Global offset table section has been added to the linker
scripts for all targets. This is especially important for
ExecuTorch runtime. This already exists for LLVM based
toolchains but was missing for GNU toolchains.

A minor addition to support ExecuTorch runtime logging is
also included.

Change-Id: Ia3fdf13b26b24163c3405d33c2df532359749836
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>

Adding HDLCD driver for MPS4 platform and with wiring it
up to the newly refactored display component under HAL.

Splitting the CMakeLists under `display` component with
a CMakeLists file now present in MPS3 and MPS4 specific
subdirectories auto-included by the top level but their
targets "excluded from all" by default.

Resolution for MPS4 targets is kept the same as it is for
MPS3 boards to ensure the applications using the display
interface can continue working as they are. However, this
may be updated in future if we wish to use higher
resolution offered by the HDLCD component.

Change-Id: Ib1be3ff861267e8c560429969b5475e22899c4dc
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>

Refactoring the LCD implementation under HAL to remove
unrequired redirection and briging it in line with the
camera and audio interfaces. Renaming `lcd` as a `display`
component to make it generic, in preparation for support
to be added for Arm Corstone-315 and Corstone-320 targets
in later patches.

Change-Id: I7da4d9c5ec91d2aa5ee3af7133575c47c06cae88
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>

The Arm NPU delegate library needs to be built before other
libraries higher up in the chain. Although this dependency
was captured in CMake, using only the target file (and not
the CMake target itself) can cause issues with Makefile
generator. Adding this dependency explicitly fixes the
issue.

Change-Id: Iad5f9bfa01ebfe7413f2473dbbd043a13fbfc8fa
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>

Fix for some third party toolchains failing to compile the
math library without explicitly including the cmath header.

A trivial non-related doumentation fix to say that LLVM or
Clang based toolchains may not work for ExecuTorch either.
Currently, only Arm GNU Embedded Toolchain works with this
flow.

Change-Id: I12889d796886fb0532002c01b1b603982c926e31
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>

Updating ExecuTorch submodule to the latest version. The
Arm AOT compiler within ExecuTorch source tree doesn't need
path to the shared lib as it previously did. All references
to this lib used for PTE file generation have been removed.

Additionally, the linking hierarchy for ExecuTorch libs has
been amended to make it simpler to link to.

Change-Id: I1d80e5a6d16666d5106c57f7d9b94c9d0f4f4b6a
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>

TensorFlow Lite Micro based logging uses the overrides for
generic Arm Cortex-M CPU targets. This patch ensures that
this is the case only when the framework library target is
configured internally. This will allow external library
targets to be provided without mandating the use of generic
Arm Cortex-M target files being included in the build.

Change-Id: I022efc9d884d789e22f970e1f6e5ddb5e12efe4f
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>

* Fixing the name of the setup script
  in the documentation

Signed-off-by: Alex Tawse <alex.tawse@arm.com>
Change-Id: Icf588d708164f8fa3ed816b08f9cc8f279a10353

This change allow external projects to override CMSIS-DSP
and TensorFlow Lite Micro dependencies by providing their
own aliased CMake targets. This enables reuse of the
platform-agnostic modules without requiring the default
dependency configuration.

Change-Id: I46e7c3f2906e2527b5cda2959f0bd5f3f57a47ba
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>

    *** WARNING: Unstable changes ***

This experimental branch introduces a framework abstraction
layer and adds initial support for the ExecuTorch backend
alongside TensorFlow Lite Micro (TFLM).

Highlights:
----------
- Introduced `fwk/` module with backend-agnostic interfaces
  to decouple use-case logic from ML frameworks.
- Isolated TFLM-specific logic into `fwk/tflm`; added
  ExecuTorch based implementation under `fwk/ExecuTorch`.
- Enabled selection of ML framework via `ML_FRAMEWORK_BUILD`
  CMake flag.

ExecuTorch Integration:
----------------------
- Added ExecuTorch model and tensor classes.
- Integrated native (host) inference flow.
- Enabled AOT compilation pipeline via `aot_arm_compiler`
  and `.pte` generation.
- Added portable ops shared library.
- Added data layout conversions (NHWC → NCHW) in common
  API.

Infrastructure & Setup:
----------------------
- Refactored setup scripts with improved structure and
  Pylint compliance.
- Introduced `--parallel` flag to speed up model setup.
- Extended script to generate ExecuTorch models and support
  conditional use-case setup.

Documentation:
-------------
- Updated all use-case documentation to reflect framework
  compatibility.

This lays the groundwork for future extensibility, enabling
use-cases to support multiple ML runtimes with minimal
duplication.

