TDDD56 Lab 3: Skeleton programming with SkePU August Ernstsson - - PowerPoint PPT Presentation

▶
tddd56
SMART_READER_LITE
LIVE PREVIEW

TDDD56 Lab 3: Skeleton programming with SkePU August Ernstsson - - PowerPoint PPT Presentation

Nov 18, 2023 440 likes •661 views

TDDD56 Lab 3: Skeleton programming with SkePU August Ernstsson august.ernstsson@liu.se Labs schedule WebReg Week Lab 1 v46 Load Balancing Responsible: August Lab 2 v47 CPU Non-blocking Data Structures Lesson 2 Lab 3 v48 High-level

slide-1
SLIDE 1

TDDD56

Lab 3: Skeleton programming with SkePU

August Ernstsson

august.ernstsson@liu.se

slide-2
SLIDE 2

TDDD56 Lab 3 August Ernstsson 2020

Labs schedule

CPU Lab 1 v46 Load Balancing Lab 2 v47 Non-blocking Data Structures Lab 3 v48 High-level Parallel Programming GPU Lab 4 v49 CUDA 1 Lab 5 v50 CUDA 2 Lab 6 v51 OpenCL

Responsible: August Responsible: Ingemar Lesson 2 Week WebReg

slide-3
SLIDE 3

TDDD56 Lab 3 August Ernstsson 2020

C++11

  • Shift in the labs from C to C++11 (”modern” C++)

// ”auto” type specifier auto addOneMap = skepu::Map<1>(addOneFunc); skepu::Vector<float> input(size), res(size); input.randomize(0, 9); // Lambda expression auto dur = skepu::benchmark::measureExecTime([&] { addOneMap(res, input); }); capture by reference

slide-4
SLIDE 4

TDDD56 Lab 3 August Ernstsson 2020

SkePU

  • Skeleton programming framework
  • C++11 library with skeleton and data container classes
  • A source-to-source translator tool
  • Smart containers: Vector<T>, Matrix<T>, tensors
  • For heterogeneous multicore systems and clusters
  • Multiple backends
  • Active research tool (A good topic for your thesis?)
slide-5
SLIDE 5

TDDD56 Lab 3 August Ernstsson 2020

SkePU architecture

C++ interface (skeletons, smart containers, …) C++ OpenMP OpenCL CUDA StarPU + MPI CPU Multi-core CPU Accelerator GPU Cluster

slide-6
SLIDE 6

TDDD56 Lab 3 August Ernstsson 2020

SkePU skeletons

  • Parametrizable higher-order functions implemented as C++ template classes
  • Map
  • Reduce
  • MapReduce
  • MapOverlap
  • MapPairs
  • MapPairsReduce
  • Scan

Map

Mult

slide-7
SLIDE 7

TDDD56 Lab 3 August Ernstsson 2020

SkePU skeletons

1 2 3 4 5 6 7 8 2 2 2 2 2 2 2 2

Map Mult Mult

4 2 8 6 12 10 16 14

Mult Mult Mult Mult Mult Mult Mult

Sequential algorithm

slide-8
SLIDE 8

TDDD56 Lab 3 August Ernstsson 2020

SkePU skeletons

1 2 3 4 5 6 7 8 2 2 2 2 2 2 2 2

Map Mult Mult

4 2 8 6 12 10 16 14

Mult Mult Mult Mult Mult Mult Mult

Parallel algorithm

slide-9
SLIDE 9

TDDD56 Lab 3 August Ernstsson 2020

SkePU syntax

int add(int a, int b, int m) { return (a + b) % m; } vec_sum(result, v1, v2, 5);

Map add

auto vec_sum = Map<2>(add);

add

Mult Mult Mult Mult Mult Mult Mult Mult
slide-10
SLIDE 10

TDDD56 Lab 3 August Ernstsson 2020

SkePU syntax, advanced

template<typename T> T abs(T input) { return input < 0 ? -input : input; } template<typename T> T userfunc(Index1D row, const Mat<T> m, const Vec<T> v) { T res = 0; for (size_t i = 0; i < v.size; ++i) res += m[row.i * m.cols + i] * v[i]; return abs(res); }

slide-11
SLIDE 11

TDDD56 Lab 3 August Ernstsson 2020

SkePU containers

  • Smart containers: Vector<T>, Matrix<T>, etc
  • Manages data across CPU and GPU
  • No data transfers unless necessary (lazy copying)
  • Keeps track of most recent writes
  • Memory consistency through software
slide-12
SLIDE 12

TDDD56 Lab 3 August Ernstsson 2020

SkePU build process

Executable Source-to-source compiler Backend compiler (e.g., GCC) Program sources SkePU runtime library Backend sources (C++, OpenCL, etc.)

Handled by lab Makefiles

slide-13
SLIDE 13

TDDD56 Lab 3 August Ernstsson 2020

Lab structure

  • Three exercises:
  • 1. Warm-up: dot product
  • 2. Averaging image filter + gaussian filter
  • 3. Median filter
slide-14
SLIDE 14

TDDD56 Lab 3 August Ernstsson 2020

  • 1. Dot product
  • Implement two variants of dot product:
  • With MapReduce skeleton
  • With Map + Reduce skeletons
  • Compare and contrast the variants
  • Why does SkePU have the MapReduce skeleton?
  • Measure with different backends and problem sizes
slide-15
SLIDE 15

TDDD56 Lab 3 August Ernstsson 2020

  • 2. Averaging filters
  • Averaging filter: find average color value in surrounding region
  • Gaussian filter: averaging filter with non-uniform weights
  • Use the MapOverlap skeleton

Original Average Gaussian

slide-16
SLIDE 16

TDDD56 Lab 3 August Ernstsson 2020

  • 3. Median filter
  • Median filter: find median color value in surrounding region
  • Requires sorting the pixel values in some way

Original Median

slide-17
SLIDE 17

TDDD56 Lab 3 August Ernstsson 2020

Image filters

  • Layout of image data in memory

1 pixel = 3 bytes!

slide-18
SLIDE 18

TDDD56 Lab 3 August Ernstsson 2020

Lab installation

  • Get from course website as usual
  • Slightly different from public SkePU distribution!
  • Pre-built binary
  • Runs on 64-bit Linux
slide-19
SLIDE 19

TDDD56 Lab 3 August Ernstsson 2020

Lab build process

Build lab program: > make bin/addone Run lab program: > bin/addone 100 CPU

CPU: Use sequential backend OpenMP: Use multithreaded backend OpenCL: Use GPU backend

slide-20
SLIDE 20

TDDD56 Lab 3 August Ernstsson 2020

A warning about warnings (and errors)

  • SkePU is a C++ template library
  • As such, gets very long and unreadable diagnostic

messages if used incorrectly!

  • Following the structure of the lab files should minimize

errors

  • Otherwise, be careful, and avoid using const!
slide-21
SLIDE 21

TDDD56 Lab 3 August Ernstsson 2020

Questions?