state-of-the-art electronic and photonic accelerators, respectively.

A Survey of Accelerating Parallel Sparse Linear Algebra

Sparse linear algebra includes the fundamental and important operations in various large-scale scientific computing and real-world applications. There exists performance bottleneck for sparse linear algebra since it mainly contains the memory-bound computations with low arithmetic intensity. How to improve its performance has increasingly become a focus of research efforts. Using parallel computing techniques to accelerate sparse linear algebra is currently the most popular method, while facing various challenges, e.g., large-scale data brings difficulties in storage, and the sparsity of data leads to irregular memory accesses and parallel load imbalance. Therefore, this article provides a comprehensive overview on acceleration of sparse linear algebra operations using parallel computing platforms, where we focus on four main classifications: sparse matrix-vector multiplication (SpMV), sparse matrix-sparse vector multiplication (SpMSpV), sparse general matrix-matrix multiplication (SpGEMM), and sparse tensor algebra. The takeaways from this article include the following: understanding the challenges of accelerating linear sparse algebra on various hardware platforms; understanding how structured data sparsity can improve storage efficiency; understanding how to optimize parallel load balance; understanding how to improve the efficiency of memory accesses; understanding how do the adaptive frameworks automatically select the optimal algorithms; and understanding recent design trends for acceleration of parallel sparse linear algebra.

RedMule: A mixed-precision matrix–matrix operation engine for flexible and energy-efficient on-chip linear algebra and TinyML training acceleration

The increasing interest in TinyML, i.e., near-sensor machine learning on power budgets of a few tens of mW, is currently pushing toward enabling TinyML-class training as opposed to inference only. Current training algorithms, based on various forms of error and gradient backpropagation, rely on floating-point matrix operations to meet the precision and dynamic range requirements. So far, the energy and power cost of these operations has been considered too high for TinyML scenarios. This paper addresses the open challenge of near-sensor training on a few mW power budget and presents RedMulE — Reduced-Precision Matrix Multiplication Engine, a low-power specialized accelerator conceived for multi-precision floating-point General Matrix–Matrix Operations (GEMM-Ops) acceleration, supporting FP16, as well as hybrid FP8 formats, with {sign,exponent,mantissa}=({1,4,3},{1,5,2}). We integrate RedMule into a Parallel Ultra-Low-Power (PULP) cluster containing eight energy-efficient RISC-V cores sharing a tightly-coupled data memory and implement the resulting system in a 22 nm technology. At its best efficiency point (@ 470 MHz, 0.65 V), the RedMulE-augmented PULP cluster achieves 755 GFLOPS/W and 920 GFLOPS/W during regular General Matrix–Matrix Multiplication (GEMM), and up to 1.19 TFLOPS/W and 1.67 TFLOPS/W when executing GEMM-Ops, respectively, for FP16 and FP8 input/output tensors. In its best performance point (@ 613 MHz, 0.8 V), RedMulE achieves up to 58.5 GFLOPS and 117 GFLOPS for FP16 and FP8, respectively, with 99.4% utilization of the array of Computing Elements and consuming less than 60 mW on average, thus enabling on-device training of deep learning models in TinyML application scenarios while retaining the flexibility to tackle other classes of common linear algebra problems efficiently.

A Simulation Framework for Memristor-Based Heterogeneous Computing Architectures

Memristor-based accelerator (MBA) has demonstrated its capability in accelerating matrix-vector multiplication (MVM) with high performance and energy efficiency. However, it is hard to determine whether and how well an application can benefit from MBAs in a heterogeneous computing architecture. In this article, we propose a simulation framework called MHSim to evaluate the energy efficiency and performance of applications running with both MBAs and CPUs. MHSim provides flexible system-level interfaces and circuit-level simulation models for designers to configure heterogeneous computing architectures. We design a general-purpose MBA which enables floating-point computation models for general matrix-matrix multiplication (GEMM). Our simulation framework can quantify the performance and energy efficiency of different MBA architectures for various applications. We validate our simulation framework with SPICE and evaluate the accuracy and performance of MBAs via several case studies. Experimental results demonstrate that the deviations of energy consumption and latency are only 0.47% and 0.49% on average compared with SPICE-based simulation.