Co-authored-by: Alex Tawse <alex.tawse@arm.com>
Change-Id: Ic69e4d86008036a35d1b47aeb541a2144b84d574
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>

Vivek Choradia authored Jun 09, 2025 and Alex Tawse committed Jun 13, 2025

Updating third party dependencies to 25.05 versions. The full list
can be found here:
https://gitlab.arm.com/artificial-intelligence/ethos-u/ethos-u/-/blob/main/25.05.json



+-------------------+------------------------------------------+
| Dependency        | Commit/Tag                               |
+-------------------+------------------------------------------+
| cmsis-6           | efba702bbdb5610e866a745740fdf68f2d4d4af5 |
| cmsis-nn          | 1438b3328d7f910423a7057f1abcb6d6576325f2 |
| core-driver       | 25.05                                    |
| core-platform     | 25.05                                    |
| cortex-dfp        | 36d0d3b118ab8493865044551d05248a5fe53427 |
| tensorflow        | 9d35a74b03fa313026982abb4d2c4ffe29de88bc |
+-------------------+------------------------------------------+

Vela version used by default has been bumped to 4.3.0.

Dependencies not updated with this change:
* cmsis-dsp: `v1.16.2` is still the latest stable release.
* avh: No major updates since 925049e8.

Change-Id: I21acdff4880f3ea634e398a5bf9ed9adaf36cdf9
Signed-off-by: Vivek Choradia <vivek.choradia@arm.com>
Signed-off-by: Alex Tawse <alex.tawse@arm.com>

* Fixing broken documentation links
* Removing header for removed section

Change-Id: I80b0dea4919c7ce65e8a1ea4340f011d9506f371
Signed-off-by: Alex Tawse <alex.tawse@arm.com>

Adding support for newer LLVM based Arm Toolchain for
Embedded (ATfE). This is experimental at this stage but
has been tested with versions 19.1.5, 20.0.0 and 20.1.0.
For platforms, these changes have been tested on FPGA
implementations AN552 and AN555 and all variants for FVPs
supported.

Major additions are the CMake toolchain file and linker
scripts, along with required plumbing. CMake presets and
documentation have been updated too. Linker scripts now
have toolchain ID and extensions as suffix instead of just
the extension.

With the associated CMake toolchain file changes, support
for semihosting has also been refreshed, making it possible
to build and run semihosted applications on FVPs easily for
all toolchains. To enable easy boot verification, a simple
HAL test has been added. This needs to be built explicitly
and will produce a small application that boots the target
platform and calls HAL initialisation function.

Change-Id: Id114ca45ed53411c2647a2a0f775ba76609068ac
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>

Vivek Choradia authored Apr 15, 2025 and Alex Tawse committed May 09, 2025

* This patch allows user to press a key to continue
  to the next inference controlled by a new
  CMake config option - INTERACTIVE_MODE.
* This is useful when terminal or LCD output needs
  to be inspected before continuing to the
  next inference
* Simplified stdout component and removed redundant
  method

Change-Id: I0d4b65ce82292606a31fa89fe47b7d86aa3c27e7
Signed-off-by: Vivek Choradia <vivek.choradia@arm.com>
Signed-off-by: Alex Tawse <alex.tawse@arm.com>

Vivek Choradia authored Apr 15, 2025 and Alex Tawse committed May 05, 2025

*  Updating FPGA deployment instructions to include macOS® commands.

Change-Id: I52cf648e8f1ec26bcd5a328b616109f2064fa9ef
Signed-off-by: Vivek Choradia <vivek.choradia@arm.com>
Signed-off-by: Alex Tawse <alex.tawse@arm.com>

* Replacing mlplatform.org URLs to GitLab
  equivalents

Signed-off-by: Alex Tawse <alex.tawse@arm.com>
Change-Id: I7c4bb74ff75799d08b2ce2871b389e75ced6cd73

Updated bare-metal toolchain files to always add function
and data sections for both toolchains. This will reduce
runtime code footprint. Again, it was previously disabled
for legacy reasons (timing adapter driver implementation)
that are no longer relevant.

Change-Id: I15aa57d1cadcf2b92a5b9e01bd1d113161c67e84
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>

Addressing couple of issues raised here:

* https://github.com/ARM-software/CMSIS-DSP/issues/242
* #1



CMSIS-DSP was previously using MLEK wrapper to set up the
sources and compilation flags. This was done for legacy
reasons that are no longer relevant. With these changes we
rely completely on the CMSIS-DSP project's CMake config to
avoid unnecessary duplication and maintenance cost.

Compile option passed to Arm(R) Compiler `-ffp-mode=full`
has been dropped as it is no longer required. This has been
replaced with recommended `-ffast-math` option instead.

Minor Jinja2 version bump (Python environment requirement)
is also included.

