DOCSLIB.ORG

Advanced Parallel Programming II

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Advanced Parallel Programming II Advanced Parallel Programming II Alexander Leutgeb, RISC Software GmbH RISC Software GmbH – Johannes Kepler University Linz © 2016 22.09.2016 | 1 Introduction to Vectorization RISC Software GmbH – Johannes Kepler University Linz © 2016 22.09.2016 | 2 Motivation . Increasement in number of cores – Threading techniques to improve performance . But flops per cycle of vector units increased as much as number of cores . No use of vector units wasting flops/watt . For best performance – Use all cores – Efficient use of vector units – Ignoring potential of vector units is so inefficient as using only one core RISC Software GmbH – Johannes Kepler University Linz © 2016 22.09.2016 | 3 Vector Unit . Single Instruction Multiple Data (SIMD) units . Mostly for floating point operations . Data parallelization with one instruction – 64-Bit unit 1 DP flop, 2 SP flop – 128-Bit unit 2 DP flop, 4 SP flop – … . Multiple data elements are loaded into vector registers and used by vector units . Some architectures have more than one instruction per cycle (e.g. Sandy Bridge) RISC Software GmbH – Johannes Kepler University Linz © 2016 22.09.2016 | 4 Parallel Execution Scalar version works on Vector version carries out the same instructions one element at a time on many elements at a time a[i] = b[i] + c[i] x d[i]; a[i:8] = b[i:8] + c[i:8] * d[i:8]; a[i] a[i] a[i+1] a[i+2] a[i+3] a[i+4] a[i+5] a[i+6] a[i+7] = = = = = = = = = b[i] b[i] b[i+1] b[i+2] b[i+3] b[i+4] b[i+5] b[i+6] b[i+7] + + + + + + + + + c[i] c[i] c[i+1] c[i+2] c[i+3] c[i+4] c[i+5] c[i+6] c[i+7] x x x x x x x x x d[i] d[i] d[i+1] d[i+2] d[i+3] d[i+4] d[i+5] d[i+6] d[i+7] RISC Software GmbH – Johannes Kepler University Linz © 2016 22.09.2016 | 5 Vector Registers RISC Software GmbH – Johannes Kepler University Linz © 2016 22.09.2016 | 6 Vector Unit Usage (Programmers View) Use vectorized libraries Ease of use (e.g. Intel MKL) Fully automatic vectorization Auto vectorization hints (#pragma ivdep) SIMD feature (#pragma simd and simd function annotation) Vector intrinsics (e.g. mm_add_ps()) ASM code Programmer control (e.g. addps) RISC Software GmbH – Johannes Kepler University Linz © 2016 22.09.2016 | 7 Auto Vectorization RISC Software GmbH – Johannes Kepler University Linz © 2016 22.09.2016 | 8 Auto Vectorization . Modern compilers analyse loops in serial code identification for vectorization – Perform loop transformations for identification . Usage of instruction set of target architecture RISC Software GmbH – Johannes Kepler University Linz © 2016 22.09.2016 | 9 Common Compiler Switches GCC and ICC Functionality Switch Disable optimization -O0 Optimize for speed (no code size increase) -O1 Optimize for speed (default) -O2 High-level optimizer (e.g. loop unroll) -O3 Aggressive optimizations (e.g. –ipo, -O3, …) -fast Create symbols for debugging -g Generate assembly files -S OpenMP support -openmp RISC Software GmbH – Johannes Kepler University Linz © 2016 22.09.2016 | 10 Architecture Specific Compiler Switches GCC Functionality Switch Optimize for current machine -march=native Generate SSE v1 code -msse Generate SSE v2 code (default, may also emit SSE v1 code) -msse2 Generate SSE v3 code (may also emit SSE v1 and v2 code) -msse3 Generrate SSSE v3 code (may also emit SSE v1, v2, and v3 code) -mssse3 Generate SSE4.1 code (may also emit (S)SSEE v1, v2, and v3 code) -msse4.1 Generate SSE4.2 code (may also emit (S)SSE v1, v2, v3, and v4 code) -msse4.2 Generate AVX code -mavx Generate AVX v2 code -mavx2 RISC Software GmbH – Johannes Kepler University Linz © 2016 