apache · xiaoxiang781216 · Sep 3, 2026 · Sep 3, 2026
@@ -26,7 +26,7 @@ These can be found in the `apps <https://github.com/apache/nuttx-apps>`_ reposit
    logging/index.rst
    lte/index.rst
    math/index.rst
-   mlearing/index.rst
+   mlearning/index.rst
    netutils/index.rst
    nsh/index.rst
    sdr/index.rst

diff --git a/Documentation/applications/mlearing/cmsis/index.rst b/Documentation/applications/mlearing/cmsis/index.rst
diff --git a/Documentation/applications/mlearing/darknet/index.rst b/Documentation/applications/mlearing/darknet/index.rst
diff --git a/Documentation/applications/mlearing/index.rst b/Documentation/applications/mlearing/index.rst
diff --git a/Documentation/applications/mlearing/libnnablart/index.rst b/Documentation/applications/mlearing/libnnablart/index.rst
diff --git a/Documentation/applications/mlearning/cmsis-nn/index.rst b/Documentation/applications/mlearning/cmsis-nn/index.rst
@@ -0,0 +1,36 @@
+=============================
+``cmsis-nn`` CMSIS-NN Library
+=============================
+
+``apps/mlearning/cmsis-nn`` builds the standalone `CMSIS-NN
+<https://github.com/ARM-software/CMSIS-NN>`_ kernels. This is the package
+TensorFlow Lite Micro uses when ``CONFIG_MLEARNING_CMSIS_NN`` is enabled.
+
+It is distinct from :doc:`../cmsis/index`, which vendors the older CMSIS 5
+tree (DSP plus NN) for NNABLA.
+
+Enable with ``CONFIG_MLEARNING_CMSIS_NN``. The Makefile downloads a pinned
+CMSIS-NN commit and compiles:
+
+- Activation functions
+- Basic math
+- Concatenation
+- Convolution
+- Fully connected
+- LSTM
+- NN support functions
+- Pooling
+- Reshape
+- Softmax
+- SVD
+
+When ``CONFIG_ARM_NEON`` is also set, a few kernels are omitted here
+because TFLM supplies NEON replacements under
+``apps/mlearning/tflite-micro/operators/neon``.
+
+Headers are added to the include path from
+``apps/mlearning/cmsis-nn/cmsis-nn/Include``.
+
+For TFLM, enabling this option defines ``CMSIS_NN`` and swaps TFLM
+reference kernels for the CMSIS-NN variants. See
+:doc:`../tflite-micro/index`.
diff --git a/Documentation/applications/mlearning/cmsis/index.rst b/Documentation/applications/mlearning/cmsis/index.rst
@@ -0,0 +1,33 @@
+=========================
+``cmsis`` CMSIS Libraries
+=========================
+
+``apps/mlearning/cmsis`` fetches `CMSIS 5
+<https://github.com/ARM-software/CMSIS_5>`_ and builds CMSIS-DSP and
+CMSIS-NN from that tree. Default version is ``5.8.0``
+(``CONFIG_CMSIS_VER``).
+
+This package is aimed at NNABLA and generic DSP use. For TensorFlow Lite
+Micro kernel acceleration, use :doc:`../cmsis-nn/index` instead.
+
+Enable ``CONFIG_CMSIS`` from :menuselection:`Application Configuration -->
+Machine Learning Support --> CMSIS Libraries`.
+
+``CONFIG_CMSIS_DSP``
+  CMSIS-DSP. Optional ``CONFIG_CMSIS_DSP_ARM_MATH_MATRIX_CHECK`` and
+  ``CONFIG_CMSIS_DSP_ARM_MATH_ROUNDING`` map to ``ARM_MATH_MATRIX_CHECK``
+  and ``ARM_MATH_ROUNDING``. ``CONFIG_ARCH_FPU`` defines
+  ``__FPU_PRESENT``.
+
+``CONFIG_CMSIS_NN``
+  CMSIS-NN from the CMSIS 5 tree. Depends on ``CONFIG_CMSIS_DSP``. A
+  NuttX patch (``cmsis-nn-support_nnabla.patch``) adds CHW convolution
+  helpers used by NNABLA.
