Running a Plan on an Emulated Cortex-M55

QEMU emulates an Arm Cortex-M55 well enough to run a TiGrIS plan as bare-metal firmware and report what it used. That is enough to answer the question a deployment usually starts with, which is whether a model fits a part at all, and it answers it before any hardware exists.

What it cannot answer is how fast the model runs. QEMU’s Cortex-M55 is functional, not cycle-accurate, so treat anything it says about time as meaningless. Use it for correctness and memory placement, and a board for latency.

What you need

• qemu-system-arm 8.2 or later, which carries the mps3-an547 machine, an Arm Corstone-300 with a Cortex-M55.
• arm-none-eabi-gcc 13.2 or later.
• A compiled .tgrs plan.

Check the machine is there before anything else:

qemu-system-arm -machine help | grep an547

mps3-an547           ARM MPS3 with AN547 FPGA image for Cortex-M55

The memory map is the point

The interesting decision is where each tier lives, because that is what decides whether a model fits. The AN547 gives you on-chip SRAM and an external DDR window, and a plan’s two arenas can go in either:

Region	Address	Stands for
DTCM	0x20000000	tightly coupled memory, 512 KiB
SRAM	0x21000000	on-chip SRAM, 4 MiB
DDR	0x60000000	external memory

Putting the plan itself at 0x60000000 and reading it in place models execute-in-place from external flash, so the weights never occupy SRAM. Put the fast arena in on-chip SRAM always. The slow tier is the question: on-chip if the model’s cross-stage tensors are small enough, external if they are not.

QEMU loads the plan to that address without any firmware involvement:

qemu-system-arm -M mps3-an547 -cpu cortex-m55 -nographic \
  -semihosting-config enable=on,target=native \
  -kernel firmware.elf \
  -device loader,file=model.tgrs,addr=0x60000000

Semihosting is what carries printf back to your terminal, so the firmware needs no UART.

A ready-made harness

tigris-bench carries a working firmware under cortex-m/m55-qemu/: a main.c that loads the plan from 0x60000000, allocates both arenas from linker-placed buffers, runs one inference and prints what it used. Two scripts wrap it, one to build against the runtime sources and one to run:

cd cortex-m/m55-qemu
./build.sh path/to/model.tgrs "My model" -DFAST_KB=192 -DSLOW_KB=256
./run.sh   path/to/model.tgrs

-DFAST_KB and -DSLOW_KB size the two arenas in on-chip SRAM, and -DSLOW_DDR moves the slow tier to external memory instead, for a model whose cross-stage tensors will not fit on-chip.

Build for cortex-m55+nomve. The portable int8 kernels do not use Helium, and leaving MVE out of the build keeps what you measure honest about which kernels ran.

Reading the result

A DS-CNN keyword spotter, small enough to be a single stage:

BENCH_START DS-CNN @ Cortex-M55 (QEMU mps3-an547)
Ops: 12  Stages: 1  Tensors: 13  Budget: 20480 bytes
OUTPUT n=12 checksum=10244 min=-128 max=127 nonmin=9
ARENA fast_peak=16496 slow_used=16 total_sram=16512 bytes (fast_cap=20480 slow_cap=65536)

fast_peak is the high-water mark the runtime actually reached, not the capacity you gave it. It is the number to carry into a hardware decision: this model needs 16.5 KiB of SRAM, so it fits a part with 32 KiB and leaves room for everything else the firmware does.

The same harness on MobileNetV2-0.35 at 224x224 shows why tiling is worth having:

BENCH_START MobileNetV2-0.35 @ Cortex-M55 (QEMU mps3-an547)
Ops: 65  Stages: 17  Tensors: 66  Budget: 131072 bytes
OUTPUT n=10 checksum=4869 min=-128 max=127 nonmin=6
ARENA fast_peak=130176 slow_used=16 total_sram=130192 bytes (fast_cap=131072 slow_cap=262144)

That model’s unscheduled peak is 735.00 KiB, with a single 1x48x112x112 activation accounting for 588.00 KiB of it. Seventeen stages bring the working set down to 130,192 bytes, on-chip, with the slow tier barely touched. A runtime that allocates one contiguous arena needs the full 735 KiB and does not run here at all.

Checking the numbers mean something

The emulator will happily run a plan that computes nonsense, so compare the output against a host reference on the same input rather than trusting that it ran. The firmware prints a checksum and the range of the output vector, which is enough to catch the two failures worth catching: a degenerate output, where every element is identical, and a mismatch against the host.

Do not treat a checksum printed in someone else’s README as the target. It is a property of the exact plan bytes that produced it, and a plan rebuilt from a model with a freshly initialized head will give a different one while being equally correct. Generate the reference from the plan you are running.