DOCSLIB.ORG

Intermediate Code & Local Optimizations

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Intermediate Code & Local Optimizations Lecture Outline Intermediate Code & • Intermediate code Local Optimizations • Local optimizations 2 Code Generation Summary Why Intermediate Languages? • We have so far discussed ISSUE: When to perform optimizations – Runtime organization – On abstract syntax trees – Simple stack machine code generation • Pro : Machine independent – Improvements to stack machine code generation • Con: Too high level – On assembly language • Our compiler goes directly from the abstract • Pro : Exposes most optimization opportunities syntax tree (AST) to assembly language... • Con: Machine dependent – ... and does not perform optimizations • Con: Must re-implement optimizations when re-targeting – (optimization is the last compiler phase, which is by – On an intermediate language far the largest and most complex these days) • Pro : Exposes optimization opportunities • Most real compilers use intermediate • Pro: Machine independent languages 3 4 Why Intermediate Languages? Kinds of Intermediate Languages High-level intermediate representations: • Have many front-ends into a single back-end – closer to the source language; e.g., syntax trees – gcc can handle C, C++, Java, Fortran, Ada, ... – easy to generate from the input program – each front-end translates source to the same – code optimizations may not be straightforward generic language (called GENERIC) Low-level intermediate representations: – closer to target machine; e.g., P-Code, U-Code (used • Have many back-ends from a single front-end in PA-RISC and MIPS), GCC’s RTL, 3-address code – Do most optimization on intermediate representation – easy to generate code from before emitting code targeted at a single machine – generation from input program may require effort “Mid”-level intermediate representations: – Java bytecode, Microsoft CIL, LLVM IR, ... 5 6 Intermediate Code Languages: Design Issues Intermediate Languages • Designing a good ICode language is not trivial • Each compiler uses its own intermediate • The set of operators in ICode must be rich language enough to allow the implementation of source – IL design is still an active area of research language operations • Nowadays, usually an intermediate language is • ICode operations that are closely tied to a a high-level assembly language particular machine or architecture, make – Uses register names, but has an unlimited number retargeting harder – Uses control structures like assembly language • A small set of operations – Uses opcodes but some are higher level • E.g., push translates to several assembly instructions – may lead to long instruction sequences for some • Most opcodes correspond directly to assembly opcodes source language constructs, – but on the other hand makes retargeting easier 7 8 Architecture of gcc Three-Address Intermediate Code • Each instruction is of the form x := y op z – y and z can only be registers or constants – Just like assembly • Common form of intermediate code • The expression x + y * z gets translated as t1 := y * z t2 := x + t1 – temporary names are made up for internal nodes – each sub-expression has a “home” 9 10 Generating Intermediate Code Generating Intermediate Code (Cont.) • Similar to assembly code generation • igen(e, t) function generates code to compute • Major difference the value of e in register t – Use any number of IL registers to hold • Example: intermediate results igen(e1 + e2, t) = Example: if (x + 2 > 3 * (y - 1) + 42) then z := 0; igen(e1, t 1) (t1 is a fresh register) t1 := x + 2 igen(e2, t2) (t2 is a fresh register) t2 := y - 1 t := t1 + t2 t3 := 3 * t2 • Unlimited number of registers t := t + 42 4 3 ⇒ simple code generation if t1 =< t4 goto L z := 0 L: 11 12 An Intermediate Language From 3-address Code to Machine Code P → S P | ε This is almost a macro expansion process S → id := id op id • id’s are register names 3-address code MIPS assembly code | id := op id x := A[i] load i into r1 | id := id • Constants can replace id’s la r2, A | push id • Typical operators: +, -, * add r2, r2, r1 | id := pop lw r2, (r2) • Typical relops: =, >, >= sw r2, x | if id relop id goto L x := y + z load y into r1 | L: load z into r2 | goto L add r3, r1, r2 sw r3, x if x >= y goto L load x into r1 load y into r2 bge r1, r2, L 13 14 Basic Blocks Basic Block Example • A basic block is a maximal sequence of Consider the basic block instructions with: L: (1) – no labels (except at the first instruction), and t := 2 * x (2) – no