Composition and Reuse with Compiled Domain-Specific Languages

Pervasive Parallelism (PPL) Lab (http://ppl.stanford.edu) and LAMP (http://lamp.epfl.ch) Arvind Sujeeth, Kevin Brown, HyoukJoong Lee, Hassan Chafi, Michael Wu, Victoria Popic, Kunle Olukotun

Tiark Rompf, Aleksander Prokopec, Vojin Jovanovic, Martin Odersky

Compiled Embedded Parallel DSLs Domain-Specific Languages for Heterogeneous Parallelism

Contact information Open Source Repository  http://stanford-­‐ppl.github.com/Delite/   Group Website http://ppl.stanford.edu  

E-mail Addresses (Students) asujeeth@stanford.edu,  

E-mail Address (Faculty Advisor) kunle@stanford.edu  

DSL Re-Use

Delite and Lightweight Modular Staging

OptiLA: Linear Algebra • Dense/Sparse Vector/Matrix • Can be re-used & extended by other DSLs

OptiML: Machine Learning • Extends OptiLA • Adds ML-specific constructs: e.g., TrainingSet, untilconverged

OptiCollections: Parallel Collections • Arrays, Maps - Parallel collection interface using Delite ops • Designed to be used by other DSLs

OptiQL: Data Querying • In-memory collection-based querying • Utilizes OptiCollections’ HashMap

OptiMesh: Mesh-based PDEs • Liszt on Delite • More advanced DSL analyses and optimizations

OptiGraph: Small-world Graphs • Green-Marl on Delite • Shares analyses / transformations with OptiMesh

DSL Interoperability

Evaluation

OptiQL  {      //  customers:  Array[Customer]      //  orders:  Array[Order]      val  orders  =  customers  Join(orders)            WhereEq(_.c_custkey,  _.o_custkey)          Select((c,o)  =>  new  Result  {                val  nationKey  =  c.c_nationkey                val  acctBalance  =  c.c_acctbal              val  price  =  o.o_totalprice          })      orderData.set(orders)  }    OptiML  {      //  run  linear  regression  on  price      val  data  =  Matrix(orderData.get)      val  x  =  data.sliceCols(0,1)      val  y  =  data.getCol(2)      theta.set(linreg.weighted(x,y).toArray)  }    println("theta:  "  +  theta.get)    

DSL Delite Ops Generic Optimizations

OptiML Map, ZipWith, Reduce, Foreach, Hash, Sort

CSE, DCE, code motion, fusion

OptiQL Map, Reduce, Filter, Sort, Hash, Join

CSE, DCE, code motion, fusion

OptiGraph ForeachReduce, Map, Reduce, Filter

CSE, DCE, code motion

OptiMesh ForeachReduce CSE, DCE, code motion

OptiCollections Map, FlatMap, Reduce, Filter, Sort, Hash, ZipWith

CSE, DCE, code motion, fusion

Delite: Reusable infrastructure for rapid development of compiled embedded DSLs

We modified the scala-virtualized compiler to introduce scopes, coarse-grained execution blocks that are compiled, staged and executed independently Scopes can communicate via global heap objects

Lexer% Parser% Type%checker%

Embedded%front7end%

External%front7end%Full%embedding%(composable)%Front7end%embedding%(not%composable)%

Spectrum)of)compiled)domain1specific)languages)

Stand7alone%(not%composable)% Back7end%embedding%(composable)%

Analysis% OpCmizaCon% Code%gen%

Embedded%back7end%

External%back7end%

Why Compiled Domain-Specific Languages? How do we compose them?

Delite&Execu+on&Graph&

Scala&CUDA&OpenCL&

gen&ops&parallel&

domain&

mul$view)IR) traversals) generated)code)

."."."

