A Code Generator for High-Performance Tensor Contractions on GPUs

2019 ◽  
Author(s):  
Jinsung Kim ◽  
Aravind Sukumaran-Rajam ◽  
Vineeth Thumma ◽  
Sriram Krishnamoorthy ◽  
Ajay Panyala ◽  
...  
Keyword(s):  
High Performance ◽  
Code Generator ◽  
Tensor Contractions

Domain-Specific Optimization of Two Jacobi Smoother Kernels and Their Evaluation in the ECM Performance Model

2014 ◽  
Vol 24 (03) ◽  
pp. 1441004 ◽  
Author(s):  
Stefan Kronawitter ◽  
Holger Stengel ◽  
Georg Hager ◽  
Christian Lengauer
Keyword(s):  
High Performance ◽  
Memory Performance ◽  
Performance Model ◽  
Program Transformations ◽  
Code Generator ◽  
Domain Specific ◽  
Code Performance ◽  
Execution Platform ◽  
Stencil Codes ◽  
Point Stencil

Our aim is to apply program transformations to stencil codes in order to yield the highest possible performance. We recognize memory bandwidth as a major limitation in stencil code performance. We conducted a study in which we applied optimizing transformations to two Jacobi smoother kernels: one 3D 1st-order 7-point stencil and one 3D 3rd-order 19-point stencil. To obtain high performance, the optimizations have to be customized for the execution platform at hand. We illustrate this by experiments on two consumer and two server architectures. We also verified the need for complex optimizations with the help of the Execution-Cache-Memory performance model. A code generator with knowledge about stencil codes and execution platforms should be able to apply our transformations automatically. We are working towards such a generator in project ExaStencils.

Linnea

2021 ◽  
Vol 47 (3) ◽  
pp. 1-26
Author(s):  
Henrik Barthels ◽  
Christos Psarras ◽  
Paolo Bientinesi
Keyword(s):  
Linear Algebra ◽  
High Performance ◽  
Search Algorithm ◽  
Test Problems ◽  
Code Generator ◽  
Matrix Computations ◽  
Significant Performance ◽  
High Level ◽  
Almost All ◽  
Performance Computing

The translation of linear algebra computations into efficient sequences of library calls is a non-trivial task that requires expertise in both linear algebra and high-performance computing. Almost all high-level languages and libraries for matrix computations (e.g., Matlab, Eigen) internally use optimized kernels such as those provided by BLAS and LAPACK; however, their translation algorithms are often too simplistic and thus lead to a suboptimal use of said kernels, resulting in significant performance losses. To combine the productivity offered by high-level languages, and the performance of low-level kernels, we are developing Linnea, a code generator for linear algebra problems. As input, Linnea takes a high-level description of a linear algebra problem; as output, it returns an efficient sequence of calls to high-performance kernels. Linnea uses a custom best-first search algorithm to find a first solution in less than a second, and increasingly better solutions when given more time. In 125 test problems, the code generated by Linnea almost always outperforms Matlab, Julia, Eigen, and Armadillo, with speedups up to and exceeding 10×.

scholarly journals SPOT: A DSL for Extending Fortran Programs with Metaprogramming

2014 ◽  
Vol 2014 ◽  
pp. 1-23 ◽  
Author(s):  
Songqing Yue ◽  
Jeff Gray
Keyword(s):  
Code Generation ◽  
Software Maintenance ◽  
High Performance ◽  
Program Transformations ◽  
Code Generator ◽  
Design And Implementation ◽  
Software Maintenance And Evolution ◽  
Performance Computing ◽  
Automate Program