ASA: A ccelerating S parse A ccumulation in Column-wise SpGEMM

Sparse linear algebra is an important kernel in many different applications. Among various sparse general matrix-matrix multiplication (SpGEMM) algorithms, Gustavson’s column-wise SpGEMM has good locality when reading input matrix and can be easily parallelized by distributing the computation of different columns of an output matrix to different processors. However, the sparse accumulation (SPA) step in column-wise SpGEMM, which merges partial sums from each of the multiplications by the row indices, is still a performance bottleneck. The state-of-the-art software implementation uses a hash table for partial sum search in the SPA, which makes SPA the largest contributor to the execution time of SpGEMM. There are three reasons that cause the SPA to become the bottleneck: (1) hash probing requires data-dependent branches that are difficult for a branch predictor to predict correctly; (2) the accumulation of partial sum is dependent on the results of the hash probing, which makes it difficult to hide the hash probing latency; and (3) hash collision requires time-consuming linear search and optimizations to reduce these collisions require an accurate estimation of the number of non-zeros in each column of the output matrix. This work proposes ASA architecture to accelerate the SPA. ASA overcomes the challenges of SPA by (1) executing the partial sum search and accumulate with a single instruction through ISA extension to eliminate data-dependent branches in hash probing, (2) using a dedicated on-chip cache to perform the search and accumulation in a pipelined fashion, (3) relying on the parallel search capability of a set-associative cache to reduce search latency, and (4) delaying the merging of overflowed entries. As a result, ASA achieves an average of 2.25× and 5.05× speedup as compared to the state-of-the-art software implementation of a Markov clustering application and its SpGEMM kernel, respectively. As compared to a state-of-the-art hashing accelerator design, ASA achieves an average of 1.95× speedup in the SpGEMM kernel.

Two-layer integrated photonic architectures with multiport photodetectors for high-fidelity and energy-efficient matrix multiplications.

Photonic integrated circuits (PICs) are emerging as a promising tool for accelerating matrix multiplications in deep learning. Previous PIC architectures, primarily focusing on the matrix-vector multiplication (MVM), have large hardware errors that increase with the device scale. In this work, we propose a novel PIC architecture for MVM, which features an intrinsically small hardware error that does not increase with the device scale. Moreover, we further develop this concept and propose a PIC architecture for the general matrix-matrix multiplication (GEMM), which allows the GEMM to be directly performed on a photonic chip with a high energy efficiency unattainable by parallel or sequential MVMs. This work provides a promising approach to realize a high fidelity and high energy efficiency optical computing platform.

FCNNLib: A Flexible Convolution Algorithm Library for Deep Learning on FPGAs

Convolution features huge complexity and demands high computation capability. Among hardware platforms, field programmable gate array (FPGA) emerges as a promising solution for its substantial available parallelism and energy efficiency. Besides, convolution can be implemented with different algorithms, including conventional, general matrix–matrix multiplication (GEMM), Winograd, and fast Fourier transformation (FFT) algorithms, which are diverse in arithmetic complexity, resource requirement, etc. Different convolutional neural network (CNN) models have different topologies and structures, favoring different convolution algorithms. In response, software libraries such as cuDNN provide a variety of computational primitives to support these algorithms. However, supporting such libraries on FPGAs is challenging. First, multiple algorithms can share the FPGA resources spatially as well as temporally, introducing either reconfiguration overhead or resource underutilization. Second, FPGA implementation remains a significant challenge for library developers. It typically requires significant specialized hardware knowledge. In this article, we propose <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">FCNNLib</monospace> , an efficient and scalable convolution algorithm library on FPGAs. To coordinate multiple convolution algorithms on FPGAs, we develop three schedulings: 1) spatial; 2) temporal; and 3) hybrid, which exhibit different tradeoffs in latency and throughput. We explore these schedulings by balancing the reconfiguration overhead, resource utilization, and optimization objectives of the CNNs. Then, we provide efficient and tunable algorithm templates that allow performance tuning through performance and resource models. To arm the users, <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">FCNNLib</monospace> exposes a set of interfaces to support high-level application designs. We demonstrate the usability of <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">FCNNLib</monospace> with state-of-the-art CNNs. <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">FCNNLib</monospace> achieves up to <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$44.6\times $ </tex-math></inline-formula> and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$1.76\times $ </tex-math></inline-formula> energy efficiency in various scenarios compared with software libraries for CPUs and GPUs, respectively.

