Single Instruction Multiple Data Processor Research Articles

Harnessing Manycore Processors with Distributed Memory for Accelerated Training of Sparse and Recurrent Models

Current AI training infrastructure is dominated by single instruction multiple data (SIMD) and systolic array architectures, such as Graphics Processing Units (GPUs) and Tensor Processing Units (TPUs), that excel at accelerating parallel workloads and dense vector matrix multiplications. Potentially more efficient neural network models utilizing sparsity and recurrence cannot leverage the full power of SIMD processor and are thus at a severe disadvantage compared to today's prominent parallel architectures like Transformers and CNNs, thereby hindering the path towards more sustainable AI. To overcome this limitation, we explore sparse and recurrent model training on a massively parallel multiple instruction multiple data (MIMD) architecture with distributed local memory. We implement a training routine based on backpropagation though time (BPTT) for the brain-inspired class of Spiking Neural Networks (SNNs) that feature binary sparse activations. We observe a massive advantage in using sparse activation tensors with a MIMD processor, the Intelligence Processing Unit (IPU) compared to GPUs. On training workloads, our results demonstrate 5-10x throughput gains compared to A100 GPUs and up to 38x gains for higher levels of activation sparsity, without a significant slowdown in training convergence or reduction in final model performance. Furthermore, our results show highly promising trends for both single and multi IPU configurations as we scale up to larger model sizes. Our work paves the way towards more efficient, non-standard models via AI training hardware beyond GPUs, and competitive large scale SNN models.

Partition-based SIMD Processing and its Application to Columnar Database Systems

The Single Instruction Multiple Data (SIMD) paradigm became a core principle for optimizing query processing in columnar database systems. Until now, only the LOAD/STORE instructions are considered to be efficient enough to achieve the expected speedups, while avoiding GATHER/SCATTER is considered almost imperative. However, the GATHER instruction offers a very flexible way to populate SIMD registers with data elements coming from non-consecutive memory locations. As we will discuss within this article, the GATHER instruction can achieve the same performance as the LOAD instruction, if applied properly. To enable the proper usage, we outline a novel access pattern allowing fine-grained, partition-based SIMD implementations. Then, we apply this partition-based SIMD processing to two representative examples from columnar database systems to experimentally demonstrate the applicability and efficiency of our new access pattern.

Fast global motion estimation on single instruction multiple data processors for real-time devices

Global motion estimation is a fundamental tool for image processing and computer vision. Since low-cost real-time devices have limited computational resources, the computational burden of global motion estimation remains considerable for low-cost devices. Hence, a fast global motion estimation algorithm for single instruction multiple data (SIMD) processors with which a CPU in modern mobile computing devices is commonly equipped is proposed. Conventional projection-based global motion estimation is modified to improve the prediction accuracy for videos containing large motions. To significantly increase the parallelism, the exact sizes of the variables in each processing step of the proposed algorithm are determined by considering the size of an SIMD register. Simulation results demonstrate that the proposed scheme is considerably fast and accurate, even for videos containing large motions.

CNFET-Based High Throughput SIMD Architecture

Carbon nanotube field effect transistor (CNFET), using the carbon nanotubes (CNTs) as the material for conducting, is a promising alternative of CMOS technology to overcome the “power wall” issue. Recently, a microprocessor solely based on CNFETs was fabricated and demonstrated, which is a big step forward to the industrial practice. However, CNFETs are inherently subject to much larger process variation or manufacturing defects; thereby it may cause significant design cost to build high performance processors. This is exacerbated in the large register file (RF) architectures widely used in single instruction multiple data (SIMD) architectures, e.g., general public utilities style processors, where the number of critical paths are multiplied by the SIMD width and thread count. In this paper, we seek cost-effective approaches to address the issues by judiciously exploiting the strong asymmetric spatial correlation in the variation unique to the CNFET fabrication process. This paper presents a microarchitectural model to characterize CNFET delay variation and malfunction, under which we show that the RF organizations coupled with the architectural schemes are critical to the performance and power consumption of the SIMD processor. Therefore, we propose several architectural techniques to mitigate the performance degradation and the impact of CNT metallization, leveraging the distinctive CNFET characteristics and the unique features in the SIMD processors. Experimental results verify the effectiveness of the proposed techniques and demonstrate the great opportunity offered by this new device technology.

