ARM for Wireless Applications ARM11 Microarchitecture On the ARMv6 Connie Wang.

ARM for Wireless Applications

ARM11 Microarchitecture

On the ARMv6

Connie Wang

Advanced RISC Machines

• >75% of market for 32-bit RISC microprocessors

• ARM11 Design led by Ian Devereux

Demands of Wireless Applications

• High performance

• Low power

• Small size

• Cost

RISC for Wireless

• Strengths:– Clock rate

– Pipelining

• Weaknesses:– High code density

– Power consumption

ARM11 for Wireless

• Strengths Enhanced:– Clock rate

• Optimized interrupt and exception handling

• Minimized context switch cost

• Instruction set for media

– Pipelining• Decoupled for high

bandwidth• Retire before execution

• Weaknesses Reduced:– High code density

• ISA extensions

• Optional application specific and/or VFP coprocessors

– Power consumption• Architecture and

instructions reduce clock rate

• Clock gate control

ARM11 Microarchitecture

• First implementation of ARMv6 architecture• 8-stage pipeline• 64-bit datapaths• Frequency: up to 750 MHz, 350 – 500+ MHz

worst case. 400 – 1,200 Dhrystone MIPS• Power: 0.4 mW/MHz worst case: 0.13µm 1.2V• Will be released to licensees in Q4 2002

ARMv6

• Media support: SIMD extensions

• Improved interrupt latency

• ISA extensions THUMB, DSP, Jazelle

• 100% backwards compatibility to ARMv5

THUMB Instruction Set

• 32-bit performance for 16-bit systems• 32-bit instructions re-coded to 16-bit op-

codes• 32-bit ROM stores 2 THUMB instructions

per word• Decompressed in pipeline to ARM

instruction equivalents• Improves code density by 35%

DSP Instruction Set

• Application accelerator for Digital Signal Processor performance

• Can load/store registers by pairs

• 16x16 or 32x16 MAC in one cycle

• Utilized in MAC pipeline

Jazelle Instruction Set

• Support for entering/exiting Java applications

• Fetches/decodes Java bytecodes, maintains a Java operand stack

• Creates a state that imitates a Java processor

• OS controls low-cost switch between Java and ARM/THUMB states

SIMD Instruction Set

• Parallel processing of 2x16-bit or 4x8-bit operands

• Four new Greater than or Equal to status bits (GE[3:0]) for MAC calculations

• Eliminates need for very high clock frequencies and hardware accelerators

• 2 – 4 x performance improvement for multimedia applications

Synchronization and Sharing Data

• Load-/store- Exclusive instructions (LDREX/STREX) support semaphores– Consolidates old Swap instruction and

necessary semaphore implementation

• Virtual Memory System Architecture v6 ID’s separate caches– Cache hierarchy and ordering rules

Bit/Byte Order Support

• E-bit for current endian setting of core– Set/cleared with SETEND instruction

• REV* instructions reverse bytes for unaligned data support– REV – reverses a word– REV16 – reverses both halfwords– REVSH – reverses high order halfword + sign

extend halfword

Exception and Interrupt Improvement

• Imperative for real-time tasks wherein low latency is critical

• F1 bit in CP15 register 1 designates: 0: Max performance mode, or1: Low interrupt latency mode to allow interrupts

• VE bit enables vectored interrupts to core– Direct vs. external-> system -> vector address

• A-bit aborts all unaligned accesses• U-bit (with clear A-bit) allows unaligned hardware

access

Mode Changing and Stack Improvements

• CPSID/CPSIE instructions allow changing between modes with interrupt disable/enable

• Save Return State (SRS) saves registers and state of current mode onto stack of target mode

• Return From Exception (RFE) loads registers and state of saved mode

• Reduces exception handling overhead

8-Stage Pipeline

• Single-issue• Dynamic branch prediction is 64-entry directly

mapped BTB• 64-bit data paths: read 2 registers in 1 clock• Loads/stores done in background• Out-of-order completion: can retire instructions

before execution• ALU processed in parallel with data cache access• MAC processed in lock-step with ALU

Prefetch

L1 memory access requires 2 cycles

Decode

Decode instruction bits and allocate stack

Issue Instruction

Load operands from registers

ALU and MAC

• ALU pipeline– Shift bits– Arithmetic and logical

operations– Save state and registers

• 3-stage MAC – Can issue a 16x16

operation per cycle– Processed with ALU

pipeline

Data Cache Access

• Map memory address• Data cache load/store

requires 2 cycles

Writeback

Write results of instructions to designated memory, cache, or register

8-Stage Pipeline

Diagram by Devereau:7

Power-saving features

• >95% of registers clock gated

• WFI instruction: wait for interrupt: can disable entire clock network

• Reduced clock cycles and use of transistors

Conclusions

• ARM11 will be implemented as a family of cores – Designed for maximum performance in

wireless multimedia – A new standard in efficiency and power for

embedded applications