Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Multiple processors”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 217 records · Page 12

Optimization and Parallelization of the Thermal-Hydraulic Sub-channel Code CTF for High-Fidelity Multi-physics Applications

This paper describes major improvements to the computational infrastructure of the CTF sub-channel code so that full-core sub-channel-resolved simulations can now be performed in much shorter run-times, either in stand-alone mode or as part of coupled-code multi-physics calculations. These improvements support the goals of the Department Of Energy (DOE) Consortium for Advanced Simulations of Light Water (CASL) Energy Innovation Hub to develop high fidelity multi-physics simulation tools for nuclear energy design and analysis. A set of serial code optimizations--including fixing computational inefficiencies, optimizing the numerical approach, and making smarter data storage choices--are first described and shown to reduce both execution time and memory usage by about a factor of ten. Next, a Single Program Multiple Data (SPMD) parallelization strategy targeting distributed memory Multiple Instruction Multiple Data (MIMD) platforms and utilizing domain-decomposition is presented. In this approach, data communication between processors is accomplished by inserting standard MPI calls at strategic points in the code. The domain decomposition approach implemented assigns one MPI process to each fuel assembly, with each domain being represented by its own CTF input file. The creation of CTF input files, both for serial and parallel runs, is also fully automated through use of a pre-processor utility that takes a greatly reduced set of user input over the traditional CTF input file. To run CTF in parallel, two additional libraries are currently needed; MPI, for inter-processor message passing, and the Parallel Extensible Toolkit for Scientific Computation (PETSc), which is leveraged to solve the global pressure matrix in parallel. Results presented include a set of testing and verification calculations and performance tests assessing parallel scaling characteristics up to a full core, sub-channel-resolved model of Watts Bar Unit 1 under hot full-power conditions (193 17x17 assemblies, ~56,000 pins, ~59,000 sub-channels, ~2.8 million thermal-hydraulic (TH) control volumes). Results demonstrate that CTF can now perform full core analysis (not previously possible due to excessively long runtimes and memory requirements) on the order of 20 minutes. This new capability is not only useful to standalone CTF users, but is also being leveraged in support of coupled code multi-physics calculations being done in the CASL program.

COBRA-TF↗

Dynamic wavefront creation for processing units using a hybrid compactor

A method, a non-transitory computer readable medium, and a processor for repacking dynamic wavefronts during program code execution on a processing unit, each dynamic wavefront including multiple threads are presented. If a branch instruction is detected, a determination is made whether all wavefronts following a same control path in the program code have reached a compaction point, which is the branch instruction. If no branch instruction is detected in executing the program code, a determination is made whether all wavefronts following the same control path have reached a reconvergence point, which is a beginning of a program code segment to be executed by both a taken branch and a not taken branch from a previous branch instruction. The dynamic wavefronts are repacked with all threads that follow the same control path, if all wavefronts following the same control path have reached the branch instruction or the reconvergence point.

97 MATHEMATICS AND COMPUTING↗

Reduce Operations: Send Volume Balancing While Minimizing Latency

Communication hypergraph model was introduced in a two-phase setting for encapsulating multiple communication cost metrics (bandwidth and latency), which are proven to be important in parallelizing irregular applications. In the first phase, computational-task-to-processor assignment is performed with the objective of minimizing total volume while maintaining computational load balance. In the second phase, communication-task-to-processor assignment is performed with the objective of minimizing total number of messages while maintaining communication-volume balance. The reduce-communication hypergraph model suffers from failing to correctly encapsulate send-volume balancing. We propose a novel vertex weighting scheme that enables part weights to correctly encode send-volume loads of processors for send-volume balancing. The model also suffers from increasing the total communication volume during partitioning. To decrease this increase, we propose a method that utilizes the recursive bipartitioning framework and refines each bipartition by vertex swaps. For performance evaluation, we consider column-parallel SpMV, which is one of the most widely known applications in which the reduce-task assignment problem arises. Extensive experiments on 313 matrices show that, compared to the existing model, the proposed models achieve considerable improvements in all communication cost metrics. Furthermore, these improvements lead to an average decrease of 30% in parallel SpMV time on 512 processors for 70 matrices with high irregularity.