jumps (except in the last instruction) w := t + x (3) if w > 0 goto L’ (4) • Idea: – Cannot jump into a basic block (except at beginning) • No way for (3) to be executed without (2) – Cannot jump out of a basic block (except at end) having been executed right before – Each instruction in a basic block is executed after – We can change (3) to w := 3 * x all the preceding instructions have been executed – Can we eliminate (2) as well? 15 16 Identifying Basic Blocks Control-Flow Graphs • Determine the set of leaders, i.e., the first A control-flow graph is a directed graph with instruction of each basic block: – Basic blocks as nodes – The first instruction of a function is a leader – An edge from block A to block B if the execution – Any instruction that is a target of a branch is a can flow from the last instruction in A to the first leader instruction in B – Any instruction immediately following a (conditional E.g., the last instruction in A is goto LB or unconditional) branch is a leader E.g., the execution can fall-through from block A to block B • For each leader, its basic block consists of itself and all instructions up to, but not Frequently abbreviated as CFGs including, the next leader (or end of function) 17 18 Control-Flow Graphs: Example Constructing the Control Flow Graph • Identify the basic blocks of the function x := 1 • The body of a function • There is a directed edge between block B1 to i := 1 (or procedure) can be block B2 if represented as a control- – there is a (conditional or unconditional) jump from L: flow graph the last instruction of B1 to the first instruction of x := x * x B2 or i := i + 1 – B2 immediately follows B1 in the textual order of if i < 42 goto L • There is one initial node the program, and B1 does not end in an unconditional jump. • All “return” nodes are terminal 19 20 Optimization Overview A Classification of Optimizations • Optimization seeks to improve a program’s For languages like C there are three granularities utilization of some resource of optimizations – Execution time (most often) (1) Local optimizations – Code size • Apply to a basic block in isolation – Network messages sent (2) Global optimizations – (Battery) power used, etc. • Apply to a control-flow graph (function body) in isolation (3) Inter-procedural optimizations • Apply across method boundaries • Optimization should not alter what the program computes Most compilers do (1), many do (2) and very few do (3) – The answer must still be the same – Observable behavior must be the same • this typically also includes termination behavior 21 22 Cost of Optimizations Local Optimizations • In practice, a conscious decision is made not • The simplest form of optimizations to implement the fanciest optimization known • No need to analyze the whole procedure body – Just the basic block in question • Why? – Some optimizations are hard to implement • Example: algebraic simplification – Some optimizations are costly in terms of compilation time – Some optimizations have low benefit – Many fancy optimizations are all three above! • Goal: maximum benefit for minimum cost 23 24 Algebraic Simplification Constant Folding • Some statements can be deleted • Operations on constants can be computed at x := x + 0 compile time x := x * 1 • In general, if there is a statement • Some statements can be simplified x := y op z x := x * 0 ⇒ x := 0 – And y and z are constants y := y ** 2 ⇒ y := y * y – Then yopz can be computed at compile time x := x * 8 ⇒ x := x << 3 x := x * 15 ⇒ t := x << 4; x := t - x • Example: x := 2 + 2 ⇒ x := 4 (on some machines << is faster than *; but not on all!) • Example: if 2 < 0 goto L can be deleted • When might constant folding be dangerous? 25 26 Flow of Control Optimizations Single Assignment Form • Eliminating unreachable code: • Some optimizations are simplified if each – Code that is unreachable in the control-flow graph register occurs only once on the left-hand – Basic blocks that are not the target of any jump or side of an assignment “fall through” from a conditional • Intermediate code can be rewritten to be in – Such basic blocks can be eliminated single assignment form x := z + y b := z + y • Why would such basic blocks occur? a := x ⇒ a := b x := 2 * x x := 2 * b • Removing unreachable code makes the (b is a fresh temporary) program smaller • More complicated in general, due to control – And sometimes also faster flow (e.g. loops) • Due to memory cache effects (increased spatial locality) 27 28 Common Subexpression Elimination Copy Propagation • Assume • If w := x appears in a block, all subsequent – A basic block is in single assignment form uses of w can be replaced with uses of x – A definition x := is the first use of x in a block • All assignments with same RHS compute the • Example: same value b := z + y b := z + y a := b ⇒ a := b • Example: x := 2 * a x := 2 * b x := y + z x := y + z … ⇒ … • This does not make the program smaller or w := y + z w := x faster but might enable other optimizations (the values of x, y, and z do not change in the … code) – Constant folding – Dead code elimination 29 30 Copy Propagation and Constant Folding
