Илья Перминов

Тимур Палташев

23.10.2013

Гетерогенная архитектура HSA:

экосистема для CPU/GPU/DSP

НИУ ИТМО

2 МСКФ’2013 | October 2013 |

ЧТО ТАКОЕ HSA

CPU Memory (Coherent)

CPU

1

CPU

N …

GPU Memory

GPU CPU

2

PCIe

Heterogeneous System Architecture

= HSA

3 МСКФ’2013 | October 2013 |

ЧТО ТАКОЕ HSA

CPU Memory (Coherent)

CPU

1

CPU

N …

GPU Memory

GPU CPU

2

PCIe

GPGPU = ? Гетерогенные

вычисления

4 МСКФ’2013 | October 2013 |

ЧТО ТАКОЕ HSA

GPGPU = ? Гетерогенные

вычисления

5 МСКФ’2013 | October 2013 |

ЧТО ТАКОЕ HSA

GPGPU = ? Гетерогенные

вычисления

6 МСКФ’2013 | October 2013 |

ЧТО ТАКОЕ HSA

GPGPU = ? Гетерогенные

вычисления

7 МСКФ’2013 | October 2013 |

HSA: HETEROGENEOUS SYSTEM ARCHITECTURE

ПАРАЛЛЕЛЬНАЯ

НАГРУЗКА

(ОДНО)ПОТОЧНАЯ

НАГРУЗКА

hUMA (MEMORY)

8 МСКФ’2013 | October 2013 |

HSA HIGH LEVEL ARCHITECTURE

CPU

GPU

Audio

Process

or

Video

Hardware

DSP

Image

Signal

Processi

ng

Fixed

Function

Acctr

Encode

Decode

Sh

are

d M

em

ory

Co

here

ncy,

User

Mo

de

Qu

eu

es

9 МСКФ’2013 | October 2013 |

СОСТАВ HSA FOUNDATION - 2013

Founders

Promoters

Supporters

Contributors

Academic

10 МСКФ’2013 | October 2013 |

PHYSICAL MEMORY

SHARED VIRTUAL MEMORY (TODAY)

Различные вирутальные адресные пространства

CPU0 GPU

VIRTUAL MEMORY1

PHYSICAL MEMORY

VA1->PA1 VA2->PA1

VIRTUAL MEMORY2

11 МСКФ’2013 | October 2013 |

ЧТО МЫ ИМЕЕМ СЕГОДНЯ

Application OS GPU

Transfer

buffer to GPU Copy/Map

Memory

12 МСКФ’2013 | October 2013 |

ЧТО МЫ ИМЕЕМ СЕГОДНЯ

Application OS GPU

Transfer

buffer to GPU Copy/Map

Memory

Queue Job

Schedule Job

Start Job

Finish Job

13 МСКФ’2013 | October 2013 |

ЧТО МЫ ИМЕЕМ СЕГОДНЯ

Application OS GPU

Transfer

buffer to GPU Copy/Map

Memory

Queue Job

Schedule Job

Start Job

Finish Job

Schedule

Application

Get Buffer

Copy/Map

Memory

14 МСКФ’2013 | October 2013 |

ОБЩАЯ КОГЕРЕНТНАЯ ПАМЯТЬ (HUMA)

15 МСКФ’2013 | October 23, 2013 |

Physical Memory

GPU

HW Coherency

Virtual Memory

C

P

U

Entire memory space:

Both CPU and GPU can access and

allocate any location in the system’s

virtual memory space

Cache Cache

Coherent Memory: Ensures CPU and

GPU

caches both see

an up-to-date view

of data Pageable memory:

The GPU can seamlessly

access virtual memory

addresses that are not

(yet)

present in physical

memory

HSA KEY FEATURES

16 МСКФ’2013 | October 23, 2013 |

GPU

CPU

CPU Memory GPU Memory

| |

| |

|

| |

| |

|

| |

| |

|

| |

| |

|

CPU explicitly copies data to GPU memory

GPU completes computation

CPU explicitly copies result back to CPU memory

Без HSA

17 МСКФ’2013 | October 23, 2013 |

CPU / GPU Uniform Memory

| |

| |

|

| |

| |

|

HSA CPU simply passes a pointer to GPU

GPU complete computation

CPU can read the result directly – no copying needed!

GPU

CPU

18 МСКФ’2013 | October 2013 |

ОБЩАЯ ПАМЯТЬ

Application OS GPU

Transfer

buffer to GPU Copy/Map

Memory

Queue Job

Schedule Job

Start Job

Finish Job

Schedule

Application

Get Buffer

Copy/Map

Memory

19 МСКФ’2013 | October 2013 |

ОБЩАЯ КОГЕРЕНТНАЯ ПАМЯТЬ

Application OS GPU

Transfer

buffer to GPU Copy/Map

Memory

Queue Job

Schedule Job

Start Job

Finish Job

Schedule

Application

Get Buffer

Copy/Map

Memory

20 МСКФ’2013 | October 2013 |

УПРАВЛЕНИЕ ОЧЕРЕДЯМИ ЗАДАЧ

21 МСКФ’2013 | October 2013 |

ДИСПЕТЧЕРИЗАЦИЯ В РЕЖИМЕ ПОЛЬЗОВАТЕЛЯ

Application OS GPU

Transfer

buffer to GPU Copy/Map

Memory

Queue Job

Schedule Job

Start Job

Finish Job

Schedule

Application

Get Buffer

Copy/Map

Memory

22 МСКФ’2013 | October 2013 |

HSA

Application OS GPU

Queue Job

Start Job

Finish Job

23 МСКФ’2013 | October 2013 |

ПОДДЕРЖКА ТЕХНОЛОГИЙ ПРОГРАММИРОВАНИЯ

C++ AMP

C++

C#

OpenCL

OpenMP

Java

Python

…

24 МСКФ’2013 | October 2013 |

LANGUAGE SUPPORT ALLOWS HSA TO HAVE

MULTIPLE SOFTWARE EXECUTION MODELS

Kernel Fusion

Driver (KFD)

HSA Core

Runtime HSA Finalizer

HSA Helper

Libraries

OpenCL App

OpenCL

Runtime

Java

Sumatra App

Java JVM

Runtime

Python

Fabric

App

Fabric Engine

RT

HSAIL

25 МСКФ’2013 | October 2013 |

JAVA ENABLEMENT BY APARAPI

Developer creates Java™ source

Source compiled to class files (bytecode) using standard compiler

Aparapi = Runtime capable of converting Java™ bytecode to OpenCL™

For execution on any

OpenCL™ 1.1+ capable device

OR execute via a thread pool if OpenCL™ is not available

26 МСКФ’2013 | October 2013 |

ПЛАНЫ ПО ПОДДЕРЖКЕ JAVA

CPU ISA GPU ISA

JVM

Application

APARAPI

GPU CPU

OpenCL™

CPU ISA GPU ISA

JVM

Application

APARAPI

HSA CPU HSA CPU

HSA Finalizer

HSAIL

CPU ISA GPU ISA

JVM

Application

APARAPI

HSA CPU HSA CPU

HSA Finalizer

HSAIL

HSA Runtime

LLVM Optimizer

IR

CPU ISA GPU ISA

Sumatra Enabled JVM

Application

HSA CPU HSA CPU

HSA Finalizer

HSAIL

27 МСКФ’2013 | October 2013 |

LOOKING FOR FACES IN ALL THE RIGHT PLACES

Quick HD Calculations

Search square = 21 x 21

Pixels = 1920 x 1080 = 2,073,600

Search squares = 1900 x 1060 = ~2 Million

28 МСКФ’2013 | October 2013 |

LOOKING FOR DIFFERENT SIZE FACES BY SCALING THE VIDEO FRAME

29 МСКФ’2013 | October 2013 |

LOOKING FOR DIFFERENT SIZE FACES BY SCALING THE VIDEO FRAME

More HD Calculations

70% scaling in H and V

Total Pixels = 4.07 Million

Search squares = 3.8 Million

30 МСКФ’2013 | October 2013 |

HAAR CASCADE STAGES

Feature l

Feature m

Feature p

Feature r

Feature q

Feature k

Stage N

Stage N+1

Face still possible? Yes

No

REJECT FRAME

31 МСКФ’2013 | October 2013 |

22 CASCADE STAGES, EARLY OUT BETWEEN EACH

STAGE 22 STAGE 21 STAGE 2 STAGE 1

NO FACE

FACE CONFIRMED

Final HD Calculations

Search squares = 3.8 million

Average features per square = 124

Calculations per feature = 100

Calculations per frame = 47 GCalcs

Calculation Rate

30 frames/sec = 1.4TCalcs/second

60 frames/sec = 2.8TCalcs/second

…and this only gets front-facing faces

32 МСКФ’2013 | October 2013 |

CASCADE DEPTH ANALYSIS

0

5

10

15

20

25Cascade Depth

20-25 15-20 10-15 5-10 0-5

33 МСКФ’2013 | October 2013 |

UNBALANCING DUE TO EXITS IN EARLIER

CASCADE STAGES

When running on the GPU, we run each search rectangle on a separate work item

Early out algorithms, like HAAR, exhibit divergence between work items

Some work items exit early

Their neighbors continue

SIMD packing suffers as a result

Live

Dead

34 МСКФ’2013 | October 2013 |

0

10

20

30

40

50

60

70

80

90

100

1 2 3 4 5 6 7 8 9-22

Tim

e (

ms)

Cascade Stage

A10-4600M (6CU@497Mhz, 4 cores@2700Mhz)

GPU CPU

PROCESSING TIME/STAGE

AMD A10 4600M APU with Radeon™ HD Graphics; CPU: 4 cores @ 2.3 GHz (turbo 3.2 GHz); GPU: AMD Radeon HD 7660G,

6 compute units, 685MHz; 4GB RAM; Windows 7 (64-bit); OpenCL™ 1.1 (873.1)

35 МСКФ’2013 | October 2013 |

ПРОИЗВОДИТЕЛЬНОСТЬ CPU-VS-GPU

AMD A10 4600M APU with Radeon™ HD Graphics; CPU: 4 cores @ 2.3 MHz (turbo 3.2 GHz); GPU: AMD Radeon HD 7660G,

6 compute units, 685MHz; 4GB RAM; Windows 7 (64-bit); OpenCL™ 1.1 (873.1)

0

2

4

6

8

10

12

0 1 2 3 4 5 6 7 8 22

Imag

es/S

ec

Number of Cascade Stages on GPU

AMD A10-4600M APU (6CU@497Mhz, 4 cores@2700Mhz)

CPU HSA GPU

36 МСКФ’2013 | October 2013 |

HAAR SOLUTION RUN DIFFERENT CASCADES ON GPU AND CPU

By seamlessly sharing data between CPU and GPU,

HSA allows the right processor to handle its appropriate workload

+2.5x

-60%

INCREASED

PERFORMANCE DECREASED ENERGY

PER FRAME

37 МСКФ’2013 | October 2013 |

LINES-OF-CODE AND PERFORMANCE FOR

DIFFERENT PROGRAMMING MODELS

AMD A10-5800K APU with Radeon™ HD Graphics – CPU: 4 cores, 3800MHz (4200MHz Turbo); GPU: AMD Radeon HD 7660D, 6 compute units, 800MHz; 4GB RAM. Software – Windows 7 Professional SP1 (64-bit OS); AMD OpenCL™ 1.2 AMD-APP (937.2); Microsoft Visual Studio 11 Beta

0

50

100

150

200

250

300

350

LO

C

Copy-back Algorithm Launch Copy Compile Init Performance

Serial CPU TBB Intrinsics+TBB OpenCL™-C OpenCL™ -C++ C++ AMP HSA Bolt

Pe

rform

an

ce

35.00

30.00

25.00

20.00

15.00

10.00

5.00

0 Copy-back

Algorithm

Launch

Copy

Compile

Init.

Copy-back

Algorithm

Launch

Copy

Compile

Copy-back

Algorithm

Launch

Algorithm

Launch

Algorithm

Launch

Algorithm

Launch

Algorithm

Launch

(Exemplary ISV “Hessian” Kernel)

(“Hessian” kernel)

38 МСКФ’2013 | October 2013 |

MORE INFO AT

http://hsafoundation.com

39 МСКФ’2013 | October 2013 |

HSA РЕШЕНИЯ ОТ ARM (CORTEX-A15 MALI-T600)

• Google Chromebook

• Google Nexus 10

• InSignal Arndale Community Board

40 МСКФ’2013 | October 2013 |

AMD “KAVERI” 2.0 PLATFORM DETAILS

APU Features

New “SteamrollerB” CPU Core with up to 20% performance* increase over “Richland” – Up to 4 Steamroller cores and 4 MB total L2 cache

– Temperature Smart Turbo Core

New Power Optimized Graphics Core Next with up to 30% performance* increase over “Richland” – Multiple DirectX® 11.1 GPU configurations

– Dual Graphics support with “Crystal” Series

New AMD fixed function acceleration – UVD 4.2 Universal Video Decode Engine

– VCE 2.0 Video Compression Engine

– New ACP (Audio Co-Processor)

– SAMU 2.1 Secure Asset Management Unit

New Display and I/O Features – New PCIe Gen3 x16 for discrete GPU expansion

– New Dedicated PCIe SSD interface

– PCIe Gen2 1 x4, 4 x1, 1x4 UMI

– 4096 x 2160 resolution per display output

– 16K x 16K Max Eyefinity SLS resolution

– “Lightning Bolt” Docking Solution

Power Management and Battery Life – 35W to 15W TDPs

– Targeting ~11 Hours Battery Life MM07* 62WHr (~8.5Hr 45WHr) 14” 1366x768 eDP panel

– New AMD Start Now 3.0 with smart sleep

*Pre-silicon projection and not based on actual measured data

QUESTIONS AND

ANSWERS