97 MATHEMATICS AND COMPUTING↗

Design of the SLAC RCE Platform: A General Purpose ATCA Based Data Acquisition System

The SLAC RCE platform is a general purpose clustered data acquisition system implemented on a custom ATCA compliant blade, called the Cluster On Board (COB). The core of the system is the Reconfigurable Cluster Element (RCE), which is a system-on-chip design based upon the Xilinx Zynq family of FPGAs, mounted on custom COB daughter-boards. The Zynq architecture couples a dual core ARM Cortex A9 based processor with a high performance 28nm FPGA. The RCE has 12 external general purpose bi-directional high speed links, each supporting serial rates of up to 12Gbps. 8 RCE nodes are included on a COB, each with a 10Gbps connection to an on-board 24-port Ethernet switch integrated circuit. The COB is designed to be used with a standard full-mesh ATCA backplane allowing multiple RCE nodes to be tightly interconnected with minimal interconnect latency. Multiple shelves can be clustered using the front panel 10-gbps connections. The COB also supports local and inter-blade timing and trigger distribution. An experiment specific Rear Transition Module adapts the 96 high speed serial links to specific experiments and allows an experiment-specific timing and busy feedback connection. This coupling of processors with a high performance FPGA fabric in a low latency, multiple node cluster allows high speed data processing that can be easily adapted to any physics experiment. RTEMS and Linux are both ported to the module. The RCE has been used or is the baseline for several current and proposed experiments (LCLS, HPS, LSST, ATLAS-CSC, LBNE, DarkSide, ILC-SiD, etc).

43 PARTICLE ACCELERATORS↗

Highly scalable linear solvers on thousands of processors.

In this report we summarize research into new parallel algebraic multigrid (AMG) methods. We first provide a introduction to parallel AMG. We then discuss our research in parallel AMG algorithms for very large scale platforms. We detail significant improvements in the AMG setup phase to a matrix-matrix multiplication kernel. We present a smoothed aggregation AMG algorithm with fewer communication synchronization points, and discuss its links to domain decomposition methods. Finally, we discuss a multigrid smoothing technique that utilizes two message passing layers for use on multicore processors.

97 MATHEMATICS AND COMPUTING↗

Aleph

Aleph is a PIC/DSMC (particle-in-cell/direct simulation Monte Carlo) software package for the simulation of low temperature plasmas. Aleph is intended to be a production quality code capable of handling complex 3D geometries. Unstructured triangular (2D) or tetrahedral (3D) meshes are used to represent simulation domains. Aleph runs either in serial on a single processor or desktop machine, or can be run in parallel using the MPI message-passing library, to enable faster performance on large problems. Aleph employs dynamic load balancing to efficiently harness thousands of processors on high performance computers and enable simulations of up to millions of mesh elements and billions of particles. Multiple decompositions of the same mesh are used: one for the particles, and one for the field solves. Electrostatic fields are computed using an unstructured mesh finite element method (FEM), and the resulting linear systems are solved using solvers from the Trilinos package (http://trilinos.org/).

Source record↗

Communications systems and methods for subsea processors

A subsea processor may be located near the seabed of a drilling site and used to coordinate operations of underwater drilling components. The subsea processor may be enclosed in a single interchangeable unit that fits a receptor on an underwater drilling component, such as a blow-out preventer (BOP). The subsea processor may issue commands to control the BOP and receive measurements from sensors located throughout the BOP. A shared communications bus may interconnect the subsea processor and underwater components and the subsea processor and a surface or onshore network. The shared communications bus may be operated according to a time division multiple access (TDMA) scheme.

02 PETROLEUM↗

Implementing a micro-operation cache with compaction

Systems, apparatuses, and methods for compacting multiple groups of micro-operations into individual cache lines of a micro-operation cache are disclosed. A processor includes at least a decode unit and a micro-operation cache. When a new group of micro-operations is decoded and ready to be written to the micro-operation cache, the micro-operation cache determines which set is targeted by the new group of micro-operations. If there is a way in this set that can store the new group without evicting any existing group already stored in the way, then the new group is stored into the way with the existing group(s) of micro-operations. Metadata is then updated to indicate that the new group of micro-operations has been written to the way. Additionally, the micro-operation cache manages eviction and replacement policy at the granularity of micro-operation groups rather than at the granularity of cache lines.

