Create the Alif E8 CMSIS project

Duplicate an existing Blinky project

Start by copying a working Blinky project as a template, and renaming the project file inside the new directory:

    

        
        

cd ~/mnist_alif/alif_vscode-template
cp -R blinky/ mnist_executorch
mv mnist_executorch/blinky.cproject.yml mnist_executorch/mnist_executorch.cproject.yml
  

    

    

        
        

cd ~\mnist_alif\alif_vscode-template
Copy-Item .\blinky .\mnist_executorch -Recurse
Rename-Item .\mnist_executorch\blinky.cproject.yml mnist_executorch.cproject.yml
  

    

Next, replace all internal references from blinky to mnist_executorch:

    

        
        

perl -pi -e 's/\bblinky\b/mnist_executorch/g' $(grep -RIl "blinky" mnist_executorch)
  

    

    

        
        

Get-ChildItem .\mnist_executorch -Recurse -File | Where-Object { Select-String -Path $_.FullName -Pattern '\bblinky\b' -Quiet } | ForEach-Object { $text = Get-Content $_.FullName -Raw; $text = $text -replace '\bblinky\b', 'mnist_executorch'; Set-Content $_.FullName $text -NoNewline }
  

    

Rename main.c to main.cpp

ExecuTorch is a C++ library, so the source file needs a .cpp extension.

Update the extension of main.c:

    

        
        

mv mnist_executorch/main.c mnist_executorch/main.cpp
  

    

    

        
        

Rename-Item .\mnist_executorch\main.c main.cpp
  

    

Copy model assets

Create an assets directory and copy the model header into the project:

    

        
        

mkdir -p mnist_executorch/assets
cp ~/mnist_alif/executorch-alif/output/mnist_model_data.h mnist_executorch/assets/
  

    

    

        
        

New-Item -ItemType Directory -Force -Path .\mnist_executorch\assets
Copy-Item ~\mnist_alif\executorch-alif\output\mnist_model_data.h .\mnist_executorch\assets\
  

    

Create the SEGGER RTT configuration

Real-Time Transfer (RTT) works through the J-Link debug probe, reading and writing a memory buffer through the debug interface. It’s faster than UART and doesn’t require extra wiring.

Within mnist_executorch/, create the configuration file SEGGER_RTT_Conf.h:

    

        
        
#ifndef SEGGER_RTT_CONF_H
#define SEGGER_RTT_CONF_H
#define SEGGER_RTT_MAX_NUM_UP_BUFFERS     (1)
#define SEGGER_RTT_MAX_NUM_DOWN_BUFFERS   (1)
#define SEGGER_RTT_BUFFER_SIZE_UP         (4096)
#define SEGGER_RTT_BUFFER_SIZE_DOWN       (16)
#define SEGGER_RTT_MODE_DEFAULT           SEGGER_RTT_MODE_NO_BLOCK_SKIP
#define SEGGER_RTT_PRINTF_BUFFER_SIZE     (256)
#endif

    

Update the solution file

The alif.csolution.yml file describes the whole CMSIS workspace: available projects, targets, devices, packs, and flash settings.

Open the file and make the following changes.

Update the created-for field to match your CMSIS Toolbox version:

    

        
        
  created-for: CMSIS-Toolbox@2.12.0

    

Add the required packs under the packs: section:

    

        
        
  packs:
    - pack: AlifSemiconductor::Ensemble@2.0.4
    - pack: ARM::CMSIS@6.0.0
    - pack: ARM::CMSIS-Compiler@2.1.0
    - pack: Keil::MDK-Middleware@8.2.0
    - pack: ARM::ethos-u-core-driver

    

Add a target-set block to the E8-HP target type so the Security Toolkit knows which binary to flash. Find the type: E8-HP section and add the following:

    

        
        
    - type: E8-HP
      device: Alif Semiconductor::AE822FA0E5597LS0:M55_HP
      board: Alif Semiconductor::DevKit-E8
      define:
        - "CORE_M55_HP"
      target-set:
        - set:
          images:
            - project-context: mnist_executorch.debug

    

For all other target types (E7-HE, E7-HP, E1C-HE, E8-HE), add a target-set pointing to blinky.debug instead.

