CoreMark¶

EEMBC CoreMark benchmark for Ambiq Apollo parts, packaged as a portable nsx application.

The example runs the standard PERFORMANCE_RUN with ITERATIONS=0 (auto-calibrate to ≥10 s), prints the score, and then spins. It uses the board-default linker script and the SoC's instruction cache — no custom memory placement or micro-optimization quirks — so the same source builds unmodified across Apollo5 and Apollo4 targets.

Energy/power measurement is not part of this example. See the separate power_benchmark example for the Joulescope-based power workflow.

It is a multi-target example: a single lean nsx.yml declares a targets: block, and the resolved closure for every board is recorded in one combined nsx.lock.

Build & Run¶

# Default board (apollo510_evb)
nsx configure --app-dir .
nsx build     --app-dir .
nsx flash     --app-dir .      # requires JLink + EVB

# Other targets
nsx build     --app-dir . --board apollo510b_evb
nsx build     --app-dir . --board apollo4p_blue_kxr_evb

Score Output (SEGGER RTT)¶

The score is printed over SEGGER RTT channel 0, not SWO/ITM. RTT writes to an in-SRAM ring buffer that the J-Link drains over SWD via background memory reads — no peripheral pins and no sensitivity to clock/baud setup. On Cortex-M55 targets the port cleans the D-cache after each write so the host sees fresh bytes; Cortex-M4 targets have no core D-cache and skip it.

Use any RTT viewer (e.g. JLinkRTTViewer) to read the output, which looks like:

--- CoreMark on Ambiq NSX ---
2K performance run parameters for coremark.
...
CoreMark 1.0 : <score> / GCC ... / STACK
--- CoreMark complete. ---

A minimal scripted capture helper is included:

python tools/rtt_capture.py --rtt-addr 0x<addr> --duration 25 --out cm.log

(--rtt-addr is the address of the _SEGGER_RTT control block, available from the linked coremark ELF via nm/readelf.)

How It Works¶

1. Init (portable_init): nsx_system_init() brings up the SoC at NSX_PERF_HIGH with the instruction/data caches enabled, sets up the RTT control block, and starts a microsecond timer (nsx-timer).
2. Benchmark (core_main): standard EEMBC CoreMark PERFORMANCE_RUN, timed with nsx_timer_us_read().
3. Report (portable_fini): prints the completion banner and spins in __WFI() so the score stays readable.

Cleaning Up¶

Nothing under build/, modules/, or .nsx/ is source-controlled — it is all re-created by nsx configure/nsx build.

nsx clean --app-dir .                 # ninja clean inside the active build dir
nsx clean --app-dir . --full          # delete the active build directory
nsx clean --app-dir . --reset         # full reset before `git pull`
nsx clean --app-dir . --reset --force # also discard local edits under modules/

Project Layout¶

coremark/
├── CMakeLists.txt          App build — CoreMark sources + NSX modules
├── nsx.yml                 Lean multi-target manifest (targets + requires)
├── nsx.lock                Combined resolved module locks (targets: map)
├── boards/                 Vendored board definitions (one dir per target)
├── cmake/nsx/              NSX CMake support (toolchains, modules, helpers)
├── modules/                NSX module sources (app-local, gitignored)
├── src/
│   ├── core_portme.c       Portable platform port (init, timer, RTT output)
│   ├── core_portme.h       Port configuration (types, seeds, timer)
│   └── coremark/           Upstream EEMBC CoreMark (Apache-2.0)
└── tools/
    └── rtt_capture.py      Optional scripted RTT score capture (pylink)

License¶

• Platform port (src/core_portme.c, src/core_portme.h): Apache-2.0
• EEMBC CoreMark (src/coremark/): Apache-2.0