Kotra, Jagadish B.↗

An efficient and portable SIMD algorithm for charge/current deposition in Particle-In-Cell codes

In current computer architectures, data movement (from die to network) is by far the most energy consuming part of an algorithm (≈20pJ/word on-die to ≈10,000 pJ/word on the network). To increase memory locality at the hardware level and reduce energy consumption related to data movement, future exascale computers tend to use many-core processors on each compute nodes that will have a reduced clock speed to allow for efficient cooling. To compensate for frequency decrease, machine vendors are making use of long SIMD instruction registers that are able to process multiple data with one arithmetic operator in one clock cycle. SIMD register length is expected to double every four years. As a consequence, Particle-In-Cell (PIC) codes will have to achieve good vectorization to fully take advantage of these upcoming architectures. In this paper, we present a new algorithm that allows for efficient and portable SIMD vectorization of current/charge deposition routines that are, along with the field gathering routines, among the most time consuming parts of the PIC algorithm. Our new algorithm uses a particular data structure that takes into account memory alignment constraints and avoids gather/scat;ter instructions that can significantly affect vectorization performances on current CPUs. The new algorithm was successfully implemented in the 3D skeleton PIC code PICSAR and tested on Haswell Xeon processors (AVX2-256 bits wide data registers). Results show a factor of ×2 to ×2.5 speed-up in double precision for particle shape factor of orders 1–3. The new algorithm can be applied as is on future KNL (Knights Landing) architectures that will include AVX-512 instruction sets with 512 bits register lengths (8 doubles/16 singles). Program summary Program Title: vec_deposition Program Files doi:http://dx.doi.org/10.17632/nh77fv9k8c.1 Licensing provisions: BSD 3-Clause Programming language: Fortran 90 External routines/libraries: OpenMP > 4.0 Nature of problem: Exascale architectures will have many-core processors per node with long vector data registers capable of performing one single instruction on multiple data during one clock cycle. Data register lengths are expected to double every four years and this pushes for new portable solutions for efficiently vectorizing Particle-In-Cell codes on these future many-core architectures. One of the main hotspot routines of the PIC algorithm is the current/charge deposition for which there is no efficient and portable vector algorithm. Solution method: Here we provide an efficient and portable vector algorithm of current/charge deposition routines that uses a new data structure, which significantly reduces gather/scatter operations. Vectorization is controlled using OpenMP 4.0 compiler directives for vectorization which ensures portability across different architectures. Restrictions: Here we do not provide the full PIC algorithm with an executable but only vector routines for current/charge deposition. These scalar/vector routines can be used as library routines in your 3D Particle-In-Cell code. However, to get the best performances out of vector routines you have to satisfy the two following requirements: (1) Your code should implement particle tiling (as explained in the manuscript) to allow for maximized cache reuse and reduce memory accesses that can hinder vector performances. The routines can be used directly on each particle tile. (2) You should compile your code with a Fortran 90 compiler (e.g Intel, gnu or cray) and provide proper alignment flags and compiler alignment directives (more details in README file).

97 MATHEMATICS AND COMPUTING↗

Shadow Distillation: Quantum Error Mitigation with Classical Shadows for Near-Term Quantum Processors

Mitigating errors in quantum information processing devices is especially important in the absence of fault tolerance. An effective method in suppressing state-preparation errors is using multiple copies to distill the ideal component from a noisy quantum state. Here, we use classical shadows and randomized measurements to circumvent the need for coherent access to multiple copies at an exponential cost. We study the scaling of resources using numerical simulations and find that the overhead is still favorable compared to full state tomography. We optimize measurement resources under realistic experimental constraints and apply our method to an experiment preparing a Greenberger-Horne-Zeilinger state with trapped ions. In addition to improving stabilizer measurements, the analysis of the improved results reveals the nature of errors affecting the experiment. Hence, our results provide a directly applicable method for mitigating errors in near-term quantum computers.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗

Multi-node and Multi-core Performance Studies of a Monte Carlo code RMC

Over the past 15 years the architecture of high-performance computing systems has become increasingly heterogeneous with individual computing cores being subdivided into nodes and even auxiliary co-processing units such as graphics processing units or the Intel many integrated core co-processors. In fact, future computer platforms is moving toward a larger number of nodes and cores with lower memory available per node. Today, several production Monte Carlo codes use some types of hybrid parallelism strategy that makes use of shared-memory parallelism within a computing node and message-passing parallelism between nodes, which enables the codes to utilize available computing resources as much as possible, both among and within computing nodes on a cluster. The so-called hybrid parallelization is also possible with Reactor Monte Carlo Code RMC by activating both MPI and OpenMP parallel algorithms together. On HPC' which typically consists of several or several-ten thousand nodes equipping 16-32 computing intra-cores per node, simulation of RMC can be carried out in the hybrid parallelization with the MPI inter-node parallelization and the OpenMP intra-node parallelization. RMC has adopted a hybrid message-passing and shared-memory approach to parallelism, which enables the code to utilize available computing resources, both among and within computing nodes on a cluster. Shared-memory parallelism is especially valuable for large-model simulations because it enables multiple compute cores within a node to share common model information such as the geometry description, material, cross section and tally data, thus reducing the memory usage per processor. The calculation on the Tianhe-2 supercomputer took 6144 cores, resulting in a rate of 173,314,000 neutrons/minute for BEAVRS benchmark. While pin-wise full core H-M benchmark burnup calculations with millions of depletion regions was selected using MPI/OpenMP parallelism, the speed of 6,144 cores reach 1,843,200 burnable regions per minute. Keeping up with the continuing trend for increases in the number of computing resources within modern HPC systems is a significant challenge for teams that develop and maintain high-performance MC radiation transport solvers. order to address this challenge, RMC has adopted a hybrid message-passing and shared-memory approach to parallelism, which enables the code to utilize available computing resources, both among and within computing nodes on a cluster. Shared-memory parallelism is especially valuable for large-model simulations because it enables multiple compute cores within a node to share common model information such as the geometry description, material and composition definitions, and cross section data, thus reducing the memory usage per processor. The parallel performance has been studied by using two complex and representative benchmarks, H-M and BEAVRS benchmark. All simulations were run on Intel Xeon E5-2692 12 core, 2.2 GHz processors at MilkyWay-2 Super Computer with up to 12,288 available computing cores. While employing MPI parallelism, excellent parallel efficiency is achieved even above 12,000 processors and RMC reaches the parallel efficiency of 96% on 12,288 cores for the pin-by-pin computation of the BEAVRS benchmark. While BEAVRS criticality benchmark was selected using MPI/OpenMP parallelism, the speed of 6,144 cores reach 173,314,000 particles per minute. While pin-wise full core H-M benchmark burnup calculations with millions of depletion regions was selected using MPI/OpenMP parallelism, the speed of 6,144 cores reach 1,843,200 burnable regions per minute. Results from these studies demonstrate that RMC, using the hybrid parallel model described in this paper, is able to handle pin-wise full core burnup calculations with millions of depletion regions and scales well through thousands of processors. (authors)

73 NUCLEAR PHYSICS AND RADIATION PHYSICS↗

SQMS Quantum R&D in Machine Learning, Optimization and Sensing beyond Fundamental Physics Applications

This newly formed team at SQMS under the Ecosystem Thrust is looking to develop capabilities impacting societal advances outside the core domain of HEP and condensed matter physics. We explicitly leverage the experimental and algorithmic innovations developed across all groups as well as connect to broad-scope external projects of the diverse team of PIs. As the inaugural set of projects, we are studying numerically quantum machine learning models inspired by efficiently trainable echo-state and orthogonal neural networks and developing designs for related experiments to be performed on quantum processors based on SQMS SRF cQED technology and Rigetti s transmon arrays. Investigated models exploit ideas and lessons learned from multiple prior work by SQMS team members in a variety of internal and external activities [R1]. Target initial applications include noisy signal processing, potentially captured by quantum sensors or noisy QPUs, as well as simulation and classification of healthcare data. For instance, image reconstruction of the brain s electrical properties by solving the inverse Maxwell equation problem with uncertainty [R2] through a hybrid quantum-classical physics-informed architecture for time-dependent processes [R3]. The group is also investigating the application and development of novel quantum sensors based on magnetic levitation of a superconducting sphere coupled to a superconducting qubit. This coupling enables high-precision measurements of the position of the sphere, which can be used for sensitive detection of forces, enabling practical applications such as gravimetry for geophysics analysis, or accelerometry for GPS-denied navigation [R4] [R1] Rieffel, Eleanor G., Ata Akbari Asanjan, M. Sohaib Alam, Namit Anand, David E. Bernal Neira, Sophie Block, Lucas T. Brady et al. "Assessing and advancing the potential of quantum computing: A NASA case study." Future Generation Computer Systems (2024). [R2] Yu, X., Serrall s, J.E., Giannakopoulos, I.I., Liu, Z., Daniel, L., Lattanzi, R. and Zhang, Z., 2023. Pifon-ept: Mr-based electrical property tomography using physics-informed fourier networks. IEEE Journal on Multiscale and Multiphysics Computational Techniques. [R3] Wudarski, Filip, Daniel OConnor, Shaun Geaney, Ata Akbari Asanjan, Max Wilson, Elena Strbac, P. Aaron Lott, and Davide Venturelli. "Hybrid quantum-classical reservoir computing for simulating chaotic systems." arXiv preprint arXiv:2311.14105 (2023). [R4] Higgins, Gerard, Saarik Kalia, and Zhen Liu. "Maglev for dark matter: Dark-photon and axion dark matter sensing with levitated superconductors." Physical Review D 109.5 (2024): 055024.

