blis performs
play

BLIS Performs Devangi N. Parikh Science of High Performance - PowerPoint PPT Presentation

Dec 20, 2022 •129 likes •311 views

BLIS Performs Devangi N. Parikh Science of High Performance Compu8ng The University of Texas at Aus8n ThunderX2 Architecture arm v8.1 Base frequency 2.0 GHz # sockets/node 2 # cores/socket 28 armv8a kernels in BLIS were wriOen by

  1. BLIS Performs Devangi N. Parikh Science of High Performance Compu8ng The University of Texas at Aus8n

  2. ThunderX2 Architecture arm v8.1 Base frequency 2.0 GHz # sockets/node 2 # cores/socket 28 armv8a kernels in BLIS were wriOen by Fransisco D. Igual for cortexa57 architectures.

  3. DGEMM (armv8a) DGEMM (single-threaded) 16 14 12 10 GFLOPS 8 BLIS 6 4 2 0 0 200 400 600 800 1000 1200 1400 1600 1800 2000 matrix dimension m=n=k

  4. DGEMM – Other Libraries DGEMM (single-threaded) 16 14 12 10 GFLOPS 8 6 4 BLIS OpenBLAS ARMPL 2 0 0 200 400 600 800 1000 1200 1400 1600 1800 2000 matrix dimension m=n=k

  5. GEMM – Other Datatypes SGEMM (single-threaded) 30 25 20 GFLOPS 15 10 BLIS OpenBLAS ARMPL 5 0 0 200 400 600 800 1000 1200 1400 1600 1800 2000 matrix dimension m=n=k

  6. GEMM – Other Datatypes CGEMM (single-threaded) 30 25 20 GFLOPS 15 10 5 BLIS OpenBLAS ARMPL 0 0 200 400 600 800 1000 1200 1400 1600 1800 2000 matrix dimension m=n=k

  7. GEMM – Other Datatypes ZGEMM (single-threaded) 16 14 12 10 GFLOPS 8 6 4 2 BLIS OpenBLAS ARMPL 0 0 200 400 600 800 1000 1200 1400 1600 1800 2000 matrix dimension m=n=k

  8. Level 3 SGEMM (single-threaded) SSYRK (single-threaded) SSYMM (single-threaded) STRMM (single-threaded) 30 30 30 30 25 25 25 25 20 20 20 20 GFLOPS GFLOPS GFLOPS GFLOPS 15 15 15 15 10 10 10 10 5 5 5 5 0 0 0 0 0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000 DGEMM (single-threaded) DSYRK (single-threaded) DSYMM (single-threaded) DTRMM (single-threaded) 15 15 15 15 10 10 10 10 GFLOPS GFLOPS GFLOPS GFLOPS 5 5 5 5 0 0 0 0 0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000 CGEMM (single-threaded) CSYRK (single-threaded) CHEMM (single-threaded) CTRMM (single-threaded) 30 30 30 30 25 25 25 25 20 20 20 20 GFLOPS GFLOPS GFLOPS GFLOPS 15 15 15 15 10 10 10 10 5 5 5 5 0 0 0 0 0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000 ZGEMM (single-threaded) ZSYRK (single-threaded) ZHEMM (single-threaded) ZTRMM (single-threaded) 15 15 15 15 GFLOPS 10 GFLOPS 10 GFLOPS 10 GFLOPS 10 5 5 5 5 BLIS OpenBLAS ARMPL 0 0 0 0 0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000 matrix dimension m=n=k matrix dimension m=n=k matrix dimension m=n=k matrix dimension m=n=k

