Power Parallel Programming for Mobile Devices

Transcription

1 Power Parallel Programming for Mobile Devices Calin Cascaval, Sept. 10, Qualcomm Technologies, Inc.

2 Outline Mobile compu-ng market - - what do users care about? Mobile compu-ng landscape - - how does the plahorm look like? Applica-ons and developers - - who is wri-ng code for mobile? Browser wars revisited - - why does it maler? Parallel heterogeneous programming made easier using MARE 2

3 Mobile Market 3

4 Mobile Market Size Billions of US $ Size of mobile market roughly doubling every two years > 20x (forecast) Mobile Desktop Servers HPC Source: IDC. HPC forecast Intersect360 4

5 What do users care about? 5 McKinsey s Consumer and Shopping Insights

6 What about next morning? 6 McKinsey s Consumer and Shopping Insights

7 The Internet of Everything Connec-vity Context Control An estimated 25B devices connected by 2020, talking to each other, making choices, and taking decisions Use context awareness to filter the data that reaches every person: time, location, taste Allow users to control and customize experience through new, mobile centric interfaces Enabling whole new classes of applica-ons, where power is even more cri-cal 7

8 What s in a smartphone?

9 Anatomy of a Mobile SoC: Qualcomm Snapdragon Source: hlp://

10 Mobile Constrains AA Energy Thermal 10

11 Mobile Applica-ons Connec-vity Social networking, sharing: Media rich, data heavy Gaming On device and multiplayer games with realistic graphics Browsing Information gateway Imaging Computational photography, search and processing of images and video 11

12 ahb2axi 16k I$ 8k D$ ARM926EJ-S AHB ARM Peripheral ahb2crif SG AP SP VBIF (DDR) DDR AXI FPB (AHB) VPE L1 $ / L2 $ Ctrl MEC Shared RAM pool (FIFO / Cache / Tile Buffer / Internal RAM) PPE OCMEM Data Mover & Controller TRE VBIF (OCMEM) OCMEM AXI memory PM4 Packet Buffers VS Buffer Index Buffers Vertex Objects Texture Objects Frame Buffer AXI FF (fix-function) External - other V B I F U C H E OXILI TOP arb arb arb arb L1 L1 L1 L1 2x TP 2x TP 2x TP 2x TP VSC VFD CP SP #3 PC SP #2 HLSQ RBBM CPU OXILI SS SP #1 SP #0 RB RB RB RB MA RB MA RB MA RB MA RB GMEM GMEM GMEM GMEM AHB Slave I/F V P C T S E / R A S Mobile Sodware Development Developers Think in terms of services and features Portable apps, performance not the main concern Use JavaScript frameworks Hardware Accel. DSP GPU CPU(s) VeNum VeNum CPU CPU VSP VPP L2 DMA Shared RAM Banks 4KB x 8 VSP-VPP I/F (FIFO) SS (shader system) CPU CPU VeNum VeNum 12 High%Power% Efficiency Low%Power% Efficiency

13 How can we easily write parallel applications that use all available cores in a battery-powered device? 13

14 Our Vision of a Mobile Sodware Stack Web Apps Na-ve Apps Performance: Domain Specific libraries - Exploit domain knowledge to provide composable libraries for all programmers - Hide hardware complexity JavaScript frameworks Portability: Zoomm Web App Engine - Use of concurrency to op-mize execu-on of Web Apps - Hardware exploita-on through the browser Browser Engine Domain Specific Parallel Libraries MARE Programmability: MARE - Parallel, heterogeneous programming made easier - Power and performance op-miza-ons 14

15 Browsers: why do we care?

16 Web Browser Usage Information access Web Applications 89% 85% Browsing 16 Source: Illuminas, 2013

17 Browser Execu-on Time Breakdown rendering javascript layout css Others parsing 5% 4% 19% 20% Insight: There is no one dominant component Paralleliza-on must address the en-re browser structure! 31% 21% ARM Cortex A9 ExecuDon Time Average over the top 30 Alexa sites (May 2010), WebKit browser on a 400MHz Cortex A9, Linux 17

18 Zoomm: Pervasive Concurrency URL User Interface Events Res. Manager Prefetching Image Decoding HTML code Per- Page DOM Per- Page Engine DOM Engine HTML Parsing Timers CSS Parsing CSS Parsing Events Styling JS code Layout Tree Per- Page JavaScript JavaScript Engine Engine Execu-on Execu-on Compila-on n Rendering Engine Layout Render 18 Cascaval et al.: Zoomm: a parallel web browser engine for mul-core mobile devices, PPoPP 2013