22.09.2016 | 11 Architecture Specific Compiler Switches ICC Functionality Switch * Optimize for current machine -xHOST Generate SSE v1 code -xSSE1 Generate SSE v2 code (default, may also emit SSE v1 code) -xSSE2 Generate SSE v3 code (may also emit SSE v1 and v2 code) -xSSE3 Generate SSE v3 code for Atom-based processors -xSSE_ATOM Generrate SSSE v3 code (may also emit SSE v1, v2, and v3 code) -xSSSE3 Generate SSE4.1 code (may also emit (S)SSEE v1, v2, and v3 code) -xSSE4.1 Generate SSE4.2 code (may also emit (S)SSE v1, v2, v3, and v4 code) -xSSE4.2 Generate AVX code -xAVX * For Intel processors use –x, for non-Intel processors use -m RISC Software GmbH – Johannes Kepler University Linz © 2016 22.09.2016 | 12 Example 4 Simple Vector Addition 1. Go to the directory example_4. 2. Compile simple.cpp (no vectorization) and execute the binary icc -std=c++11 -O2 -no-vec simple.cpp (g++ -std=c++11 -O2 simple.cpp) 3. Compile simple.cpp (vectorization) and execute the binary icc -std=c++11 -O2 simple.cpp (g++ -std=c++11 -O2 -ftree-vectorize) 4. What is the difference considering the execution times? 5. Can the execution times be further improved? RISC Software GmbH – Johannes Kepler University Linz © 2016 22.09.2016 | 13 SIMD Vectorization Basics . Vectorization offers good performance improvement on floating point intensive code . Vectorized code could compute slightly different results than non vectorized (x87 FPU – 80 Bit, SIMD – 64 Bit) . Even for scalar operations vector unit is used . Vectorization is only one aspect to improve performance – Efficent use of the cache is necessary RISC Software GmbH – Johannes Kepler University Linz © 2016 22.09.2016 | 14 Parallelization at No Cost . We tell the compiler to vectorize – That‘s not the whole story . There are cases, where a compiler cannot vectorize the code . How can we analyse such situations vectorization report . Generation of vectorization report via the compiler – More interesting what was not done and why – Focus on code paths, which were not vectorized RISC Software GmbH – Johannes Kepler University Linz © 2016 22.09.2016 | 15 Example 5 Vectorization Report 1. Go to the directory example_5. 2. Compile simple.cpp with enabled vector report generation: icc -std=c++11 -O2 -vec-report=2 simple.cpp (g++ -std=c++11 -O2 -ftree-vectorize –fopt-info-vec-missed simple.cpp) 3. Which positive/negative information does the vectorization report tell? 4. Insert the following code before std::cout in calcSP() and compile again. Have a look at the vectorization report. for (i = 1; i < VECSIZE; i++) { a[i] = a[i] + a[i – 1]; } RISC Software GmbH – Johannes Kepler University Linz © 2016 22.09.2016 | 16 Vectorization Report ICC . Information – Which code was vectorized? – Which code was not vectorized? . Compiler switch –vec-report<n> – n=0: no diagnostic information – n=1: (default) vectorized loops – n=2: vectorized/non vectorized loops (and why) – n=3: additional dependency information – n=4: only non vectorized loops – n=5: only non vectorized loops and dependency information – n=6: vectorized/non vectorized loops with details RISC Software GmbH – Johannes Kepler University Linz © 2016 22.09.2016 | 17 Loop Unrolling for (i = 0; i < N; i++) Unrolling allows compiler { to reconstruct loop for a[i] = b[i] * c[i]; vector operations } for (i = 0; i < N; i += 4) { Load b(i, …, i + 3) a[i ] = b[i ] * c[i ]; Load c(i, …, i + 3) a[i + 1] = b[i + 1] * c[i + 1]; Operate b * c -> a a[i + 2] = b[i + 2] * c[i + 2]; Store a(i, …, i + 3) a[i + 3] = b[i + 3] * c[i + 3]; } RISC Software GmbH – Johannes Kepler University Linz © 2016 22.09.2016 | 18 Requirements for Auto Vectorization . Countable . Single