Venturelli, Davide↗

Real Time Pulse Pile-up Recovery in a High Throughput Digital Pulse Processor

Detection and measurement of radiation is used extensively for non-invasive material characterization in a range of industries. However, many practical applications are frustrated by pulse pile-up within the detector. Pulse pile-up, which occurs when multiple radiation events arrive within the temporal resolving time of the detector, degrades the fidelity of subsequent material analysis. Traditional pulse processing techniques use fast digital filters and logic circuits to detect piled-up events and discard the corrupted data, however, this leads to substantial detector dead time. Consequently, there is considerable interest in more complex signal-processing algorithms to extend the performance of pulse processors and improve material characterization techniques. We present a technology for real-time decoding of pulse pile-up events. It is a model-based signal-processing algorithm able to accurately characterize the number, time-of-arrival and energy of all events in the detector output. Even in the presence of severe multi-pulse pile-up, the composite events are decoded and the energy and time-of-arrival recovered. The technology has been evaluated using a range of detectors, sources and count rates. An exceptional improvement over traditional pulse processing techniques is demonstrated.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗

Experience of Graphite Crucibles during the Pyrometallurgical Processing of Uranium

Throughout the history of processing molten uranium, various materials have been investigated and utilized for the containment of uranium at temperatures greater than 1135oC. Of these, graphite is the most prevalent material due to its ease of fabrication, expense, compatibility with induction systems, and robust characteristics. Graphite in contact with molten uranium will react to form uranium carbide species, thus a coating or mold is needed to prevent reaction of the uranium and the subsequent adherence of the solidified metal. During the pyrometallurgical treatment of used nuclear fuel for uranium recovery at the Idaho National Laboratory, two operational steps utilize graphite crucibles for containment of molten uranium. The cathode processor step separates molten salt from the uranium while the casting step blends uranium for de-enrichment and allows molten sampling of the uranium. Zirconia and yttria based materials on graphite are employed for the cathode processor and casting steps, respectively. For the cathode processor, zirconia coatings were applied initially to graphite crucibles but more recently a zirconia castable material has been molded into the graphite. Although not perfect, zirconia has been found to be compatible with both molten salt and uranium whereas yttria will react with molten salts if carried over to the casting operation. With development, the uranium losses associated with the reaction of uranium and zirconia have been reduced to approximately 1 wt. % of the total uranium throughput. Losses from yttria coating reaction during casting are typically an order of magnitude less than zirconia. The real benefit of these materials is their cost effectiveness given the multiple number of times the graphite crucibles can be reused and/or recoated. Nearly thirty uses have been recorded for graphite crucibles at both the cathode processor and casting operations. The amount of impurity pickup to the uranium products from the graphite crucibles has been minimal (<300 ppm) in terms of carbon, oxygen, zirconium, and yttrium content. Based on the promising results with the zirconia castable at the cathode processor, zirconia material has been cast into a graphite casting crucible and will be tested for comparison in the absence of salt.

11 NUCLEAR FUEL CYCLE AND FUEL MATERIALS↗

Scaling molecular dynamics beyond 100,000 processor cores for large-scale biophysical simulations

The growing interest in the complexity of biological interactions is continuously driving the need to increase system size in biophysical simulations, requiring not only powerful and advanced hardware but adaptable software that can accommodate a large number of atoms interacting through complex force fields. To address this, we developed and implemented strategies in the GENESIS molecular dynamics package designed for large numbers of processors. Long–range electrostatic interactions were parallelized by minimizing the number of processes involved in communication. Here, a novel algorithm was implemented for nonbonded interactions to increase single instruction multiple data (SIMD) performance, reducing memory usage for ultra large systems. Memory usage for neighbor searches in real–space nonbonded interactions was reduced by approximately 80%, leading to significant speedup. Using experimental data describing physical 3D chromatin interactions, we constructed the first atomistic model of an entire gene locus (GATA4). Taken together, these developments enabled the first billion–atom simulation of an intact biomolecular complex, achieving scaling to 65,000 processes (130,000 processor cores) with 1 ns/day performance.