Load more

Recommended publications
  • Polyhedral Compilation As a Design Pattern for Compiler Construction
    Polyhedral Compilation as a Design Pattern for Compiler Construction PLISS, May 19-24, 2019 [email protected] Polyhedra? Example: Tiles Polyhedra? Example: Tiles How many of you read “Design Pattern”? → Tiles Everywhere 1. Hardware Example: Google Cloud TPU Architectural Scalability With Tiling Tiles Everywhere 1. Hardware Google Edge TPU Edge computing zoo Tiles Everywhere 1. Hardware 2. Data Layout Example: XLA compiler, Tiled data layout Repeated/Hierarchical Tiling e.g., BF16 (bfloat16) on Cloud TPU (should be 8x128 then 2x1) Tiles Everywhere Tiling in Halide 1. Hardware 2. Data Layout Tiled schedule: strip-mine (a.k.a. split) 3. Control Flow permute (a.k.a. reorder) 4. Data Flow 5. Data Parallelism Vectorized schedule: strip-mine Example: Halide for image processing pipelines vectorize inner loop https://halide-lang.org Meta-programming API and domain-specific language (DSL) for loop transformations, numerical computing kernels Non-divisible bounds/extent: strip-mine shift left/up redundant computation (also forward substitute/inline operand) Tiles Everywhere TVM example: scan cell (RNN) m = tvm.var("m") n = tvm.var("n") 1. Hardware X = tvm.placeholder((m,n), name="X") s_state = tvm.placeholder((m,n)) 2. Data Layout s_init = tvm.compute((1,n), lambda _,i: X[0,i]) s_do = tvm.compute((m,n), lambda t,i: s_state[t-1,i] + X[t,i]) 3. Control Flow s_scan = tvm.scan(s_init, s_do, s_state, inputs=[X]) s = tvm.create_schedule(s_scan.op) 4. Data Flow // Schedule to run the scan cell on a CUDA device block_x = tvm.thread_axis("blockIdx.x") 5. Data Parallelism thread_x = tvm.thread_axis("threadIdx.x") xo,xi = s[s_init].split(s_init.op.axis[1], factor=num_thread) s[s_init].bind(xo, block_x) Example: Halide for image processing pipelines s[s_init].bind(xi, thread_x) xo,xi = s[s_do].split(s_do.op.axis[1], factor=num_thread) https://halide-lang.org s[s_do].bind(xo, block_x) s[s_do].bind(xi, thread_x) print(tvm.lower(s, [X, s_scan], simple_mode=True)) And also TVM for neural networks https://tvm.ai Tiling and Beyond 1.
    [Show full text]
  • CS153: Compilers Lecture 19: Optimization
    CS153: Compilers Lecture 19: Optimization Stephen Chong https://www.seas.harvard.edu/courses/cs153 Contains content from lecture notes by Steve Zdancewic and Greg Morrisett Announcements •HW5: Oat v.2 out •Due in 2 weeks •HW6 will be released next week •Implementing optimizations! (and more) Stephen Chong, Harvard University 2 Today •Optimizations •Safety •Constant folding •Algebraic simplification • Strength reduction •Constant propagation •Copy propagation •Dead code elimination •Inlining and specialization • Recursive function inlining •Tail call elimination •Common subexpression elimination Stephen Chong, Harvard University 3 Optimizations •The code generated by our OAT compiler so far is pretty inefficient. •Lots of redundant moves. •Lots of unnecessary arithmetic instructions. •Consider this OAT program: int foo(int w) { var x = 3 + 5; var y = x * w; var z = y - 0; return z * 4; } Stephen Chong, Harvard University 4 Unoptimized vs. Optimized Output .globl _foo _foo: •Hand optimized code: pushl %ebp movl %esp, %ebp _foo: subl $64, %esp shlq $5, %rdi __fresh2: movq %rdi, %rax leal -64(%ebp), %eax ret movl %eax, -48(%ebp) movl 8(%ebp), %eax •Function foo may be movl %eax, %ecx movl -48(%ebp), %eax inlined by the compiler, movl %ecx, (%eax) movl $3, %eax so it can be implemented movl %eax, -44(%ebp) movl $5, %eax by just one instruction! movl %eax, %ecx addl %ecx, -44(%ebp) leal -60(%ebp), %eax movl %eax, -40(%ebp) movl -44(%ebp), %eax Stephen Chong,movl Harvard %eax,University %ecx 5 Why do we need optimizations? •To help programmers… •They write modular, clean, high-level programs •Compiler generates efficient, high-performance assembly •Programmers don’t write optimal code •High-level languages make avoiding redundant computation inconvenient or impossible •e.g.