An Accelerator for Sparse Convolutional Neural Networks Leveraging Systolic General Matrix-matrix Multiplication

This article proposes a novel hardware accelerator for the inference task with sparse convolutional neural networks (CNNs) by building a hardware unit to perform Image to Column ( Im2Col ) transformation of the input feature map coupled with a systolic-array-based general matrix-matrix multiplication (GEMM) unit. Our design carefully overlaps the Im2Col transformation with the GEMM computation to maximize parallelism. We propose a novel design for the Im2Col unit that uses a set of distributed local memories connected by a ring network, which improves energy efficiency and latency by streaming the input feature map only once. The systolic-array-based GEMM unit in the accelerator can be dynamically configured as multiple GEMM units with square-shaped systolic arrays or as a single GEMM unit with a tall systolic array. This dynamic reconfigurability enables effective pipelining of Im2Col and GEMM operations and attains high processing element utilization for a wide range of CNNs. Further, our accelerator is sparsity aware, improving performance and energy efficiency by effectively mapping the sparse feature maps and weights to the processing elements, skipping ineffectual operations and unnecessary data movements involving zeros. Our prototype, SPOTS, is on average 2.16 \( \times \) , 1.74 \( \times \) , and 1.63 \( \times \) faster than Gemmini, Eyeriss, and Sparse-PE, which are prior hardware accelerators for dense and sparse CNNs, respectively. SPOTS is also 78 \( \times \) and 12 \( \times \) more energy-efficient when compared to CPU and GPU implementations, respectively.

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-on-chip. The efficiency of these accelerators comes from employing optimized dataflow (i.e., spatial/temporal partitioning of data across the PEs and fine-grained scheduling) strategies to optimize data reuse. The focus of this work is to evaluate these accelerator architectures using a tiled general matrix-matrix multiplication (GEMM) kernel. To do so, we develop a framework that finds optimized mappings (dataflow and tile sizes) for a tiled GEMM for a given spatial accelerator and workload combination, leveraging an analytical cost model for runtime and energy. Our evaluations over five spatial accelerators demonstrate that the tiled GEMM mappings systematically generated by our framework achieve high performance on various GEMM workloads and accelerators.

KernelFaRer

Well-crafted libraries deliver much higher performance than code generated by sophisticated application programmers using advanced optimizing compilers. When a code pattern for which a well-tuned library implementation exists is found in the source code of an application, the highest performing solution is to replace the pattern with a call to the library. Idiom-recognition solutions in the past either required pattern matching machinery that was outside of the compilation framework or provided a very brittle solution that would fail even for minor variants in the pattern source code. This article introduces Kernel Find &amp; Replacer ( KernelFaRer ), an idiom recognizer implemented entirely in the existing LLVM compiler framework. The versatility of KernelFaRer is demonstrated by matching and replacing two linear algebra idioms, general matrix-matrix multiplication (GEMM), and symmetric rank-2k update (SYR2K). Both GEMM and SYR2K are used extensively in scientific computation, and GEMM is also a central building block for deep learning and computer graphics algorithms. The idiom recognition in KernelFaRer is much more robust than alternative solutions, has a much lower compilation overhead, and is fully integrated in the broadly used LLVM compilation tools. KernelFaRer replaces existing GEMM and SYR2K idioms with computations performed by BLAS, Eigen, MKL (Intel’s x86), ESSL (IBM’s PowerPC), and BLIS (AMD). Gains in performance that reach 2000× over hand-crafted source code compiled at the highest optimization level demonstrate that replacing application code with library call is a performant solution.

Improving blocked matrix-matrix multiplication routine by utilizing AVX-512 instructions on intel knights landing and xeon scalable processors

In high-performance computing, the general matrix-matrix multiplication (xGEMM) routine is the core of the Level 3 BLAS kernel for effective matrix-matrix multiplication operations. The performance of parallel xGEMM (PxGEMM) is significantly affected by two main factors: the flop rate that can be achieved by calculating the operations and the communication costs for broadcasting submatrices to others. In this study, an approach is proposed to improve and adjust the parallel double-precision general matrix-matrix multiplication (PDGEMM) routine for modern Intel computers such as Knights Landing (KNL) and Xeon Scalable Processors (SKL). The proposed approach consists of two methods to deal with the aforementioned factors. First, the improvement of PDGEMM for the computational part is suggested based on a blocked GEMM algorithm that provides better fits for the architectures of KNL and SKL to perform better block size computation. Second, a communication routine adjustment with the message passing interface is proposed to overcome the settings of the basic linear algebra communication subprograms to improve the time-wise cost efficiency. Consequently, it is shown that performance improvements are achieved in the case of smaller matrix multiplications on the SKL clusters.