  9. MulG-threaded BLIS (28 cores) SGEMM (multi-threaded) SSYRK (multi-threaded) SSYMM (multi-threaded) STRMM (multi-threaded) 800 800 800 800 600 600 600 600 GFLOPS GFLOPS GFLOPS GFLOPS 400 400 400 400 200 200 200 200 0 0 0 0 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 DGEMM (multi-threaded) DSYRK (multi-threaded) DSYMM (multi-threaded) DTRMM (multi-threaded) 400 400 400 400 300 300 300 300 GFLOPS GFLOPS GFLOPS GFLOPS 200 200 200 200 100 100 100 100 0 0 0 0 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 CGEMM (multi-threaded) CSYRK (multi-threaded) CHEMM (multi-threaded) CTRMM (multi-threaded) 800 800 800 800 600 600 600 600 GFLOPS GFLOPS GFLOPS GFLOPS 400 400 400 400 200 200 200 200 0 0 0 0 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 ZGEMM (multi-threaded) ZSYRK (multi-threaded) ZHEMM (multi-threaded) ZTRMM (multi-threaded) 400 400 400 400 300 300 300 300 GFLOPS GFLOPS GFLOPS GFLOPS 200 200 200 200 BLIS 100 100 100 100 OpenBLAS ARMPL 0 0 0 0 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 matrix dimension m=n=k matrix dimension m=n=k matrix dimension m=n=k matrix dimension m=n=k

  10. MulG-threaded BLIS (56 cores) SGEMM (multi-threaded) SSYRK (multi-threaded) SSYMM (multi-threaded) STRMM (multi-threaded) 1500 1500 1500 1500 GFLOPS GFLOPS GFLOPS GFLOPS 1000 1000 1000 1000 500 500 500 500 0 0 0 0 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 DGEMM (multi-threaded) DSYRK (multi-threaded) DSYMM (multi-threaded) DTRMM (multi-threaded) 800 800 800 800 600 600 600 600 GFLOPS GFLOPS GFLOPS GFLOPS 400 400 400 400 200 200 200 200 0 0 0 0 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 CGEMM (multi-threaded) CSYRK (multi-threaded) CHEMM (multi-threaded) CTRMM (multi-threaded) 1500 1500 1500 1500 GFLOPS GFLOPS GFLOPS GFLOPS 1000 1000 1000 1000 500 500 500 500 0 0 0 0 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 ZGEMM (multi-threaded) ZSYRK (multi-threaded) ZHEMM (multi-threaded) ZTRMM (multi-threaded) 800 800 800 800 600 600 600 600 GFLOPS GFLOPS GFLOPS GFLOPS 400 400 400 400 BLIS 200 200 200 200 OpenBLAS ARMPL 0 0 0 0 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 matrix dimension m=n=k matrix dimension m=n=k matrix dimension m=n=k matrix dimension m=n=k

  11. Other Architectures SkylakeX (single core) SGEMM (single-threaded) SSYRK (single-threaded) SSYMM (single-threaded) STRMM (single-threaded) 100 100 100 100 80 80 80 80 GFLOPS GFLOPS GFLOPS GFLOPS 60 60 60 60 40 40 40 40 20 20 20 20 0 0 0 0 0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000 DGEMM (single-threaded) DSYRK (single-threaded) DSYMM (single-threaded) DTRMM (single-threaded) 50 50 50 50 40 40 40 40 GFLOPS GFLOPS GFLOPS GFLOPS 30 30 30 30 20 20 20 20 10 10 10 10 0 0 0 0 0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000 CGEMM (single-threaded) CSYRK (single-threaded) CHEMM (single-threaded) CTRMM (single-threaded) 100 100 100 100 80 80 80 80 GFLOPS GFLOPS GFLOPS GFLOPS 60 60 60 60 40 40 40 40 20 20 20 20 0 0 0 0 0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000 ZGEMM (single-threaded) ZSYRK (single-threaded) ZHEMM (single-threaded) ZTRMM (single-threaded) 50 50 50 50 40 40 40 40 GFLOPS GFLOPS GFLOPS GFLOPS 30 30 30 30 20 20 20 20 BLIS OpenBLAS 10 10 10 10 MKL 0 0 0 0 0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000 matrix dimension m=n=k matrix dimension m=n=k matrix dimension m=n=k matrix dimension m=n=k