A MATLAB Vectorizing Compiler Targeting Application-Specific Instruction Set Processors

This article discusses a MATLAB-to-C vectorizing compiler that exploits custom instructions, for example, for Single Instruction Multiple Data (SIMD) processing and instructions for complex arithmetic present in Application-Specific Instruction Set Processors (ASIPs). Custom instructions are represented via specialized intrinsic functions in the generated code, and the generated code can be used as input to any C/C++ compiler supporting the target processor. Furthermore, the specialized instruction set of the target processor is described in a parameterized way using a target processor-independent architecture description approach, thus allowing the support of any processor. The compiler has been used for the generation of application code for two different ASIPs for several benchmarks. The code generated by the compiler achieves a speedup between 2× --74× and 2× --97× compared to the code generated by the MathWorks MATLAB-to-C compiler. Experimental results also prove that the compiler efficiently exploits SIMD custom instructions achieving a 3.3 factor speedup compared to cases where no SIMD processing is used. Thus the compiler can be employed to reduce the development time/effort/cost and time to market through raising the abstraction of application design in an embedded systems/system-on-chip development context.

Modular vector processor architecture targeting at data-level parallelism

Taking advantage of DLP (Data-Level Parallelism) is indispensable in most data streaming and multimedia applications. Several architectures have been proposed to improve both the performance and energy consumption for such applications. Superscalar and VLIW (Very Long Instruction Word) processors along with SIMD (Single-Instruction Multiple-Data) and vector processor (VP) accelerators, are among the available options for designers to accomplish their desired requirements. We present an innovative architecture for a VP which separates the path for performing data shuffle and memory-indexed accesses from the data path for executing other vector instructions that access the memory. This separation speeds up the most common memory access operations by avoiding extra delays and unnecessary stalls. In our lane-based VP design, each vector lane uses its own private memory to avoid any stalls during memory access instructions. The proposed VP, which is developed in VHDL and prototyped on an FPGA, serves as a coprocessor for one or more scalar cores. Benchmarking shows that our VP can achieve very high performance. For example, it achieves a larger than 1500-fold speedup in the color space converting benchmark compared to running the code on a scalar core. The inclusion of distributed data shuffle engines across vector lanes has a spectacular impact on the execution time, primarily for applications like FFT (Fast-Fourier Transform) that require large amounts of data shuffling. Compared to running the benchmark on a VP without the shuffle engines, the speedup is 5.92 and 7.33 for the 64-point FFT without and with compiler optimization, respectively. Compared to runs on the scalar core, the achieved speedups for this benchmark are 52.07 and 110.45 without and with compiler optimization, respectively.

FPGA-based many-core System-on-Chip design

Massively parallel architectures are proposed as a promising solution to speed up data-intensive applications and provide the required computational power. In particular, Single Instruction Multiple Data (SIMD) many-core architectures have been adopted for multimedia and signal processing applications with massive amounts of data parallelism where both performance and flexible programmability are important metrics. However, this class of processors has faced many challenges due to its increasing fabrication cost and design complexity. Moreover, the increasing gap between design productivity and chip complexity requires new design methods. Nowadays, the recent evolution of silicon integration technology, on the one hand, and the wide usage of reusable Intellectual Property (IP) cores and FPGAs (Field Programmable Gate Arrays), on the other hand, are attractive solutions to meet these challenges and reduce the time-to-market. The objective of this work is to study the performances of massively parallel SIMD on-chip architectures with current design methodologies based on recent integration technologies. Flexibility offered by these new design tools allows design space exploration to search for the most effective implementations. This work introduces an IP-based design methodology for easy building configurable and flexible massively parallel SIMD processing on FPGA platforms. The proposed approach allows implementing a generic parallel architecture based on IP assembly that can be tailored in order to better satisfy the requirements of highly-demanding applications. The experimental results show effectiveness of the design methodology as well as the performances of the implemented SoC.