entry and exit while (i < 100) { a[i] = b[i] * c[i]; if (a[i] < 0.0) // data-dependent exit condition: break; ++i; } // loop not vectorized . Straight-line code (no switch, if with masking) for (int i = 0; i < length; i++) { float s = b[i] * b[i] – 4 * a[i] * c[i]; if (s >= 0) x[i] = sqrt(s); else x[i] = 0.; } // loop vectorized (because of masking) RISC Software GmbH – Johannes Kepler University Linz © 2016 22.09.2016 | 19 Requirements for Auto Vectorization . Only most inner loop (caution in case of loop interchange or loop collapsing) . No functions calls, but – Instrinsic math (sin, log, …) – Inline functions – Elemental functions __attribute__((vector)) RISC Software GmbH – Johannes Kepler University Linz © 2016 22.09.2016 | 20 Inhibitors of Auto Vectorization . Non contiguous data // arrays accessed with stride 2 for (int i=0; i<SIZE; i+=2) b[i]+=a[i]+x[i]; // inner loop accesses a with stride SIZE for (int j=0; j<SIZE; j++) for (int i=0; i<SIZE; i++) b[i]+=a[i][j]*x[j]; // indirect addressing of x using index array for (int i=0; i<SIZE; i+=2) b[i]+=a[i]*x[index[i]]; stride 1 is best. Caution in case of more dimensional arrays: do j=1,n F90 C for (j=0; j<n; j++) do i=1,n for (i=0;i<n;i++) a(i,j)=b(i,j)*s a[j][i]=b[j][i]*s; enddo enddo RISC Software GmbH – Johannes Kepler University Linz © 2016 22.09.2016 | 21 Inhibitors of Auto Vectorization . Inability to identify data with alias (or overlapping) – Runtime check possible multi versioned (vectorized/non vectorized) void my_cp(int nx, double* a, double* b) { for (int i = 0; i < nx; i++) a[i] = b[i]; } – Runtime check not possible non vectorized void my_combine(int* ioff, int nx, double* a, double* b, double* c) { for (int i = 0; i < nx; i++) { a[i] = b[i] + c[i + *ioff]; } } Would vectorize with strict aliasing (-ansi-alias) RISC Software GmbH – Johannes Kepler University Linz © 2016 22.09.2016 | 22 Inhibitors of Auto Vectorization .
Load more

Recommended publications
  • Vectorization Optimization
    Vectorization Optimization Klaus-Dieter Oertel Intel IAGS HLRN User Workshop, 3-6 Nov 2020 Optimization Notice Copyright © 2020, Intel Corporation. All rights reserved. *Other names and brands may be claimed as the property of others. Changing Hardware Impacts Software More Cores → More Threads → Wider Vectors Intel® Xeon® Processor Intel® Xeon Phi™ processor 5100 5500 5600 E5-2600 E5-2600 E5-2600 Platinum 64-bit E5-2600 Knights Landing series series series V2 V3 V4 8180 Core(s) 1 2 4 6 8 12 18 22 28 72 Threads 2 2 8 12 16 24 36 44 56 288 SIMD Width 128 128 128 128 256 256 256 256 512 512 High performance software must be both ▪ Parallel (multi-thread, multi-process) ▪ Vectorized *Product specification for launched and shipped products available on ark.intel.com. Optimization Notice Copyright © 2020, Intel Corporation. All rights reserved. 2 *Other names and brands may be claimed as the property of others. Vectorize & Thread or Performance Dies Threaded + Vectorized can be much faster than either one alone Vectorized & Threaded “Automatic” Vectorization Not Enough Explicit pragmas and optimization often required The Difference Is Growing With Each New 130x Generation of Threaded Hardware Vectorized Serial 2010 2012 2013 2014 2016 2017 Intel® Xeon™ Intel® Xeon™ Intel® Xeon™ Intel® Xeon™ Intel® Xeon™ Intel® Xeon® Platinum Processor Processor Processor Processor Processor Processor X5680 E5-2600 E5-2600 v2 E5-2600 v3 E5-2600 v4 81xx formerly codenamed formerly codenamed formerly codenamed formerly codenamed formerly codenamed formerly codenamed Westmere Sandy Bridge Ivy Bridge Haswell Broadwell Skylake Server Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors.