    [Show full text]
  • Cross-Platform Language Design
    Cross-Platform Language Design THIS IS A TEMPORARY TITLE PAGE It will be replaced for the final print by a version provided by the service academique. Thèse n. 1234 2011 présentée le 11 Juin 2018 à la Faculté Informatique et Communications Laboratoire de Méthodes de Programmation 1 programme doctoral en Informatique et Communications École Polytechnique Fédérale de Lausanne pour l’obtention du grade de Docteur ès Sciences par Sébastien Doeraene acceptée sur proposition du jury: Prof James Larus, président du jury Prof Martin Odersky, directeur de thèse Prof Edouard Bugnion, rapporteur Dr Andreas Rossberg, rapporteur Prof Peter Van Roy, rapporteur Lausanne, EPFL, 2018 It is better to repent a sin than regret the loss of a pleasure. — Oscar Wilde Acknowledgments Although there is only one name written in a large font on the front page, there are many people without which this thesis would never have happened, or would not have been quite the same. Five years is a long time, during which I had the privilege to work, discuss, sing, learn and have fun with many people. I am afraid to make a list, for I am sure I will forget some. Nevertheless, I will try my best. First, I would like to thank my advisor, Martin Odersky, for giving me the opportunity to fulfill a dream, that of being part of the design and development team of my favorite programming language. Many thanks for letting me explore the design of Scala.js in my own way, while at the same time always being there when I needed him.
    [Show full text]
  • Phase-Ordering in Optimizing Compilers
    Phase-ordering in optimizing compilers Matthieu Qu´eva Kongens Lyngby 2007 IMM-MSC-2007-71 Technical University of Denmark Informatics and Mathematical Modelling Building 321, DK-2800 Kongens Lyngby, Denmark Phone +45 45253351, Fax +45 45882673 [email protected] www.imm.dtu.dk Summary The “quality” of code generated by compilers largely depends on the analyses and optimizations applied to the code during the compilation process. While modern compilers could choose from a plethora of optimizations and analyses, in current compilers the order of these pairs of analyses/transformations is fixed once and for all by the compiler developer. Of course there exist some flags that allow a marginal control of what is executed and how, but the most important source of information regarding what analyses/optimizations to run is ignored- the source code. Indeed, some optimizations might be better to be applied on some source code, while others would be preferable on another. A new compilation model is developed in this thesis by implementing a Phase Manager. This Phase Manager has a set of analyses/transformations available, and can rank the different possible optimizations according to the current state of the intermediate representation. Based on this ranking, the Phase Manager can decide which phase should be run next. Such a Phase Manager has been implemented for a compiler for a simple imper- ative language, the while language, which includes several Data-Flow analyses. The new approach consists in calculating coefficients, called metrics, after each optimization phase. These metrics are used to evaluate where the transforma- tions will be applicable, and are used by the Phase Manager to rank the phases.