Code

&gen

erators&

data&structures&

Gen&analyses&

Gen&transformers&

Generic&op+miza+on&

DS&analyses&

DS&transformers&

DS&op+miza+on&

Example DSL: Vectors val  (v1,v2)  =  (Vector.rand(1000),  Vector.rand(1000))  val  a  =  v1+v2  println(a)  

Lightweight Modular Staging: Embedded DSL compilation with metaprogramming •  All application types are wrapped in an abstract type constructor, Rep[T]

•  The DSL returns type-safe placeholders instead of immediate results, allowing it to build an intermediate representation, optimize it, and generate code

•  The DSL is embedded in Scala, a general purpose programming language

•  The syntax is legal Scala

•  When the program is executed, it builds an intermediate representation instead of computing a result (DSL compilation)

•  The DSL implementation uses Delite ops to internally represent operations as parallel patterns: •  case  class  VectorPlus  extends  DeliteOpZipWith  {          def  func  =  (a,b)  =>  a  +  b        }  

•  Delite generates code from the DSL IR to multiple platforms (C++, Scala, CUDA, OpenCL)

Vector.rand Vector.rand

VectorPlus

Println

Scala code

C++ code

CUDA code

OpenCL code

•  We want programs that satisfy the 3 P’s: •  Performance, Productivity, Portability

•  Very hard to achieve with a general-purpose language: compilers lack semantic information necessary to efficiently translate to parallel hardware

•  Domain-specific languages have been shown to be capable of providing the 3 P’s in some domains, but: •  Embedded DSLs are easy to build and compose, but

don’t perform well •  Stand-alone DSLs are hard to build and compose, but

perform well •  We propose using compiled DSLs embedded in a

common back-end compiler

Two main ways: •  Open World: fine-grained cooperative composition

•  Embedded DSLs that are designed to be composed •  No whole program analyses or optimizations that rely on

domain assumptions or restricted semantics •  DSL composition reduces to object composition in the host lang  trait  Vectors  extends  MathOps  with  ScalarOps  with  VectorOps  trait  Matrices  extends  Vectors  with  MatrixOps  

•  Closed World: coarse-grained isolated composition •  Embedded DSLs that require isolation to guarantee safety of

optimizations •  3 steps: 1) independently parse DSL blocks and apply domain-

specific optimizations, 2) lower each DSL block to a common, high-level IR, 3) optimize and code generate resulting IR

Mix-in composition

OptiQL LINQ OptiQL LINQ OptiQL LINQ OptiQL LINQ0.91 52.73 0.188 19.29 182.7 663.76 31.54 154.56

55.65 22.83 182.03 668.7 31.82 140.0255.7 28.17 182.46 673.11 31.71 137.96

184.21 31.82183.36 31.68

0.91 54.69333 0.188 23.43 182.952 668.5233 31.714 144.18 (raw)0.016638 1 0.003437 0.428389 0.273666 1 0.047439 0.215669 (time)60.10256 1 290.922 2.334329 3.654091 1 21.07975 4.636727 (speedup)

total speedupgroup/soa 0.91 4.461538fusion 4.06 2.118227staging 8.6 2.161628array 18.59 2.065627lift date 38.4 1.424219iterable 54.69

OptiQLQ1 Q2

1 thread 8 threads 1 thread 8 threads

1.0

2.3

60.1 291 0

0.2

0.4

0.6

0.8

1

1.2

1P 8P

No

rmal

ize

d E

xecu

tio

n T

ime

5GB 1.0

4.6 3.7

21.1 0

0.2

0.4

0.6

0.8

1

1.2

1P 8P

LINQ

OptiQL

5GB

1 2 4 8 1 2 4 82.1414 1.6862 1.0656 0.6824 28.95 13.67 7.45 4.93

28.04 12.47 7.46 527.57 13.4 7.47 4.9927.62 12.71 7.36 5.0328.03 12.66 7.53 5.31

2.1414 1.6862 1.0656 0.6824 28.042 12.982 7.454 5.0520.297541 0.234292 0.148062 0.094817 0.614857 0.284647 0.163439 0.1107723.360885 4.268177 6.753941 10.5466 1.626394 3.513121 6.118505 9.02758