  12. Other Architectures SkylakeX (20 cores) SGEMM (multi-threaded) SSYRK (multi-threaded) SSYMM (multi-threaded) STRMM (multi-threaded) 2000 2000 2000 2000 1500 1500 1500 1500 GFLOPS GFLOPS GFLOPS GFLOPS 1000 1000 1000 1000 500 500 500 500 0 0 0 0 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 DGEMM (multi-threaded) DSYRK (multi-threaded) DSYMM (multi-threaded) DTRMM (multi-threaded) 1000 1000 1000 1000 800 800 800 800 GFLOPS GFLOPS GFLOPS GFLOPS 600 600 600 600 400 400 400 400 200 200 200 200 0 0 0 0 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 CGEMM (multi-threaded) CSYRK (multi-threaded) CHEMM (multi-threaded) CTRMM (multi-threaded) 2000 2000 2000 2000 1500 1500 1500 1500 GFLOPS GFLOPS GFLOPS GFLOPS 1000 1000 1000 1000 500 500 500 500 0 0 0 0 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 ZGEMM (multi-threaded) ZSYRK (multi-threaded) ZHEMM (multi-threaded) ZTRMM (multi-threaded) 1000 1000 1000 1000 800 800 800 800 GFLOPS GFLOPS GFLOPS GFLOPS 600 600 600 600 400 400 400 400 BLIS OpenBLAS 200 200 200 200 MKL 0 0 0 0 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 matrix dimension m=n=k matrix dimension m=n=k matrix dimension m=n=k matrix dimension m=n=k

Download Presentation
Download Policy: The content available on the website is offered to you 'AS IS' for your personal information and use only. It cannot be commercialized, licensed, or distributed on other websites without prior consent from the author. To download a presentation, simply click this link. If you encounter any difficulties during the download process, it's possible that the publisher has removed the file from their server.

Recommend

Packing - the next BLIS Fron5er? Tze Meng Low BLIS

Packing - the next BLIS Fron5er? Tze Meng Low BLIS

9/29/17 Packing - the next BLIS Fron5er? Tze Meng Low BLIS Retreat 2017 5 th loop around micro-kernel Looking back n C n C += C j A B j

115 views • 10 slides
Blis Connor Abbott, Wendy Pan, Klint Qinami, Jason Vaccaro Motivation: Why Blis? OpenGL is

Blis Connor Abbott, Wendy Pan, Klint Qinami, Jason Vaccaro Motivation: Why Blis? OpenGL is

Blis Connor Abbott, Wendy Pan, Klint Qinami, Jason Vaccaro Motivation: Why Blis? OpenGL is Complicated Architecture LLVM LLVM Backend Lexing, Sem. Src AST SAST Parsing Checking GLSL GLSL Backend Comparing OpenGL and Blis Code

455 views • 12 slides
Extending the BLIS Analytical Model for GPUs Elliot Binder, Claudia Kho, Doru Thom Popovici, Tze

Extending the BLIS Analytical Model for GPUs Elliot Binder, Claudia Kho, Doru Thom Popovici, Tze

Carnegie Mellon Extending the BLIS Analytical Model for GPUs Elliot Binder, Claudia Kho, Doru Thom Popovici, Tze Meng Low 18 September 2018 BLIS Retreat 1 Carnegie Mellon Many problems are MMM Popula5on Genomics k-Nearest Neighbours

181 views • 15 slides
Integrating DMA capabilities into BLIS for on-chip data movement Devangi Parikh Ilya

Integrating DMA capabilities into BLIS for on-chip data movement Devangi Parikh Ilya

Integrating DMA capabilities into BLIS for on-chip data movement Devangi Parikh Ilya Polkovnichenko Francisco Igual Pea Murtaza Ali 5 Generations of TI Multicore Processors Keystone architecture Lowers development effort

325 views • 20 slides
Mixing domains and precisions in BLIS: Ini5al thoughts Field G. Van Zee Science of High

Mixing domains and precisions in BLIS: Ini5al thoughts Field G. Van Zee Science of High

Mixing domains and precisions in BLIS: Ini5al thoughts Field G. Van Zee Science of High Performance Compu:ng The University of Texas at Aus:n The Problem gemm := + Lets simplify by omiFng scalars

504 views • 28 slides
HINARI: An Overview BY Samuel A Bello BLIS, MLIS, MIT, CLN Arcis Librarian University of

HINARI: An Overview BY Samuel A Bello BLIS, MLIS, MIT, CLN Arcis Librarian University of

15-Nov-17 HINARI: An Overview BY Samuel A Bello BLIS, MLIS, MIT, CLN Arcis Librarian University of Ibadan What is Hinari? It is an acronym meaning Health InterNetwork Access to Research Initiative. But now known as Hinari