    [Show full text]
  • Vegen: a Vectorizer Generator for SIMD and Beyond
    VeGen: A Vectorizer Generator for SIMD and Beyond Yishen Chen Charith Mendis Michael Carbin Saman Amarasinghe MIT CSAIL UIUC MIT CSAIL MIT CSAIL USA USA USA USA [email protected] [email protected] [email protected] [email protected] ABSTRACT Vector instructions are ubiquitous in modern processors. Traditional A1 A2 A3 A4 A1 A2 A3 A4 compiler auto-vectorization techniques have focused on targeting single instruction multiple data (SIMD) instructions. However, these B1 B2 B3 B4 B1 B2 B3 B4 auto-vectorization techniques are not sufficiently powerful to model non-SIMD vector instructions, which can accelerate applications A1+B1 A2+B2 A3+B3 A4+B4 A1+B1 A2-B2 A3+B3 A4-B4 in domains such as image processing, digital signal processing, and (a) vaddpd (b) vaddsubpd machine learning. To target non-SIMD instruction, compiler devel- opers have resorted to complicated, ad hoc peephole optimizations, expending significant development time while still coming up short. A1 A2 A3 A4 A1 A2 A3 A4 A5 A6 A7 A8 As vector instruction sets continue to rapidly evolve, compilers can- B1 B2 B3 B4 B5 B6 B7 B8 not keep up with these new hardware capabilities. B1 B2 B3 B4 In this paper, we introduce Lane Level Parallelism (LLP), which A1*B1+ A3*B3+ A5*B5+ A7*B8+ captures the model of parallelism implemented by both SIMD and A1+A2 B1+B2 A3+A4 B3+B4 A2*B2 A4*B4 A6*B6 A7*B8 non-SIMD vector instructions. We present VeGen, a vectorizer gen- (c) vhaddpd (d) vpmaddwd erator that automatically generates a vectorization pass to uncover target-architecture-specific LLP in programs while using only in- Figure 1: Examples of SIMD and non-SIMD instruction in struction semantics as input.
    [Show full text]
  • Exploiting Automatic Vectorization to Employ SPMD on SIMD Registers
    Exploiting automatic vectorization to employ SPMD on SIMD registers Stefan Sprenger Steffen Zeuch Ulf Leser Department of Computer Science Intelligent Analytics for Massive Data Department of Computer Science Humboldt-Universitat¨ zu Berlin German Research Center for Artificial Intelligence Humboldt-Universitat¨ zu Berlin Berlin, Germany Berlin, Germany Berlin, Germany [email protected] [email protected] [email protected] Abstract—Over the last years, vectorized instructions have multi-threading with SIMD instructions1. For these reasons, been successfully applied to accelerate database algorithms. How- vectorization is essential for the performance of database ever, these instructions are typically only available as intrinsics systems on modern CPU architectures. and specialized for a particular hardware architecture or CPU model. As a result, today’s database systems require a manual tai- Although modern compilers, like GCC [2], provide auto loring of database algorithms to the underlying CPU architecture vectorization [1], typically the generated code is not as to fully utilize all vectorization capabilities. In practice, this leads efficient as manually-written intrinsics code. Due to the strict to hard-to-maintain code, which cannot be deployed on arbitrary dependencies of SIMD instructions on the underlying hardware, hardware platforms. In this paper, we utilize ispc as a novel automatically transforming general scalar code into high- compiler that employs the Single Program Multiple Data (SPMD) execution model, which is usually found on GPUs, on the SIMD performing SIMD programs remains a (yet) unsolved challenge. lanes of modern CPUs. ispc enables database developers to exploit To this end, all techniques for auto vectorization have focused vectorization without requiring low-level details or hardware- on enhancing conventional C/C++ programs with SIMD instruc- specific knowledge.