SIMD parallel MCMC sampling with applications for big-data Bayesian analytics

Computational intensity and sequential nature of estimation techniques for Bayesian methods in statistics and machine learning, combined with their increasing applications for big data analytics, necessitate both the identification of potential opportunities to parallelize techniques such as Monte Carlo Markov Chain (MCMC) sampling, and the development of general strategies for mapping such parallel algorithms to modern CPUs in order to elicit the performance up the compute-based and/or memory-based hardware limits. Two opportunities for Single-Instruction Multiple-Data (SIMD) parallelization of MCMC sampling for probabilistic graphical models are presented. In exchangeable models with many observations such as Bayesian Generalized Linear Models (GLMs), child-node contributions to the conditional posterior of each node can be calculated concurrently. In undirected graphs with discrete-value nodes, concurrent sampling of conditionally-independent nodes can be transformed into a SIMD form. High-performance libraries with multi-threading and vectorization capabilities can be readily applied to such SIMD opportunities to gain decent speedup, while a series of high-level source-code and runtime modifications provide further performance boost by reducing parallelization overhead and increasing data locality for Non-Uniform Memory Access architectures. For big-data Bayesian GLM graphs, the end-result is a routine for evaluating the conditional posterior and its gradient vector that is 5 times faster than a naive implementation using (built-in) multi-threaded Intel MKL BLAS, and reaches within the striking distance of the memory-bandwidth-induced hardware limit. Using multi-threading for cache-friendly, fine-grained parallelization can outperform coarse-grained alternatives which are often less cache-friendly, a likely scenario in modern predictive analytics workflow such as Hierarchical Bayesian GLM, variable selection, and ensemble regression and classification. The proposed optimization strategies improve the scaling of performance with number of cores and width of vector units (applicable to many-core SIMD processors such as Intel Xeon Phi and Graphic Processing Units), resulting in cost-effectiveness, energy efficiency (‘green computing’), and higher speed on multi-core x86 processors.

New access modes of parallel memory subsystem for sub-pixel motion estimation

Accessing pixels in memory is a well-known bottleneck of SIMD (single instruction multiple data) processors in video/imaging. To tackle it, we propose new block and row access modes of parallel on-chip memory subsystem, which enable a higher processing throughput and lower energy consumption than the access modes of the state-of-the-art subsystems. The new access modes significantly reduce the number of on-chip memory accesses, and thereby accelerate one of key video/imaging kernels: sub-pixel block-matching motion estimation. The main idea is to exploit spatial overlaps of blocks/rows accessed for pixel interpolation, which are known at the subsystem design-time, and merge multiple accesses into a single one by accessing somewhat more pixels at a time than with other parallel memories. To avoid the need for a wider, and, therefore, more costly SIMD datapath, we propose new memory read operations that split all pixels accessed at a time into multiple SIMD-wide blocks/rows, in a convenient way for further processing. As a proof of concept, we describe a parametric, scalable, and cost-efficient architecture that supports the new access modes. The architecture is based on a previously proposed set of memory banks with multiple pixels per bank word, and a previously proposed shifted scheme for arranging pixels in the banks. We analytically and experimentally demonstrate advantages of this work on a case study of sub-pixel motion estimation for video frame-rate conversion. The implemented motion estimator processes 2160p video at 60 fps in real time, while clocked at 600 MHz. Compared to the implementations based on the state-of-the-art subsystems, this work enables 40–70 % higher throughput, consumes 17–44 % less energy and has similar silicon area and off-chip memory bandwidth costs. That is 1.8–2.9 times more efficient than the prior art, considering the throughput and all costs, i.e., consumption, area, and off-chip bandwidth. Such a higher efficiency is the result of the new access modes, which reduced the number of on-chip memory accesses by 1.6–2.1 times, and the cost-efficient architecture.

