Using ARM Caches

Last Author
bains
Last Updated
1015 Days Ago
Subscribers
None

Using ARM Caches

This page describes how to set up the MMU, L1 caches, and L2 cache on the Cortex-A9 MPCore processor found in the Cyclone V.

Introduction

The following documents are useful references:

The ARM processor in the Cyclone V has both L1 and L2 caches.
The L1 cache is split into separate instruction and data caches and is controlled directly by the processor.
The L2 cache is a unified cache and is controlled by the L2C-310 cache controller.

The L1 instruction cache can be enabled using a single bit in the SCTLR register using MRC/MCR instructions.
The L1 data cache can only be used when the memory management unit (MMU) is on.
The L2 cache can be enabled by programming the L2C-310 controller using memory-mapped registers.

To see how the processor performs in various configurations, see the benchmarking results at ARM Benchmark Results.

Enabling the MMU

The MMU translates virtual addresses used by the processor into physical addresses that correspond to actual memory locations.
It also controls the caching behaviour of and access to different sections of the memory space.

Several steps are involved when turning on the MMU:

1. Disable caches and branch predictor
  • Clear I, C, and Z bits in SCTLR using MRC/MCR instructions.
2. Invalidate Everything

It is important to invalidate caches, TLBs, etc. because once the MMU is turned on, and address translation begins, the cache entries, etc. will no longer be valid.

The following steps should be taken:

  • Invalidate instruction, data, and unified TLBs.
  • Invalidate L1 instruction and data caches.
  • Invalidate branch predictor array.
3. Set up translation table entries

The Cortex-A9 MPCore processor allows for two levels of translation tables.
For simplicity only Level 1 translation tables are used.
A flat one-to-one mapping is used where virtual addresses are mapped to the same physical address.
This is done using 1MB 'sections'.
Since the address space is 4GB, this requires 4096 translation table entries.
The L1 translation table must be aligned 16kB aligned in the memory.

  • Use 1MB sections
  • Set TEX[2:0], C, and B bits to use outer and inner write-back for normal memory (SDRAM)
  • Set TEX[2:0], C, and B bits to use non-shareable device memory for memory-mapped peripherals
  • Set AP[1:0] bits to allow read and write access
  • Set the domain to 0 (or whatever)
  • Set the section base address of each table entry to point to the appropriate 1MB section of physical memory
4. Set translation table control registers
  • Set Translation Table Base Control Register (TTBCR) to 0 so that TTBR0 is used
  • Set TTBR0 to point to the L1 translation table, and to use inner and outer write-back, write-allocate cacheable memory for translation table walks
5. Set domain access control register
  • Set DACR to client or master mode for the domain(s) you used in the translation table entries.
6. Enable the MMU
  • Set M bit in SCTLR using MCR/MRC instructions.

Once these steps are complete the L1 caches, and branch predictor can be turned on by setting the I, C, and Z bits in the SCTLR.

Programming the L2 Cache Controller

The L2C-310 cache controller is controlled using memory mapped registers.
For the Cyclone V SoC the base address of these registers is 0xFFFEF000.
The register descriptions can be found in the L2C-310 TRM, linked above.

The following steps are taken to enable the L2 cache controller:

  • Set the way size
  • Set the read, write, and hold delays for Tag RAM
  • Set the read, write, and hold delays for Data RAM
  • Set the prefetching behaviour
  • Invalidate the cache
  • Enable the L2C-310 cache controller

The L2C-310 also includes event counting registers that can be used to monitor hit and miss rates, and events related to speculative reads and prefetching.

NOTE: The I and C bits in the System Control Register (SCTLR) control caching at all levels. If the L2 cache is enabled, but the I and C bits are cleared, the processor cannot take advantage of the L2 cache.

Memory Performance Optimizations

The following additional settings greatly enhance memory performance:

  • Set memory region attributes in TTB entries to use inner write-back
  • Use minimum stable L2C-310 read, write, and hold delays
  • Enable L1 Data-side prefetch

Special L2C-310 + Cortex-A9 MPCore Options

Several other options area available as well.
Both the Cortex-A9 TRM and L2C-310 TRM (linked above) outline several optimizations for L2 memory accesses.
Link.

The following features are available when using the L2C-310 cache controller with a Cortex-A9 MPCore processor:

  • Exclusive caching
  • L2 prefetch hints
  • Early BRESP
  • Allow L2C-310 to write a full line of zeros
  • L2 cache speculative linefill (requires Snoop Control Unit to be enabled)

These features were all tested during benchmarking; however, none of them seemed to offer any performance gains.
Further investigation may may be required to make the best use of these options.

Other Possible Optimizations

There are other options available that may further boost memory performance.
These options have not yet been investigated.

  • L2 cache entry lockdown
  • Cache replacement policy
  • L2 cache preloading

Of these options, L2 cache preloading may offer the greatest benefits.

Source Code

This investigation resulted in two files: arm_cache.h and arm_cache.s that together provide functions to turn on the MMU and caches on the ARM Cortex-A9 MPCore in the Altera Cyclone V SoC.