    [Show full text]
  • Using Arm Scalable Vector Extension to Optimize OPEN MPI
    Using Arm Scalable Vector Extension to Optimize OPEN MPI Dong Zhong1,2, Pavel Shamis4, Qinglei Cao1,2, George Bosilca1,2, Shinji Sumimoto3, Kenichi Miura3, and Jack Dongarra1,2 1Innovative Computing Laboratory, The University of Tennessee, US 2fdzhong, [email protected], fbosilca, [email protected] 3Fujitsu Ltd, fsumimoto.shinji, [email protected] 4Arm, [email protected] Abstract— As the scale of high-performance computing (HPC) with extension instruction sets. systems continues to grow, increasing levels of parallelism must SVE is a vector extension for AArch64 execution mode be implored to achieve optimal performance. Recently, the for the A64 instruction set of the Armv8 architecture [4], [5]. processors support wide vector extensions, vectorization becomes much more important to exploit the potential peak performance Unlike other SIMD architectures, SVE does not define the size of target architecture. Novel processor architectures, such as of the vector registers, instead it provides a range of different the Armv8-A architecture, introduce Scalable Vector Extension values which permit vector code to adapt automatically to the (SVE) - an optional separate architectural extension with a new current vector length at runtime with the feature of Vector set of A64 instruction encodings, which enables even greater Length Agnostic (VLA) programming [6], [7]. Vector length parallelisms. In this paper, we analyze the usage and performance of the constrains in the range from a minimum of 128 bits up to a SVE instructions in Arm SVE vector Instruction Set Architec- maximum of 2048 bits in increments of 128 bits. ture (ISA); and utilize those instructions to improve the memcpy SVE not only takes advantage of using long vectors but also and various local reduction operations.
    [Show full text]
  • Introduction on Vectorization
    C E R N o p e n l a b - Intel MIC / Xeon Phi training Intel® Xeon Phi™ Product Family Code Optimization Hans Pabst, April 11th 2013 Software and Services Group Intel Corporation Agenda • Introduction to Vectorization • Ways to write vector code • Automatic loop vectorization • Array notation • Elemental functions • Other optimizations • Summary 2 Copyright© 2012, Intel Corporation. All rights reserved. 4/12/2013 *Other brands and names are the property of their respective owners. Parallelism Parallelism / perf. dimensions Single Instruction Multiple Data • Across mult. applications • Perf. gain simply because • Across mult. processes an instruction performs more works • Across mult. threads • Data parallel • Across mult. instructions • SIMD (“Vector” is usually used as a synonym) 3 Copyright© 2012, Intel Corporation. All rights reserved. 4/12/2013 *Other brands and names are the property of their respective owners. History of SIMD ISA extensions Intel® Pentium® processor (1993) MMX™ (1997) Intel® Streaming SIMD Extensions (Intel® SSE in 1999 to Intel® SSE4.2 in 2008) Intel® Advanced Vector Extensions (Intel® AVX in 2011 and Intel® AVX2 in 2013) Intel Many Integrated Core Architecture (Intel® MIC Architecture in 2013) * Illustrated with the number of 32-bit data elements that are processed by one “packed” instruction. 4 Copyright© 2012, Intel Corporation. All rights reserved. 4/12/2013 *Other brands and names are the property of their respective owners. Vectors (SIMD) float *restrict A; • SSE: 4 elements at a time float *B, *C; addps xmm1, xmm2 • AVX: 8 elements at a time for (i=0; i<n; ++i) { vaddps ymm1, ymm2, ymm3 A[i] = B[i] + C[i]; • MIC: 16 elements at a time } vaddps zmm1, zmm2, zmm3 Scalar code computes the above b1 with one-element at a time.