3D modeling↗

Multispectral Quantitative Phase Imaging Using a Diffractive Optical Network

As a label‐free imaging technique, quantitative phase imaging (QPI) provides optical path length information of transparent specimens for various applications in biology, materials science, and engineering. Multispectral QPI measures quantitative phase information across multiple spectral bands, permitting the examination of wavelength‐specific phase and dispersion characteristics of samples. Herein, the design of a diffractive processor is presented that can all‐optically perform multispectral quantitative phase imaging of transparent phase‐only objects within a snapshot. The design utilizes spatially engineered diffractive layers, optimized through deep learning, to encode the phase profile of the input object at a predetermined set of wavelengths into spatial intensity variations at the output plane, allowing multispectral QPI using a monochrome focal plane array. Through numerical simulations, diffractive multispectral processors are demonstrated to simultaneously perform quantitative phase imaging at 9 and 16 target spectral bands in the visible spectrum. The generalization of these diffractive processor designs is validated through numerical tests on unseen objects, including thin Pap smear images. Due to its all‐optical processing capability using passive dielectric diffractive materials, this diffractive multispectral QPI processor offers a compact and power‐efficient solution for high‐throughput quantitative phase microscopy and spectroscopy.

36 MATERIALS SCIENCE↗

Portable, heterogeneous ensemble workflows at scale using libEnsemble

libEnsemble is a Python-based toolkit for running dynamic ensembles, developed as part of the DOE Exascale Computing Project. The toolkit utilizes a unique generator–simulator–allocator paradigm, where generators produce input for simulators, simulators evaluate those inputs, and allocators decide whether and when a simulator or generator should be called. The generator steers the ensemble based on simulation results. Generators may, for example, apply methods for numerical optimization, machine learning, or statistical calibration. libEnsemble communicates between a manager and workers. Flexibility is provided through multiple manager–worker communication substrates each of which has different benefits. These include Python’s multiprocessing, mpi4py, and TCP. Multisite ensembles are supported using Balsam or Globus Compute. We overview the unique characteristics of libEnsemble as well as current and potential interoperability with other packages in the workflow ecosystem. We highlight libEnsemble’s dynamic resource features: libEnsemble can detect system resources, such as available nodes, cores, and GPUs, and assign these in a portable way. These features allow users to specify the number of processors and GPUs required for each simulation; and resources will be automatically assigned on a wide range of systems, including Frontier, Aurora, and Perlmutter. Such ensembles can include multiple simulation types, some using GPUs and others using only CPUs, sharing nodes for maximum efficiency. We also describe the benefits of libEnsemble’s generator–simulator coupling, which easily exposes to the user the ability to cancel, and portably kill, running simulations based on models that are updated with intermediate simulation output. We demonstrate libEnsemble’s capabilities, scalability, and scientific impact via a Gaussian process surrogate training problem for the longitudinal density profile at the exit of a plasma accelerator stage. In conclusion, the study uses gpCAM for the surrogate model and employs either Wake-T or WarpX simulations, highlighting efficient use of resources that can easily extend to exascale.

Dynamic ensembles↗

autoGEMM: Pushing the Limits of Irregular Matrix Multiplication on Arm Architectures

This paper presents an open-source library that pushes the limits of performance portability for irregular General Matrix Multiplication (GEMM) on the widely-used Arm architectures. Our library, autoGEMM, is designed to support a wide range of Arm processors: from edge devices to HPC-grade CPUs. autoGEMM generates optimized kernels for various hardware configurations by auto-combining fragments of autogenerated micro-kernels that employ hand-written optimizations to maximize computational efficiency. We optimize the kernel pipeline by tuning the register reuse and the data load/store overlapping. In addition, we use a dynamic tiling scheme to generate balanced tile shapes. Finally, we position autoGEMM on top of the TVM framework where our dynamic tiling scheme prunes the search space for TVM to identify the optimal combination of parameters for code optimization. Evaluations on five different classes of Arm chips demonstrate the advantages of autoGEMM. For small matrices, autoGEMM achieves 98% of peak and up to 2.0x speedup over state-of-the-art libraries such as LIBXSMM and LibShalom. For irregular matrices (i.e. tall skinny and long rectangles), autoGEMM is 1.3-2.0x faster than widely-used libraries such as OpenBLAS and Eigen. autoGEMM is publicly available at: https://github.com/wudu98/autoGEMM.

Wu, Du↗