Marat Dukhan
I am Ph.D. student in Computational Science and Engineering at the Georgia Tech's College of Computing. My research interests are in high-performance computing, machine learning, and their interaction.
Publications
Conference and Journal Publications (peer reviewed)
PMMA 16, June 23, 2016
Marat Dukhan, Richard Vuduc, Jason Riedy "Wanted: Floating-Point Add Round-off Error instruction"
Slides
Preprint on ArXiv
Source code on GitHub
Discussion on GitXiv
PyHPC 2015, November 15, 2015
Marat Dukhan "PeachPy meets Opcodes: Direct Machine Code Generation from Python"
International Journal of High-Performance Computing Applications, July 2, 2015
Edmond Chow, Xing Liu, Sanchit Misra, Marat Dukhan, Mikhail Smelyanskiy, Jeff R. Hammond, Yunfei Du, Xiang-Ke Liao, Pradeep Dubey "Scaling up Hartree–Fock calculations on Tianhe-2"
SPAA 2015, June 13-15, 2015
Oded Green, Marat Dukhan, Richard Vuduc "Branch-Avoiding Graph Algorithms"
HPSL 2015, February 7, 2015
Marat Dukhan, Robert Guthrie, Robertson Taylor, Richard Vuduc "Furious.js: a Model for Offloading Compute-Intensive JavaScript Applications"
IPDPS 2014, May 19-23, 2014
Jee Choi, Marat Dukhan, Xing Liu, Richard Vuduc "Algorithmic time, energy, and power on candidate HPC compute building blocks"
PyHPC 2013, November 18, 2013
Marat Dukhan "PeachPy: A Python Framework for Developing High-Performance Assembly Kernels"
PPAM 10, September 10, 2013
Marat Dukhan and Richard Vuduc "Methods for high-throughput computation of elementary functions"
SC 2012, November 10-16, 2012
William B. March, Kenneth Czechowski, Marat Dukhan, Thomas Benson, Dongryeol Lee, Andrew J. Connolly, Richard Vuduc, Edmond Chow, and Alexander G. Gray "Optimizing the Computation of N-Point Correlations on Large-Scale Astronomical Data."
Poster Presentations (peer reviewed)
The 7th Scientific Software Days conference, February 25–26, 2016
Marat Dukhan, Anirudh Gubba, Richard Vuduc "Crowdsourcing Code Tuning with PeachPy.io"
Hot Chips 25, August 25–28, 2013
Marat Dukhan "What a Fast FPU Means for Algorithms: A Story of Vector Elementary Functions"
Other Presentations (not peer-reviewed)
BLIS Retreat 2016 workshop, September 19–20, 2016
Marat Dukhan "BLAS for Deep Learning"
BLIS Retreat 2016 workshop, September 19–20, 2016
Marat Dukhan "PeachPy.io: a platform for crowdsourcing performance tuning"
ATIP 2015 workshop, November 16, 2015
Marat Dukhan, Nicolas Vasilache, Soumith Chintala, Richard Vuduc "FFT-based convolutional neural networks for wide-SIMD multi-core CPUs"
BLIS Retreat 2015 workshop, September 28–29, 2015
Marat Dukhan "Porting BLIS micro-kernels to PeachPy"
Atlanta Go Meetup, September 16, 2015
Marat Dukhan "Accelerating Data Processing in Go with SIMD Instructions"
BLIS Retreat 2014 workshop, September 25–26, 2014
Marat Dukhan "BLIS for the Web"
The 1st BLIS Retreat workshop, September 5–6, 2013
Marat Dukhan "Developing low-level assembly kernels with Peach-Py"
Teaching
Spring 2014
Linear Algebra — Foundations to Frontiers (MOOC class on edX)
Prepared two live demos for Robert van de Geijn's and Maggie Myers' Linear Algebra — Foundations to Frontiers MOOC course
Fall 2012
CSE 6230 — High Performance Computing: Tools and Applications
Assisted Prof. Richard Vuduc and fully covered three weeks of the class
- Week 6
- Lecture 6: SIMD Optimization
- Lecture 7: Advanced SIMD Optimization
- Lab 5: Basic SIMD Optimization
- Week 7
- Lecture 8: IPC and friends
- Lecture 9: IPC and friends, continued
- Lab 6: Extracting ILP for Fun and Profit
- Week 8
- Lecture 10: Memory Optimizations
- Lab 7: Grabbing Cycles Here and There
- Week 16