A Low-Energy Wide SIMD Architecture with Explicit Datapath

Energy efficiency has become one of the most important topics in computing. To meet the ever increasing demands of the mobile market, the next generation of processors will have to deliver a high compute performance at an extremely limited energy budget. Wide single instruction, multiple data (SIMD) architectures provide a promising solution, as they have the potential to achieve high compute performance at a low energy cost. We propose a configurable wide SIMD architecture that utilizes explicit datapath techniques to further optimize energy efficiency without sacrificing computational performance. To demonstrate the efficiency of the proposed architecture, multiple instantiations of the proposed wide SIMD architecture and its automatic bypassing counterpart, as well as a baseline RISC processor, are implemented. Extensive experimental results show that the proposed architecture is efficient and scalable in terms of area, performance, and energy. In a 128-PE SIMD processor, the proposed architecture is able to achieve an average of 206 times speed-up and reduces the total energy dissipation by 48.3 % on average and up to 94 %, compared to a reduced instruction set computing (RISC) processor. Compared to the corresponding SIMD architecture with automatic bypassing, an average of 64 % of all register file accesses is avoided by the 128-PE, explicitly bypassed SIMD. For total energy dissipation, an average of 27.5 %, and maximum of 43.0 %, reduction is achieved.

Intensive computing on a large data volume with a short‐vector single instruction multiple data processor

In this study, the authors want to evaluate the performances of the PowerXCell 8i processor, which is based on Cell Broadband Engine architecture. For this purpose, the authors chose an algorithm for the k-nearest neighbour problem. The authors optimised this algorithm for efficient exploitation of the facilities provided by this architecture. The authors evaluated the PowerXCell 8i performances by algorithm execution with single- and double-precision calculations. For both cases, the performances were evaluated with and without SIMDisation. For single-precision calculations, the authors achieved a maximum speed-up of 43.85 with SIMDisation by activating 6 synergetic processor element (SPE) processors and 39.73 without SIMDisation by activating 16 SPE processors. For double-precision calculations, the authors achieved a maximum speed-up of 34.79 with SIMDisation by activating 9 SPE processors and 32.71 without SIMDisation by activating 12 SPE processors. These values related to the execution on the PowerPC processor element processor and are due to the accessing way of the main memory by the SPE cores, through the DMA transfers who are performed in parallel with the computing operations. The authors conclude that this process can be efficiently used for the execution of algorithms that require intensive computations on huge data volume.

SIMD defragmenter

Single-instruction multiple-data (SIMD) accelerators provide an energy-efficient platform to scale the performance of mobile systems while still retaining post-programmability. The central challenge is translating the parallel resources of the SIMD hardware into real application performance. In scientific applications, automatic vectorization techniques have proven quite effective at extracting large levels of data-level parallelism (DLP). However, vectorization is often much less effective for media applications due to low trip count loops, complex control flow, and non-uniform execution behavior. As a result, SIMD lanes remain idle due to insufficient DLP. To attack this problem, this paper proposes a new vectorization pass called SIMD Defragmenter to uncover hidden DLP that lurks below the surface in the form of instruction-level parallelism (ILP). The difficulty is managing the data packing/unpacking overhead that can easily exceed the benefits gained through SIMD execution. The SIMD degragmenter overcomes this problem by identifying groups of compatible instructions (subgraphs) that can be executed in parallel across the SIMD lanes. By SIMDizing in bulk at the subgraph level, packing/unpacking overhead is minimized. On a 16-lane SIMD processor, experimental results show that SIMD defragmentation achieves a mean 1.6x speedup over traditional loop vectorization and a 31% gain over prior research approaches for converting ILP to DLP.

SIMD defragmenter

Single-instruction multiple-data (SIMD) accelerators provide an energy-efficient platform to scale the performance of mobile systems while still retaining post-programmability. The central challenge is translating the parallel resources of the SIMD hardware into real application performance. In scientific applications, automatic vectorization techniques have proven quite effective at extracting large levels of data-level parallelism (DLP). However, vectorization is often much less effective for media applications due to low trip count loops, complex control flow, and non-uniform execution behavior. As a result, SIMD lanes remain idle due to insufficient DLP. To attack this problem, this paper proposes a new vectorization pass called SIMD Defragmenter to uncover hidden DLP that lurks below the surface in the form of instruction-level parallelism (ILP). The difficulty is managing the data packing/unpacking overhead that can easily exceed the benefits gained through SIMD execution. The SIMD degragmenter overcomes this problem by identifying groups of compatible instructions (subgraphs) that can be executed in parallel across the SIMD lanes. By SIMDizing in bulk at the subgraph level, packing/unpacking overhead is minimized. On a 16-lane SIMD processor, experimental results show that SIMD defragmentation achieves a mean 1.6x speedup over traditional loop vectorization and a 31% gain over prior research approaches for converting ILP to DLP.