    [Show full text]
  • Dataflow Analysis: Constant Propagation
    Dataflow Analysis Monday, November 9, 15 Program optimizations • So far we have talked about different kinds of optimizations • Peephole optimizations • Local common sub-expression elimination • Loop optimizations • What about global optimizations • Optimizations across multiple basic blocks (usually a whole procedure) • Not just a single loop Monday, November 9, 15 Useful optimizations • Common subexpression elimination (global) • Need to know which expressions are available at a point • Dead code elimination • Need to know if the effects of a piece of code are never needed, or if code cannot be reached • Constant folding • Need to know if variable has a constant value • So how do we get this information? Monday, November 9, 15 Dataflow analysis • Framework for doing compiler analyses to drive optimization • Works across basic blocks • Examples • Constant propagation: determine which variables are constant • Liveness analysis: determine which variables are live • Available expressions: determine which expressions are have valid computed values • Reaching definitions: determine which definitions could “reach” a use Monday, November 9, 15 Example: constant propagation • Goal: determine when variables take on constant values • Why? Can enable many optimizations • Constant folding x = 1; y = x + 2; if (x > z) then y = 5 ... y ... • Create dead code x = 1; y = x + 2; if (y > x) then y = 5 ... y ... Monday, November 9, 15 Example: constant propagation • Goal: determine when variables take on constant values • Why? Can enable many optimizations • Constant folding x = 1; x = 1; y = x + 2; y = 3; if (x > z) then y = 5 if (x > z) then y = 5 ... y ... ... y ... • Create dead code x = 1; y = x + 2; if (y > x) then y = 5 ..
    [Show full text]
  • Compiler-Based Code-Improvement Techniques
    Compiler-Based Code-Improvement Techniques KEITH D. COOPER, KATHRYN S. MCKINLEY, and LINDA TORCZON Since the earliest days of compilation, code quality has been recognized as an important problem [18]. A rich literature has developed around the issue of improving code quality. This paper surveys one part of that literature: code transformations intended to improve the running time of programs on uniprocessor machines. This paper emphasizes transformations intended to improve code quality rather than analysis methods. We describe analytical techniques and specific data-flow problems to the extent that they are necessary to understand the transformations. Other papers provide excellent summaries of the various sub-fields of program analysis. The paper is structured around a simple taxonomy that classifies transformations based on how they change the code. The taxonomy is populated with example transformations drawn from the literature. Each transformation is described at a depth that facilitates broad understanding; detailed references are provided for deeper study of individual transformations. The taxonomy provides the reader with a framework for thinking about code-improving transformations. It also serves as an organizing principle for the paper. Copyright 1998, all rights reserved. You may copy this article for your personal use in Comp 512. Further reproduction or distribution requires written permission from the authors. 1INTRODUCTION This paper presents an overview of compiler-based methods for improving the run-time behavior of programs — often mislabeled code optimization. These techniques have a long history in the literature. For example, Backus makes it quite clear that code quality was a major concern to the implementors of the first Fortran compilers [18].
    [Show full text]
  • 3. Constant Propagation 4. Copy Propagation 5. Constant Folding
    3. Constant Propagation If a variable is known to contain a particular constant value at a particular point in the program, replace references to the variable at that point with that constant value. 4. Copy Propagation After the assignment of one variable to another, a reference to one variable may be replaced with the value of the other variable (until one or the other of the variables is reassigned). (This may also “set up” dead code elimination. Why?) 5. Constant Folding An expression involving constant (literal) values may be evaluated and simplified to a constant result value. Particularly useful when constant propagation is performed. © CS 701 Fall 2007 20 6. Dead Code Elimination Expressions or statements whose values or effects are unused may be eliminated. 7. Loop Invariant Code Motion An expression that is invariant in a loop may be moved to the loop’s header, evaluated once, and reused within the loop. Safety and profitability issues may be involved. 8. Scalarization (Scalar Replacement) A field of a structure or an element of an array that is repeatedly read or written may be copied to a local variable, accessed using the local, and later (if necessary) copied back. This optimization allows the local variable (and in effect the field or array component) to be allocated to a register. © CS 701 Fall 2007 21 9. Local Register Allocation Within a basic block (a straight line sequence of code) track register contents and reuse variables and constants from registers. 10. Global Register Allocation Within a subprogram, frequently accessed variables and constants are allocated to registers.