Change-Id: I826d0b4dd3f248cb4d91fe095717805b8d341467
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>

Updating third party dependencies to 25.02 versions. The
full list can be found here:
https://gitlab.arm.com/artificial-intelligence/ethos-u/ethos-u/-/blob/9c339ddd5369e37ed554f7e39647f4e0967a3e2d/25.02.json

Remote URLs for core-driver and core-platform have been
updated to track these sources from https://gitlab.arm.com/


as mlplatform is a read-only mirror for now and will not
have any new changes.

+--------------+------------------------------------------+
| Dependency   |  Commit/Tag                              |
+--------------+------------------------------------------+
|avh           | 925049e8ab010388e48fab01fc0411cd9bd38899 |
|cmsis-6       | 1a1799c6c58c47e737bb19523483344a25d29895 |
|cmsis-nn      | e096196a0c49f065abc03d943c583cd50de424ba |
|core-driver   | 25.02                                    |
|core-platform | 25.02                                    |
|tensorflow    | ef64591270691022a329cf04ba9e73ecfb15ddb8 |
+--------------+------------------------------------------+

Vela version used by default has been bumped to 4.2.0.

Dependencies not updated with this change:

* cmsis-dsp: `v1.16.2` is still the latest stable release
* cortex-dfp: No major source change to current `c2c70a97`

Change-Id: If5eccbb509e919dd9d101957d0915116a4e036e5
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>

* Arm Virtual Hardware (AVH) AMI is being deprecated so
  we are removing references to this in our documentation
* Minor typo fix in FVP arguments documentation

Change-Id: Id7a7af64d44a4b5295280c5808533040932145c2
Signed-off-by: Alex Tawse <alex.tawse@arm.com>

Minor change to `cmsis_device` target with the `CMSIS_device_header`
definition used instead of `CPU_HEADER_FILE`. This is done to help
with portability as other related repositories use this specific
definition.

Some other changes have been made to address concerns about cyclic
dependency on platform headers from VSI camera interface. Platform
driver target will not add header path to cmsis_device target so the
VSI CMake file doesn't need to know/guess path to platform-drivers
target to get to IRQ definitions it needs.

Change-Id: Ia6b4d22bcdfb1f44886be7afbd00f53fcbc057ee
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>

Fix for recent change to use CMake wrapper for building TensorFlow
Lite Micro - CMSIS-NN kernels were not being included in the build
heavily impacting performance for CPU-only path. This patch has a
minor change to correctly set `CORE_SOFTWARE_ACCELERATOR` CMake
variable to ensure the CMake wrapper actually adds CMSIS-NN to the
build.

Also, with the CMake build flow for CMSIS-NN and TensorFlow Lite
Micro together, the linker scripts have been modified to allow object
file name based memory placement to still work as it was for Makefile
based flow earlier.

Change-Id: Iedd91d4db7bbaa40603df6a7e93e99274be547d6
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>

A minor patch to fix an error being wrongly flagged when the audio
stream indicated completion.

Change-Id: If8e0a1e82322d64fd60aa3351caefeff068c95cc
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>

Updating submodules to pinned 24.11 components

* core-platform:	24.11 tag
* core-driver:		24.11 tag
* cmsis-nn:		22080c68d040c98139e6cb1549473e3149735f4d
* cmsis-dsp:		v1.16.2 tag
* cmsis-6:		8c4dc58928b3347f6aa98b6fb2bf6770f32a72b7
* tensorflow:		f91dd916e5de773fd2150136d025ae2dfd31d48f

Vela compiler (fetched from PyPi) version updated to 4.1.0.

Change-Id: I07d3d92fa70bbfb8551cf0a2c133ee9101795d29
Signed-off-by: Alex Tawse <alex.tawse@arm.com>

For `Sram_Only` memory mode the "weights" RO traffic was previously
wired through the second set of AXI ports and the timing adapter
settings were kept the same as the first set so this traffic also
sees the same bandwidth and lateny envelope. However, it is quite
possible to see slightly better performance if we only use one set
of AXI ports. For Sram_Only memory mode, the Vela compiler schedules
the R/W expecting to use only one set so this change also aligns
with how Vela intends such a model to be deployed on the NPU.

Change-Id: I7b37c3548e2ab00dda55ac1e13a7883f8e588f9f
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>

* Removed conditional check to prohibit Sram_Only
  builds for Arm Ethos-U65 and Arm Ethos-U85
* Updated documentation accordingly

Change-Id: I3a1473494a746409af3e9a727c8105881d701e68
Signed-off-by: Alex Tawse <alex.tawse@arm.com>

Alex Tawse authored Oct 30, 2024 and Kshitij Sisodia committed Nov 05, 2024

* Moved project to use C11 and C++17

Signed-off-by: Alex Tawse <alex.tawse@arm.com>
Change-Id: I18297ae829df7c54d39d482654915e6e9adbaa87

Adding basic Virtual Streaming Interface (VSI) support for HAL camera
interface alongside the default `hal_camera_static_images` library.
Platform drivers for MPS3 and MPS4 targets can enable use of this
library based on a new CMake configuration parameter `FVP_VSI_ENABLED`
By default this is turned OFF.

This will allow all vision-based examples to stream images from camera
connected to a host machine. Although VSI could also be used to read
frames from an input video file on disk (on host), provision for this
has not been added.

Change-Id: I56bd2fa07742170b146cae51e6ff10ee1900a63f
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>

* Moving to Vela 4.0.1
* Done primarily to use prebuilt wheels
  for non-Linux platforms

Signed-off-by: Alex Tawse <alex.tawse@arm.com>
Change-Id: I9f55462a2a48d849ec2b5584e3cb5269842167ab

Adding a HAL camera and audio interface to allow these as standard
APIs used across use cases. This exercise is primarily to pave the
way for adding Virtual Streaming Interface (VSI) support for Arm
Fixed Virtual Platforms (FVPs) in future work. However, this also
makes the application more portable and easier to add support for
physical (camera or microphone) devices, should we choose to do so.

This patch is big as it cuts across use cases and HAL layers with
platform support. The implementation of baked-in sample images and
audio data is quite old in this repository. To wire things up via
HAL for something which was only application level required touching
quite a few files.

General outline of changes:
--------------------------
 * File generation part has been changed so it generates a self
   contained set of C files in a separate directory that provide
   strongly linked functions that HAL camera or audio interface
   relies on.
 * Two new "components" (camera and audio) under HAL. They have
   weak implementations of functions that should be overridden by
   application sources. The reason for doing it this way is different
   files are generated for different use cases but the there is only
   one target we need under HAL - it is built only once but different
   application sources can override and use this module differently.
 * All the use cases changed to move away from using the generated
   files' functions directly. Instead the static baked-in data is now
   accessed via the HAL provided API's.
 * Tests for all use cases changed for same reason as above.

Byproduct of above changes:
--------------------------
 * Apps don't need to be interactive (user input). The application
   level cannot have knowledge of how many baked in samples we have
   or the filenames. This knowledge only lives below HAL.
 * Tests which were validating inferences by index have been removed.

Change-Id: I09964ce87ea5cea5700d141eef199d3d2be6686c
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>

Updating the requirements.txt file to fetch later versions of Python
packages. Also, the default build script was still using Arm MPS3
targets for latest NPU. Minor update to this script to pick Arm MPS4
based target instead.

Change-Id: I3815188a81ae34ffbc85acdfb80b97c9b96cdb42
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>

Updates to troubleshooting document.

Change-Id: I6ffebb20bb3905ab8ec9124f10f49f939918ca07

Allowing some minor customizations for LCD display by passing some
additional configuration parameters to the `lcd` component under HAL.

 - LCD title display can be turned ON/OFF (default is ON). CMake
   configuration parameter: `LCD_TITLE_EN`.
 - The title string itself can be changed. CMake configuration
   parameter: `LCD_TITLE_STRING`.

Additionally, at application level, ImageUtils.hpp now has some more
colours for background and text colours to choose from. These could
potentially be used to drive a "light" or "dark" theme.

Change-Id: Ia7c7b4f0e818ce02ba5bc94beac0da390efdf985
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>

Moving to the latest NPU driver version. This contains fix for Arm
Ethos-U85 NPU's PMU.

Minor fixes also included for displaying of these counters for the
latest Arm Corstone-320 FVP.

Change-Id: I281866b407cc133898cca97ad3b775931c367e82
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>

Alex Tawse authored Oct 01, 2024 and Alex Tawse committed Oct 02, 2024

* Adds support for Arm Corstone-320 reference system
* Includes support for Arm Ethos-U85 NPU

Change-Id: I049d582379d8feb5e969ada82db12f6747d2fd69
Signed-off-by: Alex Tawse <alex.tawse@arm.com>
Co-authored-by: Codrut-Grigore Irimie <codrut-grigore.irimie@arm.com>

Adding 24.08 version of CMSIS-pack with updated dependencies for
use-case-api.yml. This pack has been tested with examples in
https://github.com/Arm-Examples/mlek-cmsis-pack-examples

.

Change-Id: I75f58ab7fffd46460442746583a0119ffda19ce8
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>

The floating point mode flag `-ffp-mode` was incorrectly set for GNU
toolchain in the TensorFlow Lite Micro CMake file. This patch adds
check for compiler.

Change-Id: I8295d23a4b91f8fec5470ebe036dc358ecc7f595
Signed-off-by: Kshitij Sisodia <kshitij.sisodia@arm.com>