19 Zoomm: Page Load Performance WebKit Zoomm Seconds CNN BBC Yahoo Guardian NYT Facebook Engadget QQ HTC Jetstream, MSM8660, 2- core, 1.5GHz, Aug

20 Qualcomm MARE Mul-core Asynchronous Run-me Environment

21 The Smartphone of 2010 All applica-ons run on a single- core Qualcomm Snapdragon processor at 1GHz App 4 App 3 App 2 App 1 Core 21

22 The Smartphone of 2013 Mul-ple applica-ons can simultaneously run on a quad- core Snapdragon at 2+GHz App 1 App 2 App 3 App 4 Core 1 Core 2 Core 3 Core 4 22

23 What is Qualcomm MARE? MARE is a programming model and a run-me system that provides simple yet powerful abstrac-ons for parallel, power- efficient sodware Simple C++ API allows developers to express concurrency User- level library that runs on any Android device, and on Linux, Mac OS X, and Windows plahorms The goal of MARE is to reduce the effort required to write apps that fully u-lize heterogeneous SoCs 23

24 MARE Workflow Focus on your application and not on the hardware Understand algorithms Par--on algorithms into independent units of work Setup task dependencies Link with MARE Run-me Algorithms MARE Application MARE Runtime

25 Qualcomm MARE API Concepts Tasks are units of work that can be asynchronously executed Groups are sets of tasks that can be canceled or waited on 25

26 Advantages of using Qualcomm MARE Simple Produc-ve Efficient Tasks are a natural way to express parallelism Focus on application logic, not on thread management Task dependences allow the MARE runtime to perform more intelligent scheduling decisions 26

27 Hello World! #include <stdio.h> #include <mare/mare.h> int main() { mare::runtime::init(); // Initialize MARE auto hello = mare::create_task([] {printf("hello );}); // Create task auto world = mare::create_task([] {printf( World! );}); // Create task hello >> world; // Set dependency mare::launch(hello); mare::launch(world); // Launch hello task // Launch world task mare::wait_for(world); runtime::shutdown(); return 0; } // Wait to complete // Shutdown MARE 27

28 Parallel Programming PaLerns in MARE MARE offers several commonly used parallel palerns: mare::pfor_each processes the elements of a collec-on mare::pscan performs an in- place parallel prefix opera-on for the elements of a collec-on. mare::transform performs a map opera-on on the elements of a collec-on Programmers can also create their own palerns using the MARE API Parallel data structures: concurrent queues, hash tables, etc. 28

29 Bullet Physics Paralleliza-on using MARE Three major components: about 75% of total execu-on -me Constraint Solver Finds non- interac-ng groups of objects and solves as independent tasks Coarse- grained task- paralleliza-on Rendering Very coarse- grained task- paralleliza-on: offload to a separate thread to allow for con-nuous simula-on Narrow Phase Collision Detec-on Fine- grained data- paralleliza-on About 2x FPS improvement on typical mobile devices Drama-cally improves the smoothness of rendering and overall user experience 29

30 Bullet Physics Paralleliza-on MARE Serial Frames Per Second Frame Number 30

31 How can You try MARE? Stay tuned! Planning preview availability on Qualcomm Developer Network Fall 2013 Beta release for the mul-core version We Want Your Feedback! What scenarios will you use it for? Usability and features of the library 31

32 Acknowledgments Manticore team Pablo Montesinos Ortego, Michael Weber, Wayne Piekarski, Behnam Robatmili, Dario Suarez-Gracia, Vrajesh Bhavsar, Jimi Xenidis, Han Zhao, Kishore Puskuri Interns Madhukar Kedlaya, Christian Delozier, Freark Van Der Berg, Christoph Kershbaumer Former members Mehrdad Reshadi, Seth Fowler, Alex Shye, Dilma DaSilva Qualcomm Nayeem Islam, Mark Bapst, Charles Bergan, Matt Grob, Ben Gaster MulticoreWare Wen-Mei Hwu, Mark Allender, Kevin Wu 32

33 Challenges Power and performance Custom hardware for power efficiency. Heterogeneity is the game. Expertly designed and optimized frameworks that hide hardware complexity Programmability Tools and languages to enable programmers to understand and express concurrency and application semantics Composable building blocks Mobile has exci-ng applica-ons and a huge impact 33

34 Legal Disclaimers All opinions expressed in this presentation are mine and may not necessarily reflect those of my employer. Qualcomm and Snapdragon are trademarks of Qualcomm Incorporated, registered in the United States and other countries. All Qualcomm Incorporated trademarks are used with permission. Other products and brand names may be trademarks or registered trademarks of their respective owners. 34

35 35