    [Show full text]
  • Vbcc Compiler System
    vbcc compiler system Volker Barthelmann i Table of Contents 1 General :::::::::::::::::::::::::::::::::::::::::: 1 1.1 Introduction ::::::::::::::::::::::::::::::::::::::::::::::::::: 1 1.2 Legal :::::::::::::::::::::::::::::::::::::::::::::::::::::::::: 1 1.3 Installation :::::::::::::::::::::::::::::::::::::::::::::::::::: 2 1.3.1 Installing for Unix::::::::::::::::::::::::::::::::::::::::: 3 1.3.2 Installing for DOS/Windows::::::::::::::::::::::::::::::: 3 1.3.3 Installing for AmigaOS :::::::::::::::::::::::::::::::::::: 3 1.4 Tutorial :::::::::::::::::::::::::::::::::::::::::::::::::::::::: 5 2 The Frontend ::::::::::::::::::::::::::::::::::: 7 2.1 Usage :::::::::::::::::::::::::::::::::::::::::::::::::::::::::: 7 2.2 Configuration :::::::::::::::::::::::::::::::::::::::::::::::::: 8 3 The Compiler :::::::::::::::::::::::::::::::::: 11 3.1 General Compiler Options::::::::::::::::::::::::::::::::::::: 11 3.2 Errors and Warnings :::::::::::::::::::::::::::::::::::::::::: 15 3.3 Data Types ::::::::::::::::::::::::::::::::::::::::::::::::::: 15 3.4 Optimizations::::::::::::::::::::::::::::::::::::::::::::::::: 16 3.4.1 Register Allocation ::::::::::::::::::::::::::::::::::::::: 18 3.4.2 Flow Optimizations :::::::::::::::::::::::::::::::::::::: 18 3.4.3 Common Subexpression Elimination :::::::::::::::::::::: 19 3.4.4 Copy Propagation :::::::::::::::::::::::::::::::::::::::: 20 3.4.5 Constant Propagation :::::::::::::::::::::::::::::::::::: 20 3.4.6 Dead Code Elimination::::::::::::::::::::::::::::::::::: 21 3.4.7 Loop-Invariant Code Motion
    [Show full text]
  • Optimization
    Introduction to Compilers and Language Design Copyright © 2020 Douglas Thain. Paperback ISBN: 979-8-655-18026-0 Second edition. Anyone is free to download and print the PDF edition of this book for per- sonal use. Commercial distribution, printing, or reproduction without the author’s consent is expressly prohibited. All other rights are reserved. You can find the latest version of the PDF edition, and purchase inexpen- sive hardcover copies at http://compilerbook.org Revision Date: January 15, 2021 195 Chapter 12 – Optimization 12.1 Overview Using the basic code generation strategy shown in the previous chapter, you can build a compiler that will produce perfectly usable, working code. However, if you examine the output of your compiler, there are many ob- vious inefficiencies. This stems from the fact that the basic code generation strategy considers each program element in isolation and so must use the most conservative strategy to connect them together. In the early days of high level languages, before optimization strategies were widespread, code produced by compilers was widely seen as inferior to that which was hand-written by humans. Today, a modern compiler has many optimization techniques and very detailed knowledge of the underlying architecture, and so compiled code is usually (but not always) superior to that written by humans. Optimizations can be applied at multiple stages of the compiler. It’s usually best to solve a problem at the highest level of abstraction possible. For example, once we have generated concrete assembly code, about the most we can do is eliminate a few redundant instructions.