Metaprogramming has shown much promise for improving the quality of software by offering programming language techniques to address issues of modularity, reusability, maintainability, and extensibility. Thus far, the power of metaprogramming has not been explored deeply in the area of high performance computing (HPC). There is a vast body of legacy code written in Fortran running throughout the HPC community. In order to facilitate software maintenance and evolution in HPC systems, we introduce a DSL that can be used to perform source-to-source translation of Fortran programs by providing a higher level of abstraction for specifying program transformations. The underlying transformations are actually carried out through a metaobject protocol (MOP) and a code generator is responsible for translating a SPOT program to the corresponding MOP code. The design focus of the framework is to automate program transformations through techniques of code generation, so that developers only need to specify desired transformations while being oblivious to the details about how the transformations are performed. The paper provides a general motivation for the approach and explains its design and implementation. In addition, this paper presents case studies that illustrate the potential of our approach to improve code modularity, maintainability, and productivity.

Automatic Code Generator for a Customized High Performance Microprocessor Simulator

2018 ◽  
pp. 11-23
Author(s):  
Alfredo Cristóbal-Salas ◽  
Juan D. Santiago-Domínguez ◽  
Bardo Santiago-Vicente ◽  
Marco Antonio Ramírez-Salinas ◽  
Luis Alfonso Villa-Vargas ◽  
...  
Keyword(s):  
High Performance ◽  
Code Generator ◽  
Automatic Code

scholarly journals High-performance Tensor Contractions for GPUs

2016 ◽  
Vol 80 ◽  
pp. 108-118 ◽  
Author(s):  
A. Abdelfattah ◽  
M. Baboulin ◽  
V. Dobrev ◽  
J. Dongarra ◽  
C. Earl ◽  
...  
Keyword(s):  
High Performance ◽  
Tensor Contractions

A High Performance Energy Analyzer for Use in Electron Scanning Microscopy

1969 ◽  
Vol 27 ◽  
pp. 14-15
Author(s):  
A. V. Crewe ◽  
M. Isaacson ◽  
D. Johnson
Keyword(s):  
High Performance ◽  
Vertical Plane ◽  
Median Plane ◽  
Electron Gun ◽  
Uniform Field ◽  
Loss Mechanism ◽  
Electron Scanning Microscopy ◽  
Sector Type ◽  
Double Focusing ◽  
Electron Energy Loss

A double focusing magnetic spectrometer has been constructed for use with a field emission electron gun scanning microscope in order to study the electron energy loss mechanism in thin specimens. It is of the uniform field sector type with curved pole pieces. The shape of the pole pieces is determined by requiring that all particles be focused to a point at the image slit (point 1). The resultant shape gives perfect focusing in the median plane (Fig. 1) and first order focusing in the vertical plane (Fig. 2).

Vacuum System to Minimize the Specimen Contamination of High-Performance EM

1977 ◽  
Vol 35 ◽  
pp. 68-69
Author(s):  
N. Yoshimura ◽  
K. Shirota ◽  
T. Etoh
Keyword(s):  
Electron Microscope ◽  
High Speed ◽  
High Performance ◽  
High Vacuum ◽  
Vacuum System ◽  
Pump System ◽  
Pumping System ◽  
Diffusion Pump ◽  
Almost All ◽  
Cascade Type

One of the most important requirements for a high-performance EM, especially an analytical EM using a fine beam probe, is to prevent specimen contamination by providing a clean high vacuum in the vicinity of the specimen. However, in almost all commercial EMs, the pressure in the vicinity of the specimen under observation is usually more than ten times higher than the pressure measured at the punping line. The EM column inevitably requires the use of greased Viton O-rings for fine movement, and specimens and films need to be exchanged frequently and several attachments may also be exchanged. For these reasons, a high speed pumping system, as well as a clean vacuum system, is now required. A newly developed electron microscope, the JEM-100CX features clean high vacuum in the vicinity of the specimen, realized by the use of a CASCADE type diffusion pump system which has been essentially improved over its predeces- sorD employed on the JEM-100C.