- Lecture 18: Distribution of High-Performance Codes and the Future of HPC
Fall 2013
CSE 6230 — High Performance Computing: Tools and Applications
Assisted Prof. Richard Vuduc and fully covered four weeks of the class
- Week 6
- Performance Tuning Dive-In: Profiling, Compiling, and Libraries lecture
- Lab 5: Profiling
- Week 10
- Performance Tuning for CPU: SIMD Optimization lecture
- Performance Tuning for CPU: Advanced SIMD Optimization lecture
- Lab 8: SIMD For Image Processing and Computer Vision
- Week 11
- Assembly 101, Microarchitecture 101, IPC and Friends lecture
- Lab 9: Floating-Point SIMD
- Week 12
- Performance Tuning for CPU: Memory Optimizations lecture
- Lab 10: Sparse Operations
- Week 16
- MapReduce and Special Topics in HPC lecture
Fall 2013
CX 4010 / CSE 8803 — Computational Problem Solving for Scientists and Engineers
Assisted Prof. Alberto Apostolico and presented one lecture each week
- Macros lecture slides
- CPU Instructions lecture slides
- Assembly and Calling Conventions lecture slides
- C and C++ Practice lecture slides
- More C++ Practice lecture slides
- Debugging of C and C++ Codes lecture slides
- Using Command-Line Arguments in C++ lecture slides
- Working with Remote Machines lecture slides
- C++ STL lecture slides
- Least Squares Problem lecture slides
- Data Processing in C++ lecture slides
- C++ for Simulation lecture slides
- C++ Interfaces lecture slides
- Python Basics lecture slides
- Python Extensions lecture slides
- Revisiting Interfaces, Challenges in CSE, and MapReduce lecture slides
Software
Furious.js JavaScript Library
JavaScript library for hardware-accelerated scientific computing
- JavaScript tensor library similar to NumPy/SciPy
- Asynchronous computations without excessive callbacks
- Transparent offload of computations to Portable Native Client, Web Worker, WebCL or cloud server (via Web Sockets)
PeachPy: Portable Efficient Assembly Code-generation in High-level Python
Python Framework for Automating Development of High-Performance Assembly Kernels
- Supports automatic register allocation
- Provides stack frame management, including re-aligning of stack frame as needed
- Generates versions of a function for different calling conventions from the same source (e.g. functions for Microsoft x64 ABI and System V x86-64 ABI can be generated from the same source)
- Allows to define constants in the place where they are used (just like in high-level languages)
- Tracks of instruction extensions used in the function.
- Can multiplex multiple instruction streams (helpful for software pipelining)
Yeppp! High-performance library
Provides a collection of low-level functions optimized for modern processors
- Library functions have multiple implementations, optimized for different processor architectures
- The optimal function is chosen at run-time depending on the processor
- Detects processor microarchitecture and instruction set extensions
- Provides portable access to CPU cycle counter and high-resolution system timer
- Available for Windows, Linux, Mac OS X, and Android
- Supports x86, x86-64, ARM, MIPS, and PowerPC architectures
- C and C++-compatible header files, and bindings for FORTRAN, Java and .Net/Mono
- BSD license
Yeppp! CPUID for Android
Shows detailed information about the mobile CPU:
- CPU architecture (ARM, x86, or MIPS)
- CPU vendor (e.g. ARM, Qualcomm, Intel, MIPS)
- CPU microarchitecture (e.g. ARM11, Cortex-A9, Atom, XBurst)
- Minimum and maximum frequency
- Number of logical cores
- Supported instruction set extensions (e.g. NEON, VFPv4, SSSE3, MIPS3D)
- Size of level-1, level-2, and level-3 caches.