Efficient Image-Vector-Generation Processor for Edge-Based Complementary Feature Representations

A digital processor dedicated to edge-based image vector generation has been developed aiming at real-time image recognition. The processor consists of an on-chip memory and 16 single instruction multiple data (SIMD) processing elements. The capacity of the on-chip memory as well as the overhead for starting the processing have been minimized by introducing a seamless data transferring scheme from memory to processing elements. The 16 SIMD processing elements work together either as accumulators or as shift registers, thus achieving a very efficient generation of two different kinds of feature vector: projected principal-edge distribution (PPED)[3,4] and averaged principal-edge distribution (APED).[5] Concurrent use of these two vectors is shown to be very important for robust image recognition.[5] The chip was fabricated using 0.18-µm complementary metal oxide semiconductor (CMOS) technology and the generation of 64-dimension PPED and APED vectors at 84.7 and 83.9 fps, respectively, from video graphics array (VGA) size images was demonstrated at 62.5 MHz.

A Scalable Massively Parallel Processor for Real-Time Image Processing

This paper describes a high performance scalable massively parallel single-instruction multiple-data (SIMD) processor and power/area efficient real-time image processing. The SIMD processor combines 4-bit processing elements (PEs) with SRAM on a small area and thus enables at the same time a high performance of 191 GOPS, a high power efficiency of 310 GOPS/W, and a high area efficiency of 31.6 GOPS/mm <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> . The applied pipeline architecture is optimized to reduce the number of controller overhead cycles so that the SIMD parallel processing unit can be utilized during up to 99% of the operating time of typical application programs. The processor can be also optimized for low cost, low power, and high performance multimedia system-on-a-chip (SoC) solutions. A combination of custom and automated implementation techniques enables scalability in the number of PEs. The processor has two operating modes, a normal frequency (NF) mode for higher power efficiency and a double frequency (DF) mode for higher performance. The combination of high area efficiency, high power efficiency, high performance, and the flexibility of the SIMD processor described in this paper expands the application of real-time image processing technology to a variety of electronic devices.

From Xetal-II to Xetal-Pro: On the Road Toward an Ultralow-Energy and High-Throughput SIMD Processor

Looking forward to the next generation of mobile streaming computing, the demanded energy efficiency of end-user terminals will become ever stringent. The Xetal-Pro processor, which is the latest member of the Xetal low-power single-instruction multiple data (SIMD) processor family from Philips, is presented in this paper. The predecessor of Xetal-Pro, known as Xetal-II, already ranks as one of the most computational-efficient [in terms of giga operations per second (GOPS)/Watt] processors available today, however, it cannot yet achieve the demanded energy efficiency (less than 1 pJ per operation). Unlike Xetal-II, Xetal-Pro supports ultrawide supply voltage (Vdd) scaling from the nominal supply to the subthreshold region. Although aggressive Vdd scaling causes severe throughput degradation, this can be partly compensated for by the massive parallelism in the Xetal family. Xetal-II includes a large on-chip frame memory (FM), which cannot be scaled well to an ultralow Vdd hence creating a big obstacle to increase energy efficiency. Therefore, we investigate both different FM realizations and memory organization alternatives. A hybrid memory system (HMS), which reduces the non-local memory traffic and enables further Vdd scaling, is proposed. For design space exploration of the right number of the scratchpad memory (SM) entries, the corresponding data locality analysis is provided, too. Moreover, some unique circuit implementation issues of Xetal-Pro such as the customized level-shifter are also discussed. Compared to Xetal-II operating at the nominal voltage, Xetal-Pro provides up to two times energy efficiency improvement even without Vdd scaling (essentially a consequence of data localization in the SM) when delivering the same amount of ultrahigh throughput. With Vdd scaling into the sub/near threshold region, Xetal-Pro could gain more than ten times energy reduction while still delivering a high throughput of 0.69 GOPS (counting multiply and add operations only). The new insight of Xetal-Pro sheds light on the direction of future ultralow-energy SIMD processors.

Compilation Techniques for High Level Parallel Code

This paper describes methods to adapt existing optimizing compilers for sequential languages to produce code for parallel processors. In particular it looks at targeting data-parallel processors using SIMD (single instruction multiple data) or vector processors where users need features similar to high-level control flow across the data-parallelism. The premise of the paper is that we do not want to write an optimizing compiler from scratch. Rather, a method is described that allows a developer to take an existing compiler for a sequential language and modify it to handle SIMD extensions. As well as modifying the front-end, the intermediate representation and the code generation to handle the parallelism, specific optimizations are described to target the architecture efficiently.

Dynamic warp formation

Recent advances in graphics processing units (GPUs) have resulted in massively parallel hardware that is easily programmable and widely available in today's desktop and notebook computer systems. GPUs typically use single-instruction, multiple-data (SIMD) pipelines to achieve high performance with minimal overhead for control hardware. Scalar threads running the same computing kernel are grouped together into SIMD batches, sometimes referred to as warps. While SIMD is ideally suited for simple programs, recent GPUs include control flow instructions in the GPU instruction set architecture and programs using these instructions may experience reduced performance due to the way branch execution is supported in hardware. One solution is to add a stack to allow different SIMD processing elements to execute distinct program paths after a branch instruction. The occurrence of diverging branch outcomes for different processing elements significantly degrades performance using this approach. In this article, we propose dynamic warp formation and scheduling, a mechanism for more efficient SIMD branch execution on GPUs. It dynamically regroups threads into new warps on the fly following the occurrence of diverging branch outcomes. We show that a realistic hardware implementation of this mechanism improves performance by 13%, on average, with 256 threads per core, 24% with 512 threads, and 47% with 768 threads for an estimated area increase of 8%.

Access-Pattern-Aware On-Chip Memory Allocation for SIMD Processors

The number of cycles for each external memory access in Single Instruction Multiple Data (SIMD) processors is heavily affected by the access pattern, such as aligned, unaligned, or stride. We developed a high-performance dynamic on-chip memory-allocation method for SIMD processors by considering the memory access pattern as well as the access frequency. The access pattern and the access count for an array of a loop are determined by both code analysis and profiling, which are performed on a developed compiler framework. This framework not only conducts dynamic on-chip memory allocation but also generates optimized codes for a target processor. The proposed allocation method has been tested with several multimedia benchmarks including motion estimation, 2-D discrete cosine transform, and MPEG2 encoder programs.

Compiler-Based Performance Evaluation of an SIMD Processor with a Multi-Bank Memory Unit

The single instruction multiple data (SIMD) architecture is very efficient for executing arithmetic intensive programs, but frequently suffers from data-alignment problems. The data-alignment problem not only induces extra time overhead but also hinders automatic vectorization of the SIMD compiler. In this paper, we compare three on-chip memory systems, which are single-bank, multi-bank, and multi-port, for the SIMD architecture to resolve the data-alignment problems. The single-bank memory is the simplest, but supports only the aligned accesses. The multi-bank memory requires a little higher complexity, but enables the unaligned accesses and the stride accesses with a bank-conflict limitation. The multi-port memory is capable of both the unaligned and stride accesses without any restriction, but needs quite much expensive hardware. We also developed a vectorizing compiler that can conduct dynamic memory allocation and SIMD code generation. The performances of the three memory systems with our SIMD compiler are evaluated using several digital signal processing kernels and the MPEG2 encoder. The experimental results show that the multi-bank memory can carry out MPEG2 encoding 5.8 times faster, whereas the single-bank memory only achieves 2.9 times speed-up when employed in a multimedia system with a 2-issue host processor and an 8-way SIMD coprocessor. The multi-port memory obviously shows the best performance, which is however an impractical improvement over the multi-bank memory when the hardware cost is considered.