Use of the Magnetostatic Analogue In Electromagnetic Lens Engineering

1970 ◽  
Vol 28 ◽  
pp. 360-361
Author(s):  
John W. Coleman
Keyword(s):  
Boundary Conditions ◽  
High Performance ◽  
Force Fields ◽  
Optical Design ◽  
Direct Conversion ◽  
Design Engineering ◽  
Design Data ◽  
Scalar Potentials ◽  
Geometric Elements ◽  
At Will

In the design engineering of high performance electromagnetic lenses, the direct conversion of electron optical design data into drawings for reliable hardware is oftentimes difficult, especially in terms of how to mount parts to each other, how to tolerance dimensions, and how to specify finishes. An answer to this is in the use of magnetostatic analytics, corresponding to boundary conditions for the optical design. With such models, the magnetostatic force on a test pole along the axis may be examined, and in this way one may obtain priority listings for holding dimensions, relieving stresses, etc..The development of magnetostatic models most easily proceeds from the derivation of scalar potentials of separate geometric elements. These potentials can then be conbined at will because of the superposition characteristic of conservative force fields.

Prospects for convergent beam electron diffraction with a 200kV cold field-emission transmission electron microscope

1991 ◽  
Vol 49 ◽  
pp. 466-467
Author(s):  
J W Steeds ◽  
R Vincent
Keyword(s):  
Field Emission ◽  
High Performance ◽  
Small Diameter ◽  
Convergent Beam Electron Diffraction ◽  
Zone Axis ◽  
Medium Voltage ◽  
Transmission Electron ◽  
Good Crystallinity ◽  
Special Modification ◽  
Convergent Beam

We review the analytical powers which will become more widely available as medium voltage (200-300kV) TEMs with facilities for CBED on a nanometre scale come onto the market. Of course, high performance cold field emission STEMs have now been in operation for about twenty years, but it is only in relatively few laboratories that special modification has permitted the performance of CBED experiments. Most notable amongst these pioneering projects is the work in Arizona by Cowley and Spence and, more recently, that in Cambridge by Rodenburg and McMullan.There are a large number of potential advantages of a high intensity, small diameter, focussed probe. We discuss first the advantages for probes larger than the projected unit cell of the crystal under investigation. In this situation we are able to perform CBED on local regions of good crystallinity. Zone axis patterns often contain information which is very sensitive to thickness changes as small as 5nm. In conventional CBED, with a lOnm source, it is very likely that the information will be degraded by thickness averaging within the illuminated area.

Current State of Biological High-Resolution Scanning Electron Microscopy

1988 ◽  
Vol 46 ◽  
pp. 180-181 ◽  
Author(s):  
Klaus-Ruediger Peters
Keyword(s):  
High Performance ◽  
Low Voltage ◽  
Backscattered Electrons ◽  
Lens Position ◽  
Low Voltage Operation ◽  
Signal Collection ◽  
Probe Size ◽  
Electron Microscopes ◽  
Scanning Electron Microscopes ◽  
Scanning Electron

A new generation of high performance field emission scanning electron microscopes (FSEM) is now commercially available (JEOL 890, Hitachi S 900, ISI OS 130-F) characterized by an "in lens" position of the specimen where probe diameters are reduced and signal collection improved. Additionally, low voltage operation is extended to 1 kV. Compared to the first generation of FSEM (JE0L JSM 30, Hitachi S 800), which utilized a specimen position below the final lens, specimen size had to be reduced but useful magnification could be impressively increased in both low (1-4 kV) and high (5-40 kV) voltage operation, i.e. from 50,000 to 200,000 and 250,000 to 1,000,000 x respectively.At high accelerating voltage and magnification, contrasts on biological specimens are well characterized1 and are produced by the entering probe electrons in the outmost surface layer within -vl nm depth. Backscattered electrons produce only a background signal. Under these conditions (FIG. 1) image quality is similar to conventional TEM (FIG. 2) and only limited at magnifications >1,000,000 x by probe size (0.5 nm) or non-localization effects (%0.5 nm).

Sign in / Sign up

Export Citation Format

Share Document