    [Show full text]
  • MMX and SSE MMX Data Types
    MMX and SSE Extensions to the instruction set for parallel SIMD operations on packed data SIMD – Single Instruction stream Multiple Data stream MMX – Multimedia Extensions SSE – Streaming SIMD Extension SSE2 – Streaming SIMD Extension 2 Designed to speed up multimedia and communication applications graphics and image processing video and audio processing speech compression and recognition 1 MMX data types MMX instructions operate on 8, 16, 32 or 64-bit integer values, packed into a 64-bit field 4 MMX data types 63 0 packed byte b7 b6 b5 b4 b3 b2 b1 b0 8 bytes packed into a 64-bit quantity packed word 4 16-bit words packed into a 63 0 w3 w2 w1 w0 64-bit quantity packed doubleword 63 0 2 32-bit doublewords packed into a dw1 dw0 64-bit quantity quadword 63 0 one 64-bit quantity qw Operates on integer values only 2 1 MMX registers Floating-point registers 8 64-bit MMX registers aliased to the x87 floating-point MM7 registers MM6 no stack-organization MM5 MM4 The 32-bit general-purouse MM3 registers (EAX, EBX, ...) can also MM2 MM1 be used for operands and adresses MM0 MMX registers can not hold memory addresses 63 0 MMX registers have two access modes 64-bit access y 64-bit memory access, transfer between MMX registers, most MMX operations 32-bit access y 32-bit memory access, transfer between MMX and general-purpose registers, some unpack operations 3 MMX operation SIMD execution performs the same operation in parallel on 2, 4 or 8 values MMX instructions perform arithmetic and logical operations in parallel on
    [Show full text]
  • Compiler Auto-Vectorization with Imitation Learning
    Compiler Auto-Vectorization with Imitation Learning Charith Mendis Cambridge Yang Yewen Pu MIT CSAIL MIT CSAIL MIT CSAIL [email protected] [email protected] [email protected] Saman Amarasinghe Michael Carbin MIT CSAIL MIT CSAIL [email protected] [email protected] Abstract Modern microprocessors are equipped with single instruction multiple data (SIMD) or vector instruction sets which allow compilers to exploit fine-grained data level parallelism. To exploit this parallelism, compilers employ auto-vectorization techniques to automatically convert scalar code into vector code. Larsen & Amaras- inghe (2000) first introduced superword level parallelism (SLP) based vectorization, which is a form of vectorization popularly used by compilers. Current compilers employ hand-crafted heuristics and typically only follow one SLP vectorization strategy which can be suboptimal. Recently, Mendis & Amarasinghe (2018) for- mulated the instruction packing problem of SLP vectorization by leveraging an integer linear programming (ILP) solver, achieving superior runtime performance. In this work, we explore whether it is feasible to imitate optimal decisions made by their ILP solution by fitting a graph neural network policy. We show that the learnt policy, Vemal, produces a vectorization scheme that is better than the well-tuned heuristics used by the LLVM compiler. More specifically, the learnt agent produces a vectorization strategy that has a 22.6% higher average reduction in cost compared to the LLVM compiler when measured using its own cost model, and matches the runtime performance of the ILP based solution in 5 out of 7 applications in the NAS benchmark suite. 1 Introduction Modern microprocessors have introduced single instruction multiple data (SIMD) or vector units (e.g.