    [Show full text]
  • CS 4120 Introduction to Compilers Andrew Myers Cornell University
    CS 4120 Introduction to Compilers Andrew Myers Cornell University Lecture 19: Introduction to Optimization 2020 March 9 Optimization • Next topic: how to generate better code through optimization. • Tis course covers the most valuable and straightforward optimizations – much more to learn! –Other sources: • Appel, Dragon book • Muchnick has 10 chapters of optimization techniques • Cooper and Torczon 2 CS 4120 Introduction to Compilers How fast can you go? 10000 direct source code interpreters (parsing included!) 1000 tokenized program interpreters (BASIC, Tcl) AST interpreters (Perl 4) 100 bytecode interpreters (Java, Perl 5, OCaml, Python) call-threaded interpreters 10 pointer-threaded interpreters (FORTH) simple code generation (PA4, JIT) 1 register allocation naive assembly code local optimization global optimization FPGA expert assembly code GPU 0.1 3 CS 4120 Introduction to Compilers Goal of optimization • Compile clean, modular, high-level source code to expert assembly-code performance. • Can’t change meaning of program to behavior not allowed by source. • Different goals: –space optimization: reduce memory use –time optimization: reduce execution time –power optimization: reduce power usage 4 CS 4120 Introduction to Compilers Why do we need optimization? • Programmers may write suboptimal code for clarity. • Source language may make it hard to avoid redundant computation a[i][j] = a[i][j] + 1 • Architectural independence • Modern architectures assume optimization—hard to optimize by hand! 5 CS 4120 Introduction to Compilers Where to optimize? • Usual goal: improve time performance • But: many optimizations trade off space vs. time. • Example: loop unrolling replaces a loop body with N copies. –Increasing code space speeds up one loop but slows rest of program down a little.
    [Show full text]
  • Launch-Time Optimization of Opencl Kernels by Andrew Siu Doug Lee A
    Launch-time Optimization of OpenCL Kernels by Andrew Siu Doug Lee A thesis submitted in conformity with the requirements for the degree of Master of Applied Science Graduate Department of Electrical and Computer Engineering University of Toronto c Copyright 2017 by Andrew Siu Doug Lee Abstract Launch-time Optimization of OpenCL Kernels Andrew Siu Doug Lee Master of Applied Science Graduate Department of Electrical and Computer Engineering University of Toronto 2017 OpenCL kernels are compiled first before kernel arguments and launch geometry are provided later at launch time. Although some of these values remain constant during execution, the compiler is unable to optimize for them since it has no access to them. We propose and implement a novel approach that identifies such arguments, geom- etry, and optimizations at compile time using a series of annotations. At launch time the annotations, combined with the actual values of the arguments and geometry, are processed and used to optimize the code, thereby improving kernel performance. We measure the execution times of 16 benchmarks our approach. The results show that annotation processing is fast and that kernel performance is improved by a factor of up to 2.06X and on average by 1.22X. When considering the entire compilation flow, improvement can be up to 1.13X, but can also potentiality degrade to 0.44X, depending on the launch scenario. ii Acknowledgements My most humble thanks go to my advisor, Tarek S. Abdelrahman, for his invaluable expertise, candid advice, and generous patience. The document you are now reading, the research contained within, and the many related presentations are all the products of his guidance.
    [Show full text]
  • Iterative Dataflow Analysis
    Iterative Dataflow Analysis ¯ Many dataflow analyses have the same structure. ¯ They “interpret” the statements in program, collecting information as they proceed. ¯ Ideally, we would like to do “perfect interpretation,” collecting exact information about how the program executes. ¯ We can’t because we don’t have the program inputs at compile time, and we aren’t going to make our compiler run the program to completion during compilation–that just takes too long (why the hell do you think we are compiling the program in the first place?) Computer Science 320 Prof. David Walker -1- Iterative Dataflow Analysis ¯ Instead of interpreting the program exactly, we approximate the program behavior as we execute. ¯ This is called Abstract Interpretation and it is often a useful way of understanding different program analyses. ¯ Analysis via abstract interpretation is often defined by: – A transfer function — ´Òµ — that simulates/approximates execution of instruc- tion Ò on its inputs – A joining operator — — since you don’t know which branch of an if state- ment (for example) will be taken at compile time, you need to interpret both and combine the results. – A direction — FORWARD or REVERSE — In the reverse direction, we are interpreting the program backwards. Computer Science 320 Prof. David Walker -2- Iterative Dataflow Analysis To code up a particular analysis we need to take the following steps. ¯ First, we decide what sort of information we are interested in processing. This is going to determine the transfer function and the joining operator, as well as any initial conditions that need to be set up. ¯ Second, we decide on the appropriate direction for the analysis.
    [Show full text]