Education
2011 – now
Georgia Institute of Technology, College of Computing
Candidate for Ph.D. in Computational Science and Engineering
- Research advisor: Richard Vuduc
2009 – 2011
New Economic School in Moscow
M.A. in Economics
- Master's thesis: "Regime Switching Autoregression and Volatility Jumps"
2005 – 2009
Moscow Institute of Physics and Technology
B.Sc. in Applied Mathematics and Physics
- Bachelor's thesis: "Entropy Coding Optimization for H.264 Video Codec"
Contact Information
E-mail: 
Marat Dukhan, Richard Vuduc, Jason Riedy "Wanted: Floating-Point Add Round-off Error instruction"
Slides Preprint on ArXiv Source code on GitHub Discussion on GitXiv
Marat Dukhan "PeachPy meets Opcodes: Direct Machine Code Generation from Python"
Edmond Chow, Xing Liu, Sanchit Misra, Marat Dukhan, Mikhail Smelyanskiy, Jeff R. Hammond, Yunfei Du, Xiang-Ke Liao, Pradeep Dubey "Scaling up Hartree–Fock calculations on Tianhe-2"
Oded Green, Marat Dukhan, Richard Vuduc "Branch-Avoiding Graph Algorithms"
Marat Dukhan, Robert Guthrie, Robertson Taylor, Richard Vuduc "Furious.js: a Model for Offloading Compute-Intensive JavaScript Applications"
Jee Choi, Marat Dukhan, Xing Liu, Richard Vuduc "Algorithmic time, energy, and power on candidate HPC compute building blocks"
Marat Dukhan "PeachPy: A Python Framework for Developing High-Performance Assembly Kernels"
Marat Dukhan and Richard Vuduc "Methods for high-throughput computation of elementary functions"
William B. March, Kenneth Czechowski, Marat Dukhan, Thomas Benson, Dongryeol Lee, Andrew J. Connolly, Richard Vuduc, Edmond Chow, and Alexander G. Gray "Optimizing the Computation of N-Point Correlations on Large-Scale Astronomical Data."
Marat Dukhan, Anirudh Gubba, Richard Vuduc "Crowdsourcing Code Tuning with PeachPy.io"
Marat Dukhan "What a Fast FPU Means for Algorithms: A Story of Vector Elementary Functions"
Other Presentations (not peer-reviewed)
BLIS Retreat 2016 workshop, September 19–20, 2016
Marat Dukhan "BLAS for Deep Learning"
BLIS Retreat 2016 workshop, September 19–20, 2016
Marat Dukhan "PeachPy.io: a platform for crowdsourcing performance tuning"
ATIP 2015 workshop, November 16, 2015
Marat Dukhan, Nicolas Vasilache, Soumith Chintala, Richard Vuduc "FFT-based convolutional neural networks for wide-SIMD multi-core CPUs"
BLIS Retreat 2015 workshop, September 28–29, 2015
Marat Dukhan "Porting BLIS micro-kernels to PeachPy"
Atlanta Go Meetup, September 16, 2015
Marat Dukhan "Accelerating Data Processing in Go with SIMD Instructions"
BLIS Retreat 2014 workshop, September 25–26, 2014
Marat Dukhan "BLIS for the Web"
The 1st BLIS Retreat workshop, September 5–6, 2013
Marat Dukhan "Developing low-level assembly kernels with Peach-Py"
Teaching
Spring 2014
Linear Algebra — Foundations to Frontiers (MOOC class on edX)
Prepared two live demos for Robert van de Geijn's and Maggie Myers' Linear Algebra — Foundations to Frontiers MOOC course
Fall 2012
CSE 6230 — High Performance Computing: Tools and Applications
Assisted Prof. Richard Vuduc and fully covered three weeks of the class
- Week 6
- Lecture 6: SIMD Optimization
- Lecture 7: Advanced SIMD Optimization
- Lab 5: Basic SIMD Optimization
- Week 7
- Lecture 8: IPC and friends
- Lecture 9: IPC and friends, continued
- Lab 6: Extracting ILP for Fun and Profit
- Week 8
- Lecture 10: Memory Optimizations
- Lab 7: Grabbing Cycles Here and There
- Week 16
- Lecture 18: Distribution of High-Performance Codes and the Future of HPC
Fall 2013
CSE 6230 — High Performance Computing: Tools and Applications
Assisted Prof. Richard Vuduc and fully covered four weeks of the class
- Week 6
- Performance Tuning Dive-In: Profiling, Compiling, and Libraries lecture
- Lab 5: Profiling
- Week 10
- Performance Tuning for CPU: SIMD Optimization lecture
- Performance Tuning for CPU: Advanced SIMD Optimization lecture
- Lab 8: SIMD For Image Processing and Computer Vision
- Week 11
- Assembly 101, Microarchitecture 101, IPC and Friends lecture
- Lab 9: Floating-Point SIMD
- Week 12
- Performance Tuning for CPU: Memory Optimizations lecture
- Lab 10: Sparse Operations
- Week 16
- MapReduce and Special Topics in HPC lecture
Fall 2013
CX 4010 / CSE 8803 — Computational Problem Solving for Scientists and Engineers
