Accelerating performance on Qt and WebKit for the MIPS architecture
- 1. Accelerating performance on Qt and WebKit for the MIPS architecture Delivering the Best Web Experience for Embedded Devices
- 2. Agenda • MIPS Introduction • Emerging Trends • Web Connectivity • JIT Optimizations for WebKit • Benchmarks • Summary
- 3. MIPS Technologies Overview MIPS®: Number One Processor Architecture for the Digital Home MIPS develops & licenses processor IP for the embedded semiconductor market Hundreds of licenses Billions of units shipped At the core of today’s most innovative digital consumer products
- 4. MIPS Segments & Markets Consumer Communication DTV DVD STB VoIP WiFi Modem/RG DPF Home Home NAS Entertainment Networking DMA DSC/DVC Printers & Copiers Personal Other Entertainment Segments Internet Games Infra PND/PMP MID/Netbook Telematics Appliances Mobile Microcontrollers, Enterprise
- 5. MIPS Processors: Owning “The Digital Home” -The Growth Sector of Semiconductors for 2009- End Market Share & Customers Digital TV 60% Set Top Box -Cable 72% -IPTV 75% Our customers include DVD -Blu-ray 75% -DVD 72% Digital Camera 31% Mobile TV & GPS PMP 60%
- 6. MIPS 32-Bit Processor Core Families 1004K: Multi-threaded (34K), Multiprocessor (1-4 cores) Broad range Coherence Management Unit 1004K of 1.3+ GHz (40nm), 1.5DMIPS/MHz/Core synthesizable processors, 74K: Superscalar 15-stage pipeline optimized for 1.1 GHz in 65nm (prod’n frequency) 74K low-power Up to 2.7GHz in 40nm convergent consumer 34K: Multi-threading 34K multimedia applications 24KE: DSP extensions 24KE 24K: 8-stage pipeline 800 MHz (65nm) 24K m M14Kc i c M14Kc 4KE: Cache, MMU 4KE r o M4K: MCU, Low Cost M M4K M14K I P S 4KS: Security 4KS microMIPS, MIPS32 Reduced interrupt latency AHB, advanced debug
- 7. Agenda • MIPS Introduction • Emerging Trends • Web Connectivity • JIT Optimizations for WebKit • Benchmarks • Summary
- 8. MIPS in Converging Devices Universal Media Player IP Phone Networked Multi-Protocol Media Tank IP Phone IPOD Video LCD TV Recorder with Widgets IP Set-top Box HD Home Entertainment Camcorder Device Movie Player STB Portable Media Player Networked Digital Photo Frame Digital Media Player Personal Media Retail Shopping Player Terminal
- 9. Qt Is In Homes Today Devices –Roku Netflix Player –Neuros OSD –Blu-Ray Players Software –MythTV PVR –LinuxMCE 9 © 2009 Nokia
- 10. Bring Internet Media into the Home Internet Connected Web Streaming Infotainment Portals “HD” Devices INTERNET HTTP H.264 Flash Etc. 10
- 11. Typical STB Software Architecture 3rd Party OEM App App 3rd Party Middleware Qt WebKit Browser Toolkit STB Adobe® Qt GUI ToolKit Java VM Middleware Flash Lite® Qt for Embedded Linux DRMs Audio Video Demux HDMI CAs µCode µCode µCode Library Linux Kernel MIPS MIPS Audio Video MIPS Peripheral IPU Security CPU DSPs Decoders CPU I/O MIPS 74K
- 12. Agenda • MIPS Introduction • Emerging Trends • Web Connectivity • JIT Optimizations for WebKit • Benchmarks • Summary
- 13. What Drives Web Performance? • Highly web content specific – Interactive AJAX content vs Simple HTML – Image heavy vs. text • Page rendering • Image rendering • Javascript • Interactive web content is heavily dependent on Javascript performance – PC: Chrome vs. IE8 vs. Firefox “performance” judged mainly on Javascript engine performance
- 14. Faster Javascript Leads to “Faster” Web Experience • PC browser performance differences directly related to Javascript engine – SquirrelFish Extreme (Nitro) • Javascript Virtual Machine engine used in Webkit Open Source Project • Some projects that make use of Webkit: Safari, Qt, Arora, Midori – V8 • Chrome – Tracemonkey (w/ nanoJIT) • Firefox 3.5.x – IE8
- 15. JIT and the Javascript Engine • Developing a JIT speeds up browsing performance on Javascript-based webpages and applications using it as a scripting language • JIT techniques also helpful for Tamarin – Flash 10 – Yahoo Widgets • Status: – JIT completed and submitted to Webkit project for review/acceptance – MIPS incorporating into MIPS-Android HD development
- 16. Agenda • MIPS Introduction • Emerging Trends • Web connectivity • JIT optimizations for WebKit • Benchmarks • Summary
- 17. MIPS JIT for WebKit JavaScriptCore • Development environment: Debian 5.0 (lenny). – Debian runs on top of Qemu simulators, MIPS Malta boards (FPGA), or Sigma boards. • Qt libraries are installed on Debian. – # apt-get install libqt4-dev • Webkit is configured and built with Qt (--qt) on Debian. • To accelerate compilation time, we copy Qt libraries, Qt headers, Qt generated files, and Makefile to an x86 Linux host. Then, we use MIPS linux-cross toolchains from CodeSourcery to cross-compile.
- 18. Implementing JIT • JIT frameworks need to assign registers for special purposes. The selection of registers is important. – Some registers are used for returned values. – Some registers must be preserved across function calls. • The design for trampoline is important. – We need to set up parameters in the stack to be passed to functions. – We need to support exceptions. – For MIPS position-independent code, we need to preserve the value of global-pointer register ($28) in the trampoline code.
- 19. Implementing JIT (cont.) • MIPS PC-relative branches cover 16-bit ranges. – When branches are too far away, direct jump (J) covers 26-bit ranges. – LUI/ORI/JR covers 32-bit ranges. • J and LUI/ORI/JR are not PC-relative instructions. • When the buffer to hold instructions is moved, the target of J and LUI/ORI/JR must be re-calculated and updated. • The relaxation of branches need four extra words. A short branch is created to jump over four extra words to speed up. • After JIT code is generated, we need to flush data cache using synci (a user-level instruction).
- 20. All WebKit JIT optimizations are supported • ENABLE_JIT • ENABLE_JIT_OPTIMIZE_CALL • ENABLE_JIT_OPTIMIZE_NATIVE_CALL • ENABLE_JIT_OPTIMIZE_PROPERTY_ACCESS • ENABLE_JIT_OPTIMIZE_METHOD_CALLS • YARR_JIT
- 21. Agenda • MIPS Introduction • Emerging Trends • Web Connectivity • JIT Optimizations for WebKit • Benchmarks • Summary
- 22. Web Performance Benchmarks Relative Javascript Benchmark Performance Interpreted JIT H/W Platforms: 74Kf = Sigma Designs SMP8642 @ 660 MHz A8 = TI OMAP 3530 @ 500 MHz Atom = Intel Atom 330 @ 1.6 GHz >2x faster than ARM Same performance as Atom @ a fraction of the power and cost 74Kf_32K/32K_0KB 74Kf_32K/32K_256K (est) A8_16K/16K_256K Atom_32K/24K_512K V8/Sunspider benchmark suite All results normalized to 1 GHz
- 23. Agenda • MIPS Introduction • Emerging Trends • Web Connectivity • JIT Optimizations for WebKit • Benchmarks • Summary
- 24. Best Web and Multimedia Experience Driving Performance 2x performance of ARM! Fraction of the cost and power consumption of Intel!