463 MB75 MBOptiCollections

1.0

1.9

3.4

5.5

1.6

3.5 6.1

9.0

0

0.2

0.4

0.6

0.8

1

1.2

1P 2P 4P 8P

Scala ParallelCollectionsOptiCollections

463 MB 1.0

1.7

2.4 2.8

3.4 4.3

6.8 10.5

0

0.2

0.4

0.6

0.8

1

1.2

1P 2P 4P 8P

No

rma

lize

d E

xecu

tio

n T

ime

75 MB

C++1 2 4 8 1 2 4 80.05 0.03 0.03 0.030.05 0.03 0.03 0.030.05 0.03 0.03 0.030.05 0.03 0.03 0.030.05 0.03 0.03 0.03

0.05 0.03 0.03 0.03 0.09 0.05 0.03 0.03 0.03 (raw)1 0.6 0.6 0.6 1.8 1 0.6 0.6 0.6 (time)1 1.666667 1.666667 1.666667 0.555556 1 1.666667 1.666667 1.666667 (speedup)

0.555556 0.333333 0.333333 0.333333 1 0.555556 0.333333 0.333333 0.333333 (time rel. c++)1.8 3 3 3 1 1.8 3 3 3 (speedup rel. c++)

100k x 800kGreen MarlOptiGraph

1.7 1.7 1.7

0

0.2

0.4

0.6

0.8

1

1P 2P 4P 8P

No

rma

lize

d E

xe

cuti

on

Tim

e

100k nodes x 800k edges

1.0

1.8

2.9 3.3

.75

1.5

2.9 3.6

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1P 2P 4P 8P

Green Marl

OptiGraph

8M nodes x 64M edges

GPU Liszt GPU Liszt CPU GPU Liszt GPU Liszt CPU2.6 2.199043 73.9821 1.641 1.39132 5.03738 (raw)

0.035144 0.029724 1 0.325765 0.276199 1 (time) 0.035144 0.02972428.45465 33.64287 1 3.069701 3.620576 1 (speedup) 28.45465 33.64287

OptiMeshShallow Water Scalar Convection

1.0

33.6 28.5 0

0.2

0.4

0.6

0.8

1

1.2

No

rma

lize

d E

xe

cuti

on

Tim

e

28.5

33.6

0

0.01

0.02

0.03

0.04

Shallow Water

No

rma

lize

d E

xe

cuti

on

Tim

e

OptiMesh GPU

3.1

3.6

0

0.1

0.2

0.3

0.4

Scalar Convection

Liszt GPU

1P GPU GPU

1.0

3.6 3.1

0

0.2

0.4

0.6

0.8

1

1.2

Liszt CPU

Liszt GPU

OptiMesh

1P GPU GPU

1.0

3.9

8.0 44.9

8.1 54.4

0

0.2

0.4

0.6

0.8

1

1.2

1P 8P

No

rmal

ize

d E

xecu

tio

n T

ime

Scala Library

OptiDSLs No Cross Opt

OptiDSLs

1.0

6.4

17.3 51.1 31.4 79.9 0

0.2

0.4

0.6

0.8

1

1.2

1P 8P

No

rma

lize

d E

xecu

tio

n T

ime

Scala Library

OptiDSLs No Cross Opt

OptiDSLs

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

OptiDSLs

OptiDSLsNo Cross Opt

Scala Library

OptiDSLs

OptiDSLsNo Cross Opt

Scala Library

Normalized Execution Time

OptiQL

OptiGraph

OptiML

4.2

7.2

1.6

14.1

19.7

1 P

8 P

OptiQL vs. LINQ OptiCollections vs. Scala

OptiGraph vs. Green Marl OptiMesh vs. Liszt

Open World Composition

Small dataset Large dataset

TPCH Q1 TPCH Q2

Shallow Water Simulation Scalar Convection

Reverse web-link benchmark

Page Rank

Closed World Composition

Twitter Data Analysis