Add mnist_executorch to the projects list at the end of the file:

    

        
        
  projects:
    - project: blinky/blinky.cproject.yml
    - project: hello/hello.cproject.yml
    - project: hello_rtt/hello_rtt.cproject.yml
    - project: mnist_executorch/mnist_executorch.cproject.yml

    

Configure the project file

The mnist_executorch/mnist_executorch.cproject.yml file describes the MNIST application: its source files, components, include paths, and build options.

Locate this file and replace the contents with the following configuration:

    

        
        
# yaml-language-server: $schema=https://raw.githubusercontent.com/Open-CMSIS-Pack/devtools/tools/projmgr/2.6.0/tools/projmgr/schemas/cproject.schema.json
project:
  groups:
    - group: App
      files:
        - file: main.cpp
        - file: executorch_runner.cpp
        - file: executorch_runner.h
        - file: SEGGER_RTT_Conf.h
        - file: assets/mnist_model_data.h

    - group: SEGGER_RTT
      files:
        - file: ../libs/SEGGER_RTT_V796h/RTT/SEGGER_RTT.c
        - file: ../libs/SEGGER_RTT_V796h/RTT/SEGGER_RTT_printf.c
        - file: ../libs/SEGGER_RTT_V796h/Syscalls/SEGGER_RTT_Syscalls_GCC.c

  output:
    base-name: $Project$
    type:
      - elf
      - bin

  layers:
    - layer: ../device/ensemble/alif-ensemble.clayer.yml

  packs:
    - pack: AlifSemiconductor::Ensemble
    - pack: ARM::ethos-u-core-driver

  components:
    - component: ARM::Machine Learning:NPU Support:Ethos-U Driver&Generic U85

  define:
    - C10_USING_CUSTOM_GENERATED_MACROS
    - ETHOSU
    - ETHOSU85

  add-path:
    - .
    - assets
    - ../libs/SEGGER_RTT_V796h/RTT
    - ../third_party/executorch/et_bundle/include
    - ../third_party/executorch/et_bundle/include/executorch/runtime/core/portable_type/c10

  misc:
    - for-compiler: GCC
      C-CPP:
        - -std=c++17
        - -fno-rtti
        - -fno-exceptions
        - -DETHOSU_ARCH=u85
        - -DETHOSU_MACS=256
        - -DETHOSU_MODEL=1
      Link:
        - -L/absolute/path/to/third_party/executorch/et_bundle/lib
        - -Wl,--whole-archive
        - -lexecutorch
        - -lexecutorch_core
        - -lexecutorch_delegate_ethos_u
        - -lcortex_m_ops_lib
        - -Wl,--no-whole-archive
        - -Wl,--start-group
        - -lextension_runner_util
        - -lcortex_m_kernels
        - -lportable_ops_lib
        - -lportable_kernels
        - -lquantized_ops_lib
        - -lquantized_kernels
        - -lkernels_util_all_deps
        - -lcmsis-nn
        - -lflatccrt
        - -Wl,--end-group

    

There are several important details in this configuration:

• --whole-archive is required for libexecutorch, libexecutorch_core, libexecutorch_delegate_ethos_u, and libcortex_m_ops_lib. These libraries contain static registration constructors — for operator registration and PAL symbols — that the linker would otherwise discard as unused.
• Don’t add portable_ops_lib or quantized_ops_lib to --whole-archive. They are large and will overflow the microcontroller’s ITCM or MRAM.
• --start-group and --end-group resolve circular dependencies among the remaining libraries.
• C10_USING_CUSTOM_GENERATED_MACROS tells ExecuTorch to skip looking for a cmake_macros.h header that doesn’t exist in the bare-metal build.
• The c10 include path provides the tensor type definitions that ExecuTorch’s headers depend on.

Add the source files

The application code is split across three files:

• main.cpp is the firmware entry point. It initializes the board, loads the embedded model from mnist_model_data.h, passes the MNIST input image to the runner, and prints the predicted digit and output scores.
• executorch_runner.h declares the small C-style interface used by main.cpp to initialize ExecuTorch and run inference.
• executorch_runner.cpp implements the ExecuTorch integration. It initializes the Ethos-U NPU driver and loads the .pte program. It also prepares input and output tensors, allocates runtime buffers, and executes the model.

