[MOSUT20150131] Linux Runs on SoCKit Board with the GPGPU
Download as PPTX, PDF2 likes1,325 views
本文档探讨了在Sockit开发板上运行Linux操作系统与Nyuzi处理器技术的集成。详细描述了Nyuzi处理器的特性、引导过程以及FPGA设计中软件IP的开发与接口设计。作者提供了一些开发流程和操作细节，以协助开发者在FPGA上实现高效运算。
![[MOSUT20150131] Linux Runs on SoCKit Board with the GPGPU](https://image.slidesharecdn.com/linuxrunsonsockitboardwiththegpgpu-150127112230-conversion-gate02/85/MOSUT20150131-Linux-Runs-on-SoCKit-Board-with-the-GPGPU-23-320.jpg)
![經過下列步驟,就可以撰寫程式來測試FPGA是
否有加成功? Yes, it works.
• Converting .sof to .rbf
• Device tree generation
• 給寫driver用的。概念上就是知道I/O的name，就可以透過查詢
Device Tree得到address來操作[跨板子設計]
• Preloader image generation
• 會產生出preloader的source code ，所以有些FPGA初始化設定
可以在這裡撰寫編譯
• Configure Pinmux
• Preloader and U-Boot Customization - v13.1
• Updating SD card
過程與細節可參考我的研究紀錄](https://image.slidesharecdn.com/linuxrunsonsockitboardwiththegpgpu-150127112230-conversion-gate02/85/MOSUT20150131-Linux-Runs-on-SoCKit-Board-with-the-GPGPU-24-320.jpg)
[MOSUT20150131] Linux Runs on SoCKit Board with the GPGPU
- 1. Linux Runs on the SoCKit Board with the GPGPU AJ (張仁傑) <ajblane0612@gmail.com> 2015.1.31
- 2. Outline • Nyuzi processor • Global System Reference Design (GSRD)
- 3. Nyuzi Processor特色-SIMD • A custom RISC instruction set (load-store architecture) • 30 general purpose scalar registers • 32 vector registers, each with 16 lanes • Instructions are 32-bit fixed length and must be 32-bit aligned des op2 op1 mask add_i_mask v1, s2, v2, v3
- 5. 對Nyuzi processor看法: 如果在Nyuzi processor上面加上OS，context switch感覺是個悲劇 不過作者為了3D render，有在上面實作一 個小型OS來切換(?)
- 7. $ jload boot.hex $ bin/serial_boot program.elf GPGPU Niso II The DE2-115 Board Ubutun14.10 USB Blaster-JTAG Clock rate: ~60 Mhz
- 8. 如何實現?
- 9. Niso II The DE2-115 Board GPGPU Memory UART 0 0xffffffff serial_boot.c 解譯program.ELF 4. boot.c等待serial_boot.c 的command並開始寫入 hex code至memory 5. serial_boot.c告知 boot.c,code已經搬移結束.跳 至start.s stack 0x2000 code JTAG Loader PC start.s boot.c stack 0x200000 1. GPGPU reset後從address 0開 始執行 2. start.s 設定sp = 0x2000 3. 跳至boot.c 6. 設置PC開始執行,並把 sp設置0x200000 (?) boot.hex Ubutun14.10 step1 step6 step4 告知GPGPU把HEX code 寫至board memory start.s boot.c
- 10. $ jload boot.hex $ bin/serial_boot program.elf 怎那麼麻煩? 作者沒有在板子上使用 OS(提供服務)
- 11. 我們有什麼? • 硬體: the SoCKit board • RocketBoards.org組織提供的GSRD 14.1 for the SoCKit board 我怎會有這個板子? • 開放原始碼專案協作 (2014 年暑期)
- 12. GSRD • 提供開發者有一個Linux 環境可以操作FPGA • Software IP • Linux Driver • File I/O來操作Software IP • 開發流程概念: • 先有個Image,先run.再針 對各部分update • 細節可參考 • Booting Linux Using Prebuilt SD Card Image • Creating and Updating SD Card
- 13. 來看Altera Hard Process System (HPS)
- 14. • H2F • High bandwidth • Light weight • F2H (FPGA to HPS system) • F2S (FPGA-to-HPS SDRAM Interface)
- 15. Cyclone V HPS Memory Map • 在SoCKit版中，Cortex-A9可主動透過 H2F傳資料給software IP • LWAXI • 0xFF200000 ~0xFF3FFFFF(2MB) • HPS2FPGA (high bandwidth) • 0xC000000開始
- 16. Golden Hardware Reference Design (GHRD) • On-Chip RAM需高頻寬傳輸， 控制周邊就不須使用高頻寬
- 19. 參考SoCKit Video Server and Video IP (VIP) Reference Design設 LWAXI F2H 使用Qsys連接AXI master/slave
- 20. Software IP使用Qsys,有這麼簡單@@? • 必須了解AXI master slave • AMBA AXI and ACE Protocol Specification AXI3, AXI4, and AXI4-Lite, ACE and ACE-Lite • 我不了解,但我會改 Demo AXI Memory Design Example • Qsys System Design Tutorial • 注意細節: • 所有AXI master或者slave的pin都要宣告 • AXI slave's ID bit width • If an AXI slave's ID bit width is smaller than required for your system, the AXI slave response might not reach all AXI • maximum_master_id_width_in_the_interconnect + log2(number_of_masters_in_the_same_interconnect) • 所以我把Nyuzi Processor製作成Qsys元件
- 24. 經過下列步驟,就可以撰寫程式來測試FPGA是 否有加成功? Yes, it works. • Converting .sof to .rbf • Device tree generation • 給寫driver用的。概念上就是知道I/O的name，就可以透過查詢 Device Tree得到address來操作[跨板子設計] • Preloader image generation • 會產生出preloader的source code ，所以有些FPGA初始化設定 可以在這裡撰寫編譯 • Configure Pinmux • Preloader and U-Boot Customization - v13.1 • Updating SD card 過程與細節可參考我的研究紀錄
- 26. Niso II The DE2-115 Board GPGPU Memory UART 0 0xffffffff serial_boot.c 解譯program.ELF 4. boot.c等待serial_boot.c 的command並開始寫入 hex code至memory 5. serial_boot.c告知 boot.c,code已經搬移結束.跳 至start.s stack 0x2000 code JTAG Loader PC start.s boot.c stack 0x200000 1. GPGPU reset後從address 0開 始執行 2. start.s 設定sp = 0x2000 3. 跳至boot.c 6. 設置PC開始執行,並把 sp設置0x200000 (?) boot.hex Ubutun14.04 step1 step6 step4 告知GPGPU把HEX code 寫至board memory start.s boot.c
- 27. Cortex-A9 The SoCKit Board GPGPU Memory 0 0xffffffff stackcode PC Ubutun14.04 LINUX HPS SD card Program.hex produced by elf2hex LWAXI Address data 0xFF200000 RESET 0xFF200004 Code base address 0xFF200008 Stack address (*highest* stack) 0xFF20000C Heap address 0xFF200010 Frame address 1. 透過Linux宣告code空間,並設置code base address、stack address • fd = open(“/dev/mem”,..) • used_ptr = mmap(..,fd,0xff200000) 2. cortex-A9從SD card搬code至指定memory 3. 透過cortex-A9來控制GPGPU RESET signal step 1 step 2 F2H LWAXI F2H
- 28. boot.hex • How to build • Elf2hex is as a loader • Tutorial: Creating an LLVM Backend for the Cpu0 Architecture
- 29. LLVM-nyuzi is built to the target
- 30. Cortex-A9 The Sockit Board LINUX HPS SD card ld.mcld llvm-ar elf2hex clang .s, .c, .cpp .o .a.o .elf .hex .hex