Efficient Mixed-Precision Tall-and-Skinny Matrix-Matrix Multiplication for GPUs

General matrix-matrix multiplication (GEMM) is a commonly used BLAS level-3 routine in big data analysis and scientific computations. To further enhance the capability for GEMM computation on GPUs, manufacturers have introduced dedicated hardware for tensor and matrix operations into modern GPU architectures, which is called the Tensor Core unit. Mixed-precision GEMM based on the Tensor Core units has been introduced into many BLAS libraries and deep learning frameworks. However, these implementations are usually designed for large square matrices while these implementations tend to have a low performance for irregular-shaped matrices, especially for tall-and-skinny matrices. This paper discusses on optimizing the GEMM computation suited for tall-and-skinny matrices on GPUs with three optimization methods: task mapping, memory access, and efficient use of Tensor core units by filling multiple fragments. First, the task mapping pattern of GEMM is optimized to make the implementation avoid launching too many thread blocks even when the sizes of input matrices are large. Second, the memory access pattern is optimized for half-precision tall-and-skinny matrices stored in the row-major layout. Third, Tensor Core units are effectively used even for extremely skinny matrices by filling multiple fragments into a Tensor Core operation. To examine the effectiveness of the proposed optimization methods, the experiments are conducted in two cases of GEMM that take tall-and-skinny matrices as input. With the proposed optimization methods, the evaluation results show that the optimized GEMM algorithms can make 1.07x to 3.19x and 1.04x to 3.70x speedups compared with the latest cuBLAS library on NVIDIA V100 and NVIDIA A100, respectively. By reducing the usage of the Tensor Core operations and utilizing the optimized memory access pattern, the optimized GEMM algorithms can save the energy consumptions of V100 and A100 by 34% to 74% and 62% to 82%, respectively.

Optimizing Hardware Accelerated General Matrix-Matrix Multiplication for CNNs on FPGAs

Convolution is inarguably the most complex operation utilized in Convolutional Neural Networks (convnets). Owing to the billions of independent multiply-adds involved, convolution is being massively parallelized by the simultaneous utilization of many cores of Graphical Processing Units (GPUs). Although GPUs have shown significant performance improvements in both training and inference stages, they are not well-suited for mobile vision applications where both energy and real-time constraints need to be satisfied. In contrast, Field Programmable Gate Arrays (FPGAs) have demonstrated massive parallelization capabilities, with fast DSPs and on-chip memory, at a lower energy cost than GPUs. Hence, they are being utilized to design convnet accelerators for embedded applications. In this brief, we design an FPGA-based accelerator for general matrix-matrix multiplication (GeMM) to improve the efficiency of convolutional layers of Shufflenet, an efficient convnet architecture. Experimental results show significant performance improvements against the state-of-the-art FPGA-based implementations of both efficient convnets that are tailored towards mobile vision applications, and complex convnets that are used in traditional applications.

Yet Another Tensor Toolbox for Discontinuous Galerkin Methods and Other Applications

The numerical solution of partial differential equations is at the heart of many grand challenges in supercomputing. Solvers based on high-order discontinuous Galerkin (DG) discretisation have been shown to scale on large supercomputers with excellent performance and efficiency if the implementation exploits all levels of parallelism and is tailored to the specific architecture. However, every year new supercomputers emerge and the list of hardware-specific considerations grows simultaneously with the list of desired features in a DG code. Thus, we believe that a sustainable DG code needs an abstraction layer to implement the numerical scheme in a suitable language. We explore the possibility to abstract the numerical scheme as small tensor operations, describe them in a domain-specific language (DSL) resembling the Einstein notation, and to map them to small General Matrix-Matrix Multiplication routines. The compiler for our DSL implements classic optimisations that are used for large tensor contractions, and we present novel optimisation techniques such as equivalent sparsity patterns and optimal index permutations for temporary tensors. Our application examples, which include the earthquake simulation software SeisSol, show that the generated kernels achieve over 50% peak performance of a recent 48-core Skylake system while the DSL considerably simplifies the implementation.