+
+Public headers are added from:
+
+- ``CMSIS_5/CMSIS/Core/Include``
+- ``CMSIS_5/CMSIS/DSP/Include`` (when DSP is enabled)
+- ``CMSIS_5/CMSIS/NN/Include`` (when NN is enabled)
+
+See https://arm-software.github.io/CMSIS_5/General/html/index.html.
diff --git a/Documentation/applications/mlearning/darknet/index.rst b/Documentation/applications/mlearning/darknet/index.rst
@@ -0,0 +1,18 @@
+============================================
+``darknet`` YOLO: Real-Time Object Detection
+============================================
+
+``apps/mlearning/darknet`` vendors `Darknet
+<https://github.com/pjreddie/darknet>`_, the framework behind YOLO (You
+Only Look Once) object detection.
+
+Enable ``CONFIG_DARKNET_YOLO``. ``CONFIG_DARKNET_YOLO_VER`` is the Git
+branch fetched at build time (default ``master``). The Makefile compiles
+the core Darknet sources (network, parser, convolutional and connected
+layers, YOLO/region/detection layers, RNN/GRU/LSTM, and image helpers).
+
+Include path: ``apps/mlearning/darknet/darknet/include``.
+
+Darknet is a larger, floating-point oriented stack compared with
+TensorFlow Lite Micro. Prefer :doc:`../tflite-micro/index` for
+microcontroller-class inference.
diff --git a/Documentation/applications/mlearning/index.rst b/Documentation/applications/mlearning/index.rst
@@ -0,0 +1,33 @@
+========================
+Machine Learning Support
+========================
+
+NuttX machine learning support lives in the `apps
+<https://github.com/apache/nuttx-apps>`_ repository under
+``apps/mlearning``. These packages wrap upstream TinyML runtimes and
+optimized kernels. The build system fetches the selected upstream sources
+when the corresponding Kconfig option is enabled.
+
+Enable the libraries from::
+
+  Application Configuration  --->
+    Machine Learning Support  --->
+
+The following packages are available:
+
+- :doc:`cmsis/index` — ARM CMSIS 5 DSP and NN libraries
+- :doc:`cmsis-nn/index` — standalone ARM CMSIS-NN kernels for TensorFlow Lite Micro
+- :doc:`darknet/index` — Darknet YOLO object detection
+- :doc:`libnnablart/index` — Sony NNABLA C inference runtime
+- :doc:`tflite-micro/index` — TensorFlow Lite for Microcontrollers
+
+A ready-to-build simulator configuration is ``sim:tflm``. See
+:doc:`/platforms/sim/sim/boards/sim/index`.
+
+.. toctree::
+   :glob:
+   :maxdepth: 1
+   :titlesonly:
+   :caption: Contents
+
+   */index*
diff --git a/Documentation/applications/mlearning/libnnablart/index.rst b/Documentation/applications/mlearning/libnnablart/index.rst
@@ -0,0 +1,23 @@
+========================================
+``libnnablart`` NNABLA Runtime Libraries
+========================================
+
+``apps/mlearning/libnnablart`` builds the Sony `NNABLA C Runtime
+<https://github.com/sony/nnabla-c-runtime>`_ so NuttX applications can run
+networks produced by Neural Network Libraries.
+
+Enable ``CONFIG_NNABLA_RT``. ``CONFIG_NNABLA_RT_VER`` selects the upstream
+tag (default ``1.24.0``). The Makefile downloads that release and compiles
+the runtime plus a subset of operators, including:
+
+- Activations (ReLU, sigmoid, tanh, softmax, ELU, leaky ReLU, Swish, ...)
+- Arithmetic and reduction
+- Affine (float and fixed-point)
+- Convolution, depthwise convolution, deconvolution, and pooling
+- Batch normalization, dropout, reshape, concatenate, and related array ops
+- Binary-weight / binary-connect quantization helpers
+
+Include path: ``apps/mlearning/libnnablart/nnabla-c-runtime/include``.
+
+CMSIS 5 NN kernels with the NNABLA CHW helpers can be enabled together
+via :doc:`../cmsis/index`.
diff --git a/Documentation/applications/mlearning/tflite-micro/index.rst b/Documentation/applications/mlearning/tflite-micro/index.rst
@@ -0,0 +1,180 @@
+======================================
+``tflite-micro`` TensorFlow Lite Micro
+======================================
+
+TensorFlow Lite for Microcontrollers (TFLM) is a C++ interpreter for
+running ``.tflite`` models on memory-constrained targets. NuttX integrates
+it from ``apps/mlearning/tflite-micro``.