    [Show full text]
  • Vector Parallelism on Multi-Core Processors
    Vector Parallelism on Multi-Core Processors Steve Lantz, Cornell University CoDaS-HEP Summer School, July 25, 2019 PART I: Vectorization Basics (author: Steve Lantz) Vector Parallelism: Motivation • CPUs are no faster than they were a decade ago – Power limits! “Slow” transistors are more efficient, cooler • Yet process improvements have made CPUs denser – Moore’s Law! Add 2x more “stuff” every 18–24 months • One way to use extra transistors: more cores – Dual-core Intel chips arrived in 2005; counts keep growing – 6–28 in Skylake-SP (doubled in 2-chip Cascade Lake module) • Another solution: SIMD or vector operations – First appeared on Pentium with MMX in 1996 – Vectors have ballooned: 512 bits (16 floats) in Intel Xeon Die shot of 28-core Skylake-SP – Can vectorization increase speed by an order of magnitude? Source: wikichip.org 3 What Moore’s Law Buys Us, These Days… 4 A Third Dimension of Scaling • Along with scaling out and up, you can “scale deep” – Arguably, vectorization can be as important as multithreading • Example: Intel processors in TACC Stampede2 cluster In the cluster: 48 or 68 cores per node 5936 nodes 64 ops/cycle/core – 1736 Skylake-SP + 4200 Xeon Phi (KNL) nodes; 48 or 68 cores each – Each core can do up to 64 operations/cycle on vectors of 16 floats 5 How It Works, Conceptually 1 2 3 4 5 6 7 8 6 8 10 12 SIMD: Single Instruction, Multiple Data 6 Three Ways to Look at Vectorization 1. Hardware Perspective: Run vector instructions involving special registers and functional units that allow in-core parallelism for operations on arrays (vectors) of data.
    [Show full text]
  • VECTORIZATION-Slides
    Vectorization V. Ruggiero ([email protected]) Roma, 19 July 2017 SuperComputing Applications and Innovation Department Outline Topics Introduction Data Dependencies Overcoming limitations to SIMD-Vectorization Vectorization Parallelism Data Thread/Task Process Vectorization Multi-Threading Message Passing Automatic OpenMP MPI Directives/Pragmas TBB, CilkTM Plus Libraries OpenCL pthreads + Cluster + SIMD + Multicore Manycore Topics covered I What are the microprocessor vector extensions or SIMD (Single Instruction Multiple Data Units) I How to use them I Through the compiler via automatic vectorization I Manual transformations that enable vectorization I Directives to guide the compiler I Through intrinsics I Main focus on vectorizing through the compiler I Code more readable I Code portable Outline Topics Introduction Data Dependencies Overcoming limitations to SIMD-Vectorization Vectorization What is Vectorization? I Hardware Perspective: Specialized instructions, registers, or functional units to allow in-core parallelism for operations on arrays (vectors) of data. I Compiler Perspective: Determine how and when it is possible to express computations in terms of vector instructions I User Perspective: Determine how to write code in a manner that allows the compiler to deduce that vectorization is possible. What Happened To Clock Speed? I Everyone loves to misquote Moore’s Law: I "CPU speed doubles every 18 months." I Correct formulation: I "Available on-die transistor density doubles every 18 months." I For a while, this meant easy increases in clock speed I Greater transistor density means more logic space on a chip Clock Speed Wasn’t Everything I Chip designers increased performance by adding sophisticated features to improve code efficiency. I Branch-prediction hardware.
    [Show full text]
  • Impact of Vectorization and Multithreading on Performance and Energy Consumption on Jetson Boards
    Impact of vectorization and multithreading on performance and energy consumption on Jetson boards Sylvain Jubertie, Emmanuel Melin, Naly Raliravaka, Emmanuel Bodèle, Pablo Escot Bocanegra To cite this version: Sylvain Jubertie, Emmanuel Melin, Naly Raliravaka, Emmanuel Bodèle, Pablo Escot Bocanegra. Impact of vectorization and multithreading on performance and energy consumption on Jetson boards. The 2018 International Conference on High Performance Computing & Simulation (HPCS 2018) - HPCS 2018, Jul 2018, Orléans, France. hal-01795146v3 HAL Id: hal-01795146 https://hal-brgm.archives-ouvertes.fr/hal-01795146v3 Submitted on 9 Apr 2020 HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non, lished or not. The documents may come from émanant des établissements d’enseignement et de teaching and research institutions in France or recherche français ou étrangers, des laboratoires abroad, or from public or private research centers. publics ou privés. Impact of vectorization and multithreading on performance and energy consumption on Jetson boards Sylvain Jubertie Emmanuel Melin Naly Raliravaka LIFO EA 4022 LIFO EA 4022 LIFO EA 4022 Universite´ d’Orleans,´ INSA CVL Universite´ d’Orleans,´ INSA CVL Universite´ d’Orleans,´ INSA CVL Orleans,´ France Orleans,´ France Orleans,´ France [email protected] [email protected] [email protected] Emmanuel Bodele` Pablo Escot Bocanegra IUT Departement´ GTE GREMI UMR 7344 Universite´ d’Orleans´ CNRS/Universite´ d’Orleans´ Orleans,´ France Orleans,´ France [email protected] [email protected] Abstract—ARM processors are well known for their energy ef- Furthermore, multicore architectures are also interesting in ficiency and are consequently widely used in embedded platforms.