Assisted Prof. Alberto Apostolico and presented one lecture each week
- Macros lecture slides
- CPU Instructions lecture slides
- Assembly and Calling Conventions lecture slides
- C and C++ Practice lecture slides
- More C++ Practice lecture slides
- Debugging of C and C++ Codes lecture slides
- Using Command-Line Arguments in C++ lecture slides
- Working with Remote Machines lecture slides
- C++ STL lecture slides
- Least Squares Problem lecture slides
- Data Processing in C++ lecture slides
- C++ for Simulation lecture slides
- C++ Interfaces lecture slides
- Python Basics lecture slides
- Python Extensions lecture slides
- Revisiting Interfaces, Challenges in CSE, and MapReduce lecture slides
Software
Furious.js JavaScript Library
JavaScript library for hardware-accelerated scientific computing
- JavaScript tensor library similar to NumPy/SciPy
- Asynchronous computations without excessive callbacks
- Transparent offload of computations to Portable Native Client, Web Worker, WebCL or cloud server (via Web Sockets)
PeachPy: Portable Efficient Assembly Code-generation in High-level Python
Python Framework for Automating Development of High-Performance Assembly Kernels
- Supports automatic register allocation
- Provides stack frame management, including re-aligning of stack frame as needed
- Generates versions of a function for different calling conventions from the same source (e.g. functions for Microsoft x64 ABI and System V x86-64 ABI can be generated from the same source)
- Allows to define constants in the place where they are used (just like in high-level languages)
- Tracks of instruction extensions used in the function.
- Can multiplex multiple instruction streams (helpful for software pipelining)
Yeppp! High-performance library
Provides a collection of low-level functions optimized for modern processors
- Library functions have multiple implementations, optimized for different processor architectures
- The optimal function is chosen at run-time depending on the processor
- Detects processor microarchitecture and instruction set extensions
- Provides portable access to CPU cycle counter and high-resolution system timer
- Available for Windows, Linux, Mac OS X, and Android
- Supports x86, x86-64, ARM, MIPS, and PowerPC architectures
- C and C++-compatible header files, and bindings for FORTRAN, Java and .Net/Mono
- BSD license
Yeppp! CPUID for Android
Shows detailed information about the mobile CPU:
- CPU architecture (ARM, x86, or MIPS)
- CPU vendor (e.g. ARM, Qualcomm, Intel, MIPS)
- CPU microarchitecture (e.g. ARM11, Cortex-A9, Atom, XBurst)
- Minimum and maximum frequency
- Number of logical cores
- Supported instruction set extensions (e.g. NEON, VFPv4, SSSE3, MIPS3D)
- Size of level-1, level-2, and level-3 caches.
Education
2011 – now
Georgia Institute of Technology, College of Computing
Candidate for Ph.D. in Computational Science and Engineering
- Research advisor: Richard Vuduc
2009 – 2011
New Economic School in Moscow
M.A. in Economics
- Master's thesis: "Regime Switching Autoregression and Volatility Jumps"
2005 – 2009
Moscow Institute of Physics and Technology
B.Sc. in Applied Mathematics and Physics
- Bachelor's thesis: "Entropy Coding Optimization for H.264 Video Codec"
Contact Information
E-mail:
Marat Dukhan "BLAS for Deep Learning"
Marat Dukhan "PeachPy.io: a platform for crowdsourcing performance tuning"
Marat Dukhan, Nicolas Vasilache, Soumith Chintala, Richard Vuduc "FFT-based convolutional neural networks for wide-SIMD multi-core CPUs"
Marat Dukhan "Porting BLIS micro-kernels to PeachPy"
Marat Dukhan "Accelerating Data Processing in Go with SIMD Instructions"
Marat Dukhan "BLIS for the Web"
Marat Dukhan "Developing low-level assembly kernels with Peach-Py"
Spring 2014
Linear Algebra — Foundations to Frontiers (MOOC class on edX)
Prepared two live demos for Robert van de Geijn's and Maggie Myers' Linear Algebra — Foundations to Frontiers MOOC course
Fall 2012
CSE 6230 — High Performance Computing: Tools and Applications
Assisted Prof. Richard Vuduc and fully covered three weeks of the class
- Week 6
- Lecture 6: SIMD Optimization
- Lecture 7: Advanced SIMD Optimization
- Lab 5: Basic SIMD Optimization