Matrix multiplication on batches of small matrices in half and half-complex precisions

Machine learning and artificial intelligence (AI) applications often rely on performing many small matrix operations—in particular general matrix–matrix multiplication (GEMM). These operations are usually performed in a reduced precision, such as the 16-bit floating-point format (i.e., half precision or FP16). The GEMM operation is also very important for dense linear algebra algorithms, and half-precision GEMM operations can be used in mixed-precision linear solvers. Therefore, high-performance batched GEMM operations in reduced precision are significantly important, not only for deep learning frameworks, but also for scientific applications that rely on batched linear algebra, such as tensor contractions and sparse direct solvers.This paper presents optimized batched GEMM kernels for graphics processing units (GPUs) in FP16 arithmetic. The paper addresses both real and complex half-precision computations on the GPU. The proposed design takes advantage of the Tensor Core technology that was recently introduced in CUDA-enabled GPUs. With eight tuning parameters introduced in the design, the developed kernels have a high degree of flexibility that overcomes the limitations imposed by the hardware and software (in the form of discrete configurations for the Tensor Core APIs). For real FP16 arithmetic, performance speedups are observed against cuBLAS for sizes up to 128, and range between 1.5× and 2.5×. For the complex FP16 GEMM kernel, the speedups are between 1.7× and 7× thanks to a design that uses the standard interleaved matrix layout, in contrast with the planar layout required by the vendor’s solution. The paper also discusses special optimizations for extremely small matrices, where even higher performance gains are achievable.

Reducing the amount of out‐of‐core data access for GPU‐accelerated randomized SVD

SummaryWe propose two acceleration methods, namely, Fused and Gram, for reducing out‐of‐core data access when performing randomized singular value decomposition (RSVD) on graphics processing units (GPUs). Out‐of‐core data here are data that are too large to fit into the GPU memory at once. Both methods accelerate GPU‐enabled RSVD using the following three schemes: (1) a highly tuned general matrix‐matrix multiplication (GEMM) scheme for processing out‐of‐core data on GPUs; (2) a data‐access reduction scheme based on one‐dimensional data partition; and (3) a first‐in, first‐out scheme that reduces CPU‐GPU data transfer using the reverse iteration. The Fused method further reduces the amount of out‐of‐core data access by merging two GEMM operations into a single operation. By contrast, the Gram method reduces both in‐core and out‐of‐core data access by explicitly forming the Gram matrix. According to our experimental results, the Fused and Gram methods improved the RSVD performance up to 1.7× and 5.2×, respectively, compared with a straightforward method that deploys schemes (1) and (2) on the GPU. In addition, we present a case study of deploying the Gram method for accelerating robust principal component analysis, a convex optimization problem in machine learning.

Accelerating convolutional neural networks on FPGAs

In recent years, convolutional neural networks (CNNs) have become widely adopted for computer vision tasks. FPGAs have been adequately explored as a promising hardware accelerator for CNNs owing to their high performance, energy efficiency, and reconfigurability. However, previous FPGA methods, which are based on the conventional convolutional algorithm, are often bounded by the computational capability of FPGAs. This paper first introduces four convolution algorithms: 6-loop algorithm, general matrix-matrix multiplication (GEMM), Winograd algorithm, and fast Fourier transform (FFT) algorithm. Then, we present the implementations of these algorithms at home and abroad, and also introduce their corresponding optimization techniques.

Register-based implementation of the sparse general matrix-matrix multiplication on GPUs

General sparse matrix-matrix multiplication (SpGEMM) is an essential building block in a number of applications. In our work, we fully utilize GPU registers and shared memory to implement an efficient and load balanced SpGEMM in comparison with the existing implementations.

Design of a High-Performance GEMM-like Tensor–Tensor Multiplication

We present “GEMM-like Tensor–Tensor multiplication” (GETT), a novel approach for dense tensor contractions that mirrors the design of a high-performance general matrix–matrix multiplication (GEMM). The critical insight behind GETT is the identification of three index sets, involved in the tensor contraction, which enable us to systematically reduce an arbitrary tensor contraction to loops around a highly tuned “macro-kernel.” This macro-kernel operates on suitably prepared (“packed”) sub-tensors that reside in a specified level of the cache hierarchy. In contrast to previous approaches to tensor contractions, GETT exhibits desirable features such as unit-stride memory accesses, cache-awareness, as well as full vectorization, without requiring auxiliary memory. We integrate GETT alongside the so-called Transpose–Transpose-GEMM-Transpose and Loops-over-GEMM approaches into an open source “Tensor Contraction Code Generator.” The performance results for a wide range of tensor contractions suggest that GETT has the potential of becoming the method of choice: While GETT exhibits excellent performance across the board, its effectiveness for bandwidth-bound tensor contractions is especially impressive, outperforming existing approaches by up to 12.4×. More precisely, GETT achieves speedups of up to 1.41× over an equivalent-sized GEMM for bandwidth-bound tensor contractions while attaining up to 91.3% of peak floating-point performance for compute-bound tensor contractions.