83 views • 7 slides
Sharing a BLISful State Maggie Myers Devangi Parikh Robert van de Geijn Field Van Zee Our

Sharing a BLISful State Maggie Myers Devangi Parikh Robert van de Geijn Field Van Zee Our

Sharing a BLISful State Maggie Myers Devangi Parikh Robert van de Geijn Field Van Zee Our Purpose for this PresentaCon Gather your input on how to effecCvely Communicate existence of BLIS Share the idea that underlie BLIS Engage

412 views • 12 slides
Shared Memory Parallelization of MTTKRP for Dense Tensors BLIS Retreat 2017, September 18 th Koby

Shared Memory Parallelization of MTTKRP for Dense Tensors BLIS Retreat 2017, September 18 th Koby

Shared Memory Parallelization of MTTKRP for Dense Tensors BLIS Retreat 2017, September 18 th Koby Hayashi , Grey Ballard, Yujie Jiang, Michael Tobia hayakb13@,ballard@,jiany14@,tobiamj@wfu.edu Neuroimaging Application ensor: Time by Subjects by

491 views • 24 slides
An Analytical Model for BLIS Tze Meng Low 1 Francisco D. Igual 2 Tyler M. Smith 3 Enrique

An Analytical Model for BLIS Tze Meng Low 1 Francisco D. Igual 2 Tyler M. Smith 3 Enrique

An Analytical Model for BLIS Tze Meng Low 1 Francisco D. Igual 2 Tyler M. Smith 3 Enrique Quitana-Ortiz 4 1 Carnegie Mellon University 2 Universidad Complutense de Madrid, Spain 3 The University of Texas at Austin 4 Universidad Jaume I, Spain 2 nd

768 views • 28 slides
The BLIS Approach to Skinny Matrix Multiplication Field G. Van Zee Science of High Performance

The BLIS Approach to Skinny Matrix Multiplication Field G. Van Zee Science of High Performance

The BLIS Approach to Skinny Matrix Multiplication Field G. Van Zee Science of High Performance Computing The University of Texas at Austin September 19, 2019 Science of High Performance Computing (SHPC) research group Led by Robert A. van

1.21k views • 83 slides
Another year of progress for BLIS: 2017-2018 Field G. Van Zee Science of High Performance

Another year of progress for BLIS: 2017-2018 Field G. Van Zee Science of High Performance

Another year of progress for BLIS: 2017-2018 Field G. Van Zee Science of High Performance Compu:ng The University of Texas at Aus:n Science of High Performance Compu:ng (SHPC) research group Led by Robert A. van de Geijn Contributes to

771 views • 50 slides
VANGUARD Compliance Manager Support Compliance Customization Supports RACF and Performs

VANGUARD Compliance Manager Support Compliance Customization Supports RACF and Performs

VANGUARD Compliance Manager Support Compliance Customization Supports RACF and Performs critical compliance Performs specific custom ACF2 checks to adhere to industry baseline checks specific regulations VANGUARD Compliance Manager

343 views • 5 slides
Vice President Responsibilities: Acts in the absence of the President and performs other

Vice President Responsibilities: Acts in the absence of the President and performs other

Vice President Responsibilities: Acts in the absence of the President and performs other duties as assigned HSA Liaison to HSA Room Parent Coordinator Oversees HSA non-fundraising events and works with event chair(s), etc. Assists

96 views • 7 slides
Introduction to Magnetism (2) : Magnetism today or How performs the micromagnetic theory now?

Introduction to Magnetism (2) : Magnetism today or How performs the micromagnetic theory now?

Introduction to Magnetism (2) : Magnetism today or How performs the micromagnetic theory now? Andr Thiaville Laboratoire de Physique des Solides Universit Paris-Sud & CNRS, Orsay European School of Magnetism, 1 Constanta, 2005:

1.35k views • 89 slides
BLASFEO Gianluca Frison University of Freiburg BLIS retreat September 19, 2017 Gianluca Frison

BLASFEO Gianluca Frison University of Freiburg BLIS retreat September 19, 2017 Gianluca Frison

BLASFEO Gianluca Frison University of Freiburg BLIS retreat September 19, 2017 Gianluca Frison BLASFEO BLASFEO Basic Linear Algebra Subroutines For Embedded Optimization Gianluca Frison BLASFEO BLASFEO performance dgemm_nt 50 40