+
+The build downloads a pinned TFLM snapshot from
+https://github.com/tensorflow/tflite-micro and applies NuttX patches that
+add INT8-only operator registrations and an ``extern "C"`` entry point for
+the hello-world example.
+
+Dependencies
+============
+
+``CONFIG_TFLITEMICRO`` depends on all of:
+
+- ``CONFIG_SYSTEM_FLATBUFFERS`` — FlatBuffers headers (``apps/system/flatbuffers``)
+- ``CONFIG_MATH_GEMMLOWP`` — gemmlowp (``apps/math/gemmlowp``)
+- ``CONFIG_MATH_KISSFFT`` — KissFFT (``apps/math/kissfft``)
+- ``CONFIG_MATH_RUY`` — Ruy (``apps/math/ruy``)
+
+TFLM is C++, so the configuration also needs C++ support (``CONFIG_HAVE_CXX``
+and a C++ standard library such as ``CONFIG_LIBCXX``).
+
+Optional acceleration:
+
+- ``CONFIG_MLEARNING_CMSIS_NN`` — replace reference kernels with CMSIS-NN
+  implementations. See :doc:`../cmsis-nn/index`.
+- ``CONFIG_ARM_NEON`` — when CMSIS-NN is enabled, a few kernels are taken
+  from ``apps/mlearning/tflite-micro/operators/neon`` instead of CMSIS-NN.
+
+Configuration
+=============
+
+From :menuselection:`Application Configuration --> Machine Learning Support`:
+
+``CONFIG_TFLITEMICRO``
+  Enable the TFLM library.
+
+``CONFIG_TFLITEMICRO_DEBUG``
+  Keep TFLM error strings and print memory-use information
+  (``TF_LITE_SHOW_MEMORY_USE``). When disabled, ``TF_LITE_STRIP_ERROR_STRINGS``
+  is set to reduce code size.
+
+``CONFIG_TFLITEMICRO_SYSLOG``
+  Route TFLM logging through NuttX syslog via ``tflm_syslog.cc``.
+  ``CONFIG_TFLITEMICRO_SYSLOG_LEVEL`` selects the syslog priority
+  (default ``6``, ``LOG_INFO``). See ``include/syslog.h``.
+
+``CONFIG_TFLITEMICRO_TOOL``
+  Build the ``tflm`` NSH command from ``tflm_tool.cc``.
+  ``CONFIG_TFLITEMICRO_TOOL_PRIORITY`` and
+  ``CONFIG_TFLITEMICRO_TOOL_STACKSIZE`` set the task attributes
+  (defaults: priority 100, stack 4096).
+
+``CONFIG_TFLITEMICRO_HELLOWORLD``
+  Build the ``tflm_hello`` example (CMake builds only). The example runs
+  the upstream TFLM hello-world float and INT8 sine models.
+  ``CONFIG_TFLITEMICRO_HELLOWORLD_PRIORITY`` and
+  ``CONFIG_TFLITEMICRO_HELLOWORLD_STACKSIZE`` set the task attributes
+  (defaults: priority 100, stack 4096).
+
+Building
+========
+
+The ``sim:tflm`` defconfig enables TFLM, the ``tflm`` tool, the hello-world
+example, libc++, FlatBuffers, gemmlowp, KissFFT, and Ruy.
+
+Makefile build (produces the ``tflm`` tool; hello-world is CMake-only):
+
+.. code-block:: console
+
+   $ cd nuttx
+   $ make distclean
+   $ ./tools/configure.sh sim:tflm
+   $ make -j$(nproc)
+   $ ./nuttx
+
+CMake build (produces both ``tflm`` and ``tflm_hello``):
+
+.. code-block:: console
+
+   $ cd nuttx
+   $ cmake -B build -DBOARD_CONFIG=sim:tflm -GNinja
+   $ cmake --build build
+   $ ./build/nuttx
+
+The first build downloads the TFLM sources. Subsequent builds reuse the
+unpacked tree under ``apps/mlearning/tflite-micro/tflite-micro``.
+``make distclean`` in the apps tree removes that snapshot.
+
+Using the ``tflm`` tool
+=======================
+
+``tflm`` loads a ``.tflite`` file from the filesystem, constructs a
+``tflite::MicroInterpreter``, and can invoke the model once for profiling
+or emit compiled C++ (when TFLM was built with ``TFLITE_MODEL_COMPILER``).