    [Show full text]
  • A Using Machine Learning to Improve Automatic Vectorization
    A Using Machine Learning to Improve Automatic Vectorization Kevin Stock, The Ohio State University Louis-Noel¨ Pouchet, The Ohio State University P. Sadayappan, The Ohio State University Automatic vectorization is critical to enhancing performance of compute-intensive programs on modern processors. How- ever, there is much room for improvement over the auto-vectorization capabilities of current production compilers, through careful vector-code synthesis that utilizes a variety of loop transformations (e.g. unroll-and-jam, interchange, etc.). As the set of transformations considered is increased, the selection of the most effective combination of transformations becomes a significant challenge: currently used cost-models in vectorizing compilers are often unable to identify the best choices. In this paper, we address this problem using machine learning models to predict the performance of SIMD codes. In contrast to existing approaches that have used high-level features of the program, we develop machine learning models based on features extracted from the generated assembly code, The models are trained off-line on a number of benchmarks, and used at compile-time to discriminate between numerous possible vectorized variants generated from the input code. We demonstrate the effectiveness of the machine learning model by using it to guide automatic vectorization on a variety of tensor contraction kernels, with improvements ranging from 2× to 8× over Intel ICC’s auto-vectorized code. We also evaluate the effectiveness of the model on a number of stencil computations and show good improvement over auto-vectorized code. 1. INTRODUCTION With the increasing degree of SIMD parallelism in processors, the effectiveness of automatic vector- ization in compilers is crucial.
    [Show full text]
  • Selective Vectorization for Short-Vector Instructions Samuel Larsen, Rodric Rabbah, and Saman Amarasinghe
    Computer Science and Artificial Intelligence Laboratory Technical Report MIT-CSAIL-TR-2009-064 December 18, 2009 Selective Vectorization for Short-Vector Instructions Samuel Larsen, Rodric Rabbah, and Saman Amarasinghe massachusetts institute of technology, cambridge, ma 02139 usa — www.csail.mit.edu Selective Vectorization for Short-Vector Instructions Samuel Larsen∗ Rodric Rabbah∗ Saman Amarasinghe VMware IBM Research MIT CSAIL Abstract Multimedia extensions are nearly ubiquitous in today’s general-purpose processors. These extensions consist primarily of a set of short-vector instructions that apply the same opcode to a vector of operands. Vector instructions introduce a data-parallel component to processors that exploit instruction-level par- allelism, and present an opportunity for increased performance. In fact, ignoring a processor’s vector opcodes can leave a significant portion of the available resources unused. In order for software developers to find short-vector instructions generally useful, however, the compiler must target these extensions with complete transparency and consistent performance. This paper describes selective vectorization, a technique for balancing computation across a proces- sor’s scalar and vector units. Current approaches for targeting short-vector instructions directly adopt vectorizing technology first developed for supercomputers. Traditional vectorization, however, can lead to a performance degradation since it fails to account for a processor’s scalar resources. We formulate selective vectorization in the context of software pipelining. Our approach creates software pipelines with shorter initiation intervals, and therefore, higher performance. A key aspect of selective vectorization is its ability to manage transfer of operands between vector and scalar instructions. Even when operand transfer is expensive, our technique is sufficiently sophisticated to achieve significant performance gains.
    [Show full text]