605 views • 31 slides
Smart Printing Solutions LTD What is SmartPrinter? A server based print engine. Performs

Smart Printing Solutions LTD What is SmartPrinter? A server based print engine. Performs

Smart Printing Solutions LTD What is SmartPrinter? A server based print engine. Performs print management and manipulation. Works with any PCL5 printer. Receives input date from any source application. What can it do? Language

658 views • 21 slides
Data Storage and Interaction using Magnetized Fabric Justin Chan Shyam Gollakota 1 Existing

Data Storage and Interaction using Magnetized Fabric Justin Chan Shyam Gollakota 1 Existing

Data Storage and Interaction using Magnetized Fabric Justin Chan Shyam Gollakota 1 Existing approaches PROJECT JACQUARD LILYPAD ARDUINO Requires batteries 2 Can we create a smart fabric design without any onboard electronics or

514 views • 36 slides
An Energy-Aware Method for the Joint Recognition of Activities and Gestures Using Wearable S

An Energy-Aware Method for the Joint Recognition of Activities and Gestures Using Wearable S

An Energy-Aware Method for the Joint Recognition of Activities and Gestures Using Wearable S ensors Joseph Korpela 1 Kazuyuki Takase 1 Takahiro Hirashima 1 Takuya Maekawa 1 Julien Eberle 2 Dipanj an Chakraborty 3 Karl Aberer

291 views • 18 slides
Gesture Detection A camera is used to detect gestures made in the cameras field of view using

Gesture Detection A camera is used to detect gestures made in the cameras field of view using

Gesture Detection A camera is used to detect gestures made in the cameras field of view using infrared LEDs. Detected gestures will be used to control a music player. Andrew Zhong Patrick Kuczera Shahzeb Asif 1 Motivation/Goals The

230 views • 10 slides
Point Cloud based Gesture Recognition with Kinect 2 Anton Klarn, Jonathan Karlsson Kinect v2

Point Cloud based Gesture Recognition with Kinect 2 Anton Klarn, Jonathan Karlsson Kinect v2

Point Cloud based Gesture Recognition with Kinect 2 Anton Klarn, Jonathan Karlsson Kinect v2 2.5D Sensor (Depth) Time-of-flight sensor Full HD color camera 4-microphone array 30+ frame rate (Demo of Point Cloud)

640 views • 16 slides
GPU peak performance vs. CPU Squeezing GPU performance Peak Double Precision FLOPS Peak Memory

GPU peak performance vs. CPU Squeezing GPU performance Peak Double Precision FLOPS Peak Memory

GPU peak performance vs. CPU Squeezing GPU performance Peak Double Precision FLOPS Peak Memory Bandwidth GPGPU 2015: High Performance Computing with CUDA University of Cape Town (South Africa), April, 20 th -24 th , 2015 GPU 6x faster on

178 views • 4 slides
Module 4.1 Memory and Data Locality CUDA Memories Objective To learn to effectively use

Module 4.1 Memory and Data Locality CUDA Memories Objective To learn to effectively use

GPU Teaching Kit Accelerated Computing Module 4.1 Memory and Data Locality CUDA Memories Objective To learn to effectively use the CUDA memory types in a parallel program Importance of memory access efficiency Registers,

187 views • 17 slides
Application of Many-core Accelerators for Problems in Astronomy and Physics N.Nakasato

Application of Many-core Accelerators for Problems in Astronomy and Physics N.Nakasato

Application of Many-core Accelerators for Problems in Astronomy and Physics N.Nakasato (University of Aizu, Japan) in collaboration with F.Yuasa, T.Ishikawa, J.Makino, H.Daisaka No.2 Agenda Our Problems Recent Development of Many-core

641 views • 38 slides
Nikolay Khokhlov, MIPT Quasilinear equations, inverse problems and their applications Moscow

Nikolay Khokhlov, MIPT Quasilinear equations, inverse problems and their applications Moscow

Applying OpenCL technology for seismic modeling using grid-characteristic methods Andrey Ivanov, MIPT Nikolay Khokhlov, MIPT Quasilinear equations, inverse problems and their applications Moscow Institute of Physics and Technology, Dolgoprudny,

414 views • 23 slides

More recommend