+
+.. code-block:: console
+
+   nsh> tflm -h
+
+   Utility to use tflite micro on nuttx.
+   [ -C       ] Compile tflite model into c++ codes.
+   [ -E       ] Do once evaluation (for profiling).
+   [ -i <str> ] Readable model file path.
+   [ -o <str> ] Writable c++ file path.
+   [ -p <str> ] Prefix of compiled code.
+   [ -a <int> ] Arena size (mempool).
+   [ -h       ] Print this message.
+
+Both ``-i`` and ``-o`` are required. Defaults are prefix ``NXAI`` and
+arena size 8192 bytes.
+
+The built-in operator resolver registers eight INT8-oriented ops:
+
+- ``CONV_2D`` (INT8)
+- ``MAX_POOL_2D`` (INT8)
+- ``QUANTIZE`` (float32 to INT8)
+- ``DEQUANTIZE`` (INT8)
+- ``MEAN`` (INT8)
+- ``RESHAPE``
+- ``FULLY_CONNECTED`` (INT8)
+- ``SOFTMAX`` (INT8)
+
+Models that need other operators must change the resolver in
+``apps/mlearning/tflite-micro/tflm_tool.cc``.
+
+Hello-world example
+===================
+
+With a CMake ``sim:tflm`` image:
+
+.. code-block:: console
+
+   nsh> tflm_hello
+
+This runs the upstream hello-world test: it profiles memory and latency,
+then loads the float and INT8 sine models that are converted to C arrays
+at build time with ``xxd``.
+
+Embedding a model in an application
+===================================
+
+TFLM is designed for targets without a filesystem and without dynamic
+allocation for the model itself. Typical NuttX applications compile the
+``.tflite`` file into a C array and pass it to ``tflite::GetModel()``.
+
+The CMake helper ``tflite_generate_data()`` in
+``apps/mlearning/tflite-micro/CMakeLists.txt`` wraps ``xxd -i`` for that
+purpose. The same conversion can be done manually::
+
+  xxd -i model.tflite model_data.h
+
+Provide a tensor arena (the ``-a`` size in ``tflm``, or a static buffer
+in application code) large enough for the model's scratch tensors.
+Insufficient arena size makes ``MicroInterpreter::AllocateTensors()``
+fail.
+
+Patches applied by NuttX
+========================
+
+- ``tflite-micro.patch`` — guard the ``ASSERT`` macro in the TFLM signal
+  circular buffer.
+- ``0001-dequantize-int8.patch`` — ``Register_DEQUANTIZE_INT8()``.
+- ``0002-quantize-int8.patch`` — ``Register_QUANTIZE_FLOAT32_INT8()``.
+- ``0003-mean-int8.patch`` — ``Register_MEAN_INT8()``.
+- ``0004-tflite-add-extern-C-to-main-function-to-avoid-c-mang.patch`` —
+  ``extern "C"`` on the hello-world ``main`` so NuttX can call it
+  (applied by the CMake fetch).
+
+See also
+========
+
+- Upstream TFLM: https://github.com/tensorflow/tflite-micro
+- Hello-world example:
+  https://github.com/tensorflow/tflite-micro/blob/main/tensorflow/lite/micro/examples/hello_world/README.md
+- :doc:`../cmsis-nn/index` for ARM kernel acceleration
@@ -1446,6 +1446,31 @@ To use IPv4, modify these settings in the defconfig file:
    -CONFIG_NET_IPv6=y
    -CONFIG_NET_IPv6_NCONF_ENTRIES=4
 
+tflm
+----
+
+Builds TensorFlow Lite for Microcontrollers from
+``apps/mlearning/tflite-micro``, including the ``tflm`` command-line tool.
+A CMake build of this configuration also includes the ``tflm_hello``
+example.
+
+.. code-block:: console
+
+   $ ./tools/configure.sh sim:tflm
+   $ make -j$(nproc)
+   $ ./nuttx
+
+With CMake::
+
+   $ cmake -B build -DBOARD_CONFIG=sim:tflm -GNinja
+   $ cmake --build build
+   $ ./build/nuttx
+
+From NSH, ``tflm -h`` prints the tool usage. ``tflm_hello`` (CMake image)
+runs the upstream sine-model test.
+
+See :doc:`/applications/mlearning/tflite-micro/index`.
+
 touchscreen
 -----------