- Week 7
- Lecture 8: IPC and friends
- Lecture 9: IPC and friends, continued
- Lab 6: Extracting ILP for Fun and Profit
- Week 8
- Lecture 10: Memory Optimizations
- Lab 7: Grabbing Cycles Here and There
- Week 16
- Lecture 18: Distribution of High-Performance Codes and the Future of HPC
Fall 2013
CSE 6230 — High Performance Computing: Tools and Applications
Assisted Prof. Richard Vuduc and fully covered four weeks of the class
- Week 6
- Performance Tuning Dive-In: Profiling, Compiling, and Libraries lecture
- Lab 5: Profiling
- Week 10
- Performance Tuning for CPU: SIMD Optimization lecture
- Performance Tuning for CPU: Advanced SIMD Optimization lecture
- Lab 8: SIMD For Image Processing and Computer Vision
- Week 11
- Assembly 101, Microarchitecture 101, IPC and Friends lecture
- Lab 9: Floating-Point SIMD
- Week 12
- Performance Tuning for CPU: Memory Optimizations lecture
- Lab 10: Sparse Operations
- Week 16
- MapReduce and Special Topics in HPC lecture
Fall 2013
CX 4010 / CSE 8803 — Computational Problem Solving for Scientists and Engineers
Assisted Prof. Alberto Apostolico and presented one lecture each week
- Macros lecture slides
- CPU Instructions lecture slides
- Assembly and Calling Conventions lecture slides
- C and C++ Practice lecture slides
- More C++ Practice lecture slides
- Debugging of C and C++ Codes lecture slides
- Using Command-Line Arguments in C++ lecture slides
- Working with Remote Machines lecture slides
- C++ STL lecture slides
- Least Squares Problem lecture slides
- Data Processing in C++ lecture slides
- C++ for Simulation lecture slides
- C++ Interfaces lecture slides
- Python Basics lecture slides
- Python Extensions lecture slides
- Revisiting Interfaces, Challenges in CSE, and MapReduce lecture slides
Software
Furious.js JavaScript Library
JavaScript library for hardware-accelerated scientific computing
- JavaScript tensor library similar to NumPy/SciPy
- Asynchronous computations without excessive callbacks
- Transparent offload of computations to Portable Native Client, Web Worker, WebCL or cloud server (via Web Sockets)
PeachPy: Portable Efficient Assembly Code-generation in High-level Python
Python Framework for Automating Development of High-Performance Assembly Kernels
- Supports automatic register allocation
- Provides stack frame management, including re-aligning of stack frame as needed
- Generates versions of a function for different calling conventions from the same source (e.g. functions for Microsoft x64 ABI and System V x86-64 ABI can be generated from the same source)
- Allows to define constants in the place where they are used (just like in high-level languages)
- Tracks of instruction extensions used in the function.
- Can multiplex multiple instruction streams (helpful for software pipelining)
Yeppp! High-performance library
Provides a collection of low-level functions optimized for modern processors
- Library functions have multiple implementations, optimized for different processor architectures
- The optimal function is chosen at run-time depending on the processor
- Detects processor microarchitecture and instruction set extensions
- Provides portable access to CPU cycle counter and high-resolution system timer
- Available for Windows, Linux, Mac OS X, and Android
- Supports x86, x86-64, ARM, MIPS, and PowerPC architectures
- C and C++-compatible header files, and bindings for FORTRAN, Java and .Net/Mono
- BSD license
Yeppp! CPUID for Android
Shows detailed information about the mobile CPU:
- CPU architecture (ARM, x86, or MIPS)
- CPU vendor (e.g. ARM, Qualcomm, Intel, MIPS)
- CPU microarchitecture (e.g. ARM11, Cortex-A9, Atom, XBurst)
- Minimum and maximum frequency
- Number of logical cores
- Supported instruction set extensions (e.g. NEON, VFPv4, SSSE3, MIPS3D)
- Size of level-1, level-2, and level-3 caches.
Education
| 2011 – now | Georgia Institute of Technology, College of Computing |
| Candidate for Ph.D. in Computational Science and Engineering | |
|
|
| 2009 – 2011 | New Economic School in Moscow |
| M.A. in Economics | |
|
|
| 2005 – 2009 | Moscow Institute of Physics and Technology |
| B.Sc. in Applied Mathematics and Physics | |
|
Contact Information
E-mail: