1. ARM Cortex-A9 MPCore ™ processor Presented by- Chris Cai (xiaocai2) Rehana Tabassum (tabassu2) Sam Mussmann (mussmnn2)

2. Background “The architectural simplicity of ARM processors leads to very small implementations, and small implementations mean devices can have very low power consumption. Implementation size, performance, and very low power consumption are key attributes of the ARM architecture.” ARM Architecture Reference Manual ARMv7-A edition

3. Background (2) ARM is RISC Uniform register file Load/store architecture Simple addressing

4. Background (3) The ARM Cortex-A9 processor is the high performance choice in a family of low power, cost-sensitive devices. The Cortex-A9 microarchitecture is delivered either as a Cortex-A9 single core processor or a scalable multicore processor: the Cortex-A9 MPCore ™ processor

5. Where is it used? Examples: -Apple A5 (iPhone 4S, iPad 2, iPad mini) http://en.wikipedia.org/wiki/ARM_Cortex-A9_MPCore#Implementations http://en.wikipedia.org/wiki/Iphone_4s

6. Where is it used? (2) Examples: -NVIDIA Tegra 2 (Motorola Xoom, Droid X2) http://en.wikipedia.org/wiki/ARM_Cortex-A9_MPCore#Implementations http://en.wikipedia.org/wiki/Motorola_Xoom

7. Where is it used? (3) Examples: -PlayStation Vita http://en.wikipedia.org/wiki/ARM_Cortex-A9_MPCore#Implementations http://en.wikipedia.org/wiki/PlayStation_Vita

8. What are its specs? The Cortex A9 core: - Gives 2.50 DMIPS/MHz/core (Dhrystone MIPS) - Generally clocked between 800MHz and 2GHz - Possible to run > 1GHz and < 250mW http://arm.com/products/processors/cortex-a/cortex-a9.php?tab=Specifications http://www.linuxfordevices.com/c/a/News/ARM-spins-multicoreenabled-Cortex-core/

9. Presentation Overview Micro-architecture Memory System Multi-core

10. Microarchitecture Overview Variable length, out of order, superscalar pipeline – Two instructions are fetched in one cycle – Issue up to 4 instructions per cycle into: Primary data processing pipeline Secondary data processing pipeline Load-store pipeline Compute engine (FPU/NEON) pipeline Speculative execution – Supporting virtual renaming of physical registers and removing pipelines stalls due to data dependencies

11. CortexA9 Microarchitecture www.arm.com/files/pdf/armcortexa-9processors.pdf Instruction Fetch Decode IssueRename Execute Writeback Memory

12. Instruction Fetch Instruction cache size: 16KB, 32KB, or 64KB Superscalar pipeline: fetching two instructions at once Branch Prediction: – Global History Buffer: 1K ~ 16K entries – Branch-Target Address Cache: 512 ~ 4K entries – Return stack of 4 x 32 bits Fast-loop mode: instruction loop that are smaller than 64 bytes often complete without additional instruction cache accesses

13. Instruction Decode Super Scalar Decoder -Capable of decoding two full instructions per cycle

14. Rename Register Renaming -Resolving data dependencies and unroll small loops by hardware

15. Issue Issue can be fed maximum of 2 instructions per cycle Issue can dispatch up to 4 instructions per cycle Out of order selection of instructions from queue

16. Execute Variable length Executing Stage (1 ~ 3 cycles) -Most Instructions finish within 1 cycle -Instruction which folds shifts and rotates can take 3 cycles ADD r0, r1, r2 (1 cycle) ADD r0, r1, r2 LSL #2 (2 cycle) Corresponds to a = b + (c << 2); ADD r0, r1, r2 LSL r3 (3 cycle) Corresponds to a = b + (c << d);

17. Execute (2) NEON Media Processing Engine -NEON technology supports instructions targeted primarily at audio, video, 3D graphics, image and speech processing. http://www.arm.com/files/pdf/AT_-_NEON_for_Multimedia_Applications.pdf

18. Execute (3) What is NEON? – NEON is a wide SIMD data processing architecture 32 registers, 64 bit wide or 16 registers, 128 bit wide – NEON instructions perform “Packed SIMD” processing Registers can be considered as “vector” of same data type Instructions perform the same operation in all lanes http://www.arm.com/files/pdf/AT_-_NEON_for_Multimedia_Applications.pdf

19. Execute (4) NEON Media Processing Engine supports vector computations on: -half-precision (16bit), single-precision (32bit), double- precision (64bit) floating-point numbers -8, 16, 32 and 64 bit signed and unsigned integers Supported Operations Include: -addition, subtraction, multiplication -maximum or minimum of a vector of operands -Inverse square-root approximation (y = x^-(1/2)) -many more

20. Memory dependent load-store instructions forwarded for resolution within memory system 2-level TLB structure – micro TLB 32 entries on data side and 32 or 64 entries on instruction side to reduce power consumed in translation and protection look-ups – main TLB http://infocenter.arm.com/help/topic/com.arm.doc.ddi0388i/DDI0388I_cortex_a9_r4p1_trm.pdf

21. Memory (2) Data prefetcher – monitor cache line requests by processor and cache misses to determine how much data to prefetch – can prefetch up to 8 independent data streams – prefetch and allocate data in the L1 data cache, as long as it keeps hitting in the prefetched cache line – When stop prefetching?

22. Memory Hierarchy CPU Instruct ion Cache Data Cache CPU Instruct ion Cache Data Cache CPU Instruct ion Cache Data Cache CPU Instruct ion Cache Data Cache Snoop Control Unit (SCU) Accelerator Coherence Port L2 Cache Main Memory Cortex A9 MPcore http://infocenter.arm.com/help/topic/com.arm.doc.ddi0407i/DDI0407I_cortex_a9_mpcore_r4p1_trm.pdf

23. L1 caches Non-unified -32 bytes line length -can be disabled independently 16, 32 or 64KB 4 - way associative support for Security Extensions I cache: VIPT D cache: PIPT -reduce number of caches flushes and refills and save energy http://infocenter.arm.com/help/topic/com.arm.doc.ddi0407i/DDI0407I_cortex_a9_mpcore_r4p1_trm.pdf CPU D$ I$ CPU D$ I$ CPU D$ I$ CPU D$ I$ SCU ACP L2 Cache Main Memory Cortex A9 MPcore AXI RW 64-bit bus AXI RW 64-bit bus

24. L2 cache shared, unified Off-chip 128KB to 8MB 4 to 16-way associative http://infocenter.arm.com/help/topic/com.arm.doc.ddi0407i/DDI0407I_cortex_a9_mpcore_r4p1_trm.pdf CPU D$ I$ CPU D$ I$ CPU D$ I$ CPU D$ I$ SCU ACP L2 Cache Main Memory Cortex A9 MPcore AXI RW 64-bit bus AXI RW 64-bit bus

25. Snoop Control Unit Integral part of cache memory systems Connects processors to memory system through AXI interfaces http://infocenter.arm.com/help/topic/com.arm.doc.ddi0407i/DDI0407I_cortex_a9_mpcore_r4p1_trm.pdf CPU D$ I$ CPU D$ I$ CPU D$ I$ CPU D$ I$ SCU ACP L2 Cache Main Memory Cortex A9 MPcore AXI RW 64-bit bus AXI RW 64-bit bus

26. SCU functions : -maintain data cache coherency -initiate L2 memory accesses -arbitrate between processors’ simultaneous request for L2 accesses -manages accesses from ACP does not support instruction cache coherency http://infocenter.arm.com/help/topic/com.arm.doc.ddi0407i/DDI0407I_cortex_a9_mpcore_r4p1_trm.pdf Snoop Control Unit (1)

27. Accelerator Coherence Port optional AXI 64-bit slave port allows to connect to non-cached system mastering peripherals and accelerators —DMA engine or cryptographic accelerator SCU enforces memory coherency http://www.arm.com/files/pdf/ARMCortexA-9Processors.pdf

28. Multi-Core http://www.arm.com/files/pdf/ARMCortexA-9Processors.pdf

29. Cache Coherence – MESI http://en.wikipedia.org/wiki/MESI_protocol

30. Cache Coherence – MESI (2) System Level Benchmarking Analysis of the Cortex A9 MPCore Roberto Mijat, ARM Connected Community Technical Symposium, 2009 ARM MPCore has optimizations to MESI: Duplicated tag RAMs All done in the Snoop Control Unit

31. Cache Coherence – MESI (2) System Level Benchmarking Analysis of the Cortex A9 MPCore Roberto Mijat, ARM Connected Community Technical Symposium, 2009 ARM MPCore has optimizations to MESI: Duplicated tag RAMs Cache-2-Cache transfer All done in the Snoop Control Unit

32. Cache Coherence – MESI (2) System Level Benchmarking Analysis of the Cortex A9 MPCore Roberto Mijat, ARM Connected Community Technical Symposium, 2009 ARM MPCore has optimizations to MESI: Duplicated tag RAMs Cache-2-Cache transfer Migratory Lines All done in the Snoop Control Unit

33. Generalized Interrupt Control System Level Benchmarking Analysis of the Cortex A9 MPCore Roberto Mijat, ARM Connected Community Technical Symposium, 2009 Which core services interrupts? GIC gives the programmer control Centralizes interrupts, then dispatches to individual core(s)