Run these commands to download the source files and place them in mnist_executorch/:

    

        
        
cd ~/mnist_alif/alif_vscode-template/mnist_executorch
curl -o main.cpp https://raw.githubusercontent.com/arm-education/alif-ethos-u85-npu-mnist/refs/heads/main/mnist_executorch/main.cpp
curl -o executorch_runner.cpp  https://raw.githubusercontent.com/arm-education/alif-ethos-u85-npu-mnist/refs/heads/main/mnist_executorch/executorch_runner.cpp
curl -o executorch_runner.h https://raw.githubusercontent.com/arm-education/alif-ethos-u85-npu-mnist/refs/heads/main/mnist_executorch/executorch_runner.h

    

After downloading the files, open them in VS Code. Briefly review how main.cpp calls into executorch_runner.cpp to initialize ExecuTorch, pass in the MNIST input tensor, and read back the model output.

Update the linker script

Open device/ensemble/RTE/Device/AE822FA0E5597LS0_M55_HP/linker_gnu_mram.ld.src.

On Cortex-M systems, the Data Tightly Coupled Memory (DTCM) is a small, fast RAM region connected closely to the CPU. By default, the linker script sends most .bss data to DTCM.

ExecuTorch uses large memory pools, so this data must be routed to SRAM0 and SRAM1 instead. Without this change, the linker reports that .bss does not fit in DTCM.

Make the following changes.

Add SRAM1 to the zero-initialization table

Find the following block in .zero.table:

    

        
        
#if __HAS_BULK_SRAM
    LONG (ADDR(.bss.at_sram0))
    LONG (SIZEOF(.bss.at_sram0)/4)
#endif

    

Replace it with the following:

    

        
        
#if __HAS_BULK_SRAM
    LONG (ADDR(.bss.at_sram0))
    LONG (SIZEOF(.bss.at_sram0)/4)
    LONG (ADDR(.bss.at_sram1))
    LONG (SIZEOF(.bss.at_sram1)/4)
#endif

    

This block tells the startup code to clear both SRAM0 and SRAM1 .bss sections before main() runs.

Add GOT sections to .data.at_dtcm

Find the following block in .data.at_dtcm:

    

        
        
    KEEP(*(.jcr*))
    . = ALIGN(8);

    

Replace it with the following:

    

        
        
    KEEP(*(.jcr*))
    *(.got)
    *(.got.plt)
    . = ALIGN(8);

    

The ExecuTorch C++ libraries can use a GOT (Global Offset Table). These entries must be copied into RAM at startup.

Route ExecuTorch buffers to SRAM0 and SRAM1

Find the following existing block:

    

        
        
#if __HAS_BULK_SRAM
  .bss.at_sram0 (NOLOAD) : ALIGN(8)
  {
    *(.bss.lcd_crop_and_interpolate_buf)
    *(.bss.lcd_frame_buf)
    *(.bss.camera_frame_buf)
    *(.bss.camera_frame_bayer_to_rgb_buf)
  } > SRAM0
#endif

    

Replace it with the following:

    

        
        
#if __HAS_BULK_SRAM
  .bss.at_sram0 (NOLOAD) : ALIGN(8)
  {
    *(.bss.lcd_crop_and_interpolate_buf)
    *(.bss.lcd_frame_buf)
    *(.bss.camera_frame_buf)
    *(.bss.camera_frame_bayer_to_rgb_buf)
    *(.bss.at_sram0)
    *(.bss.at_sram0.*)
  } > SRAM0
  .bss.at_sram1 (NOLOAD) : ALIGN(8)
  {
    *(.bss.at_sram1)
    *(.bss.at_sram1.*)
  } > SRAM1
#endif

    

This block must appear before the generic .bss section. Otherwise, sections such as .bss.at_sram0 and .bss.at_sram1 are caught by *(.bss.*) and placed in DTCM.

The project structure is now ready.

What you’ve accomplished and what’s next

You’ve now created the mnist_executorch firmware project, configured CMSIS packs, and set up the linker to include ExecuTorch libraries with the correct archive flags.

You’ve also added the application code that loads the model and runs inference on the Ethos-U85 NPU.

Next, you’ll load and prepare a test image to copy into the firmware project.