Add NanoXplore NG-ULTRA accelerator backend

.gitignore

@@ -11,7 +11,7 @@ my-hls-test
 *.tar.gz
 docs/_build
 docs/autodoc/*
-hls4mlprj_*
+test/pytest/*_test_*
 *~
 *.ipynb
 *.ipynb_checkpoints/

.gitmodules

@@ -10,3 +10,9 @@
 [submodule "hls4ml/templates/catapult/ac_math"]
 	path = hls4ml/templates/catapult/ac_math
 	url = https://github.com/hlslibs/ac_math.git
+[submodule "hls4ml/templates/bambu/nnet_utils/gcem"]
+	path = hls4ml/templates/bambu/nnet_utils/gcem
+	url = https://github.com/kthohr/gcem
+[submodule "hls4ml/templates/bambu/ac_types"]
+	path = hls4ml/templates/bambu/ac_types
+	url = https://github.com/ferrandi/ac_types.git

docs/backend/bambu.rst

@@ -0,0 +1,64 @@
+=====
+Bambu
+=====
+
+The ``Bambu`` backend targets `Bambu/PandA <https://panda.dei.polimi.it/>`_, an
+open-source high-level synthesis compiler. It converts ``hls4ml`` models to HLS
+C++ and drives Bambu to synthesizable Verilog, so no proprietary HLS tool is
+required on the critical path. Both ``io_parallel`` and ``io_stream`` are
+supported.
+
+Quick start
+===========
+
+.. code-block:: python
+
+    import hls4ml
+
+    config = hls4ml.utils.config_from_keras_model(model, granularity='name', backend='Bambu')
+
+    hls_model = hls4ml.converters.convert_from_keras_model(
+        model,
+        hls_config=config,
+        backend='Bambu',
+        part='xc7a100tcsg324-1',
+        io_type='io_parallel',
+        output_dir='my_bambu_prj',
+    )
+
+    # Compile the bridge and check numerical accuracy against Keras
+    hls_model.compile()
+    y = hls_model.predict(X)
+
+    # Run Bambu: C-simulation, HLS synthesis, RTL co-simulation and (optionally)
+    # a Vivado logic-synthesis pass for post-route resource/timing/power numbers.
+    hls_model.build(csim=True, synth=True, cosim=True, vsynth=True)
+
+    report = hls4ml.report.parse_bambu_report('my_bambu_prj')
+
+``part`` must be a device Bambu knows about; the supported names are the keys of
+``partname_to_bambu`` in ``hls4ml/backends/bambu/bambu_backend.py``. Requiring
+``vsynth=True`` (and therefore the resource/timing/power numbers) needs Vivado on
+the ``PATH``; the rest of the flow only needs the ``bambu`` executable.
+
+Post-route utilization, timing and power numbers are parsed from the reports
+Bambu's Vivado flow produces, reusing the shared report helpers in
+``hls4ml/report/vivado_report.py``.
+
+Known limitations
+=================
+
+* Large completely-partitioned arrays crash Bambu's frontend, so dense layers
+  must be kept small.
+* The softmax inverse lookup table is emitted as a ``constexpr`` array. When
+  ``fix_softmax_table_size`` shrinks the table (i.e. when
+  ``2 ** min(input_bitwidth, table_bitwidth)`` is below the default table size),
+  Bambu's clang rejects the initializer at compile time. Other softmax
+  configurations work; if you hit this, widen the input/table precision or take
+  the argmax on the host.
+* ``-m64`` combined with ``ac_channel`` crashes ``InterfaceInfer``. The default
+  path avoids this by using the headers Bambu ships.
+
+These were reported to the PandA developers and are addressed by
+`PandA-bambu#396 <https://github.com/ferrandi/PandA-bambu/pull/396>`_; the
+limitations above apply to current Bambu versions until that release lands.

docs/backend/nanoxplore.rst

@@ -0,0 +1,64 @@
+=====================
+NanoXploreAccelerator
+=====================
+
+The **NanoXploreAccelerator** backend builds on the :doc:`Bambu <bambu>` backend
+and turns a Bambu-generated HLS core into a complete accelerator: a float I/O
+wrapper, an AXI4 slave, a PLL configuration, and a versioned ``manifest.json``
+describing the result. It targets NanoXplore's `NG-ULTRA
+<https://www.nanoxplore.com/>`_, a radiation-hardened FPGA with no HLS tool of
+its own.
+
+.. code-block:: python
+
+    hls_model = hls4ml.converters.convert_from_keras_model(
+        model,
+        hls_config=config,
+        backend='NanoXploreAccelerator',
+    )
+    hls_model.build(synth=True, bitstream=True)
+
+Defaults are the NG-ULTRA DevKit's: part ``nx2h540tsc`` and a 20 ns clock
+period, matching the board's 50 MHz oscillator.
+
+A deliberate seam
+=================
+
+``hls4ml`` never imports a vendor tool. The backend writes everything a
+place-and-route flow needs into the project directory -- the complete RTL file
+list, clock, port map and data widths, all in ``manifest.json`` and versioned so
+a mismatch fails loudly -- then shells out to a single command and reads back
+``report.json``:
+
+.. code-block:: text
+
+    hls4ml-nanoxplore-bitstream <project_dir>
+
+The command is configurable through the ``BitStreamCommand`` config value. The
+vendor-specific driver lives out of tree, which means the abstract layer can be
+built and tested with no vendor licence: ``build(synth=True, bitstream=False)``
+produces the wrapper, the RTL and the manifest, and stops before place and
+route.
+
+Structure
+=========
+
+``BambuAcceleratorBackend`` is abstract and unregistered. It provides the
+wrapper generation, the RTL templates, the manifest and the PLL patching, and
+leaves one abstract method, ``_generate_bitstream``. Other FPGA families can
+reuse the layer by subclassing it.
+
+``NanoXploreAcceleratorBackend`` is the registered concrete backend: NG-ULTRA
+defaults plus the CLI call.
+
+Clocking
+========
+
+Any requested ``ClockPeriod`` is turned into a solved ``NX_PLL_U``
+configuration, spliced into the generated top level, so the hardware clock and
+the timing constraint agree by construction rather than by convention. This
+needs OR-Tools:
+
+.. code-block:: bash
+
+    pip install hls4ml[nanoxplore]

docs/index.rst

@@ -38,6 +38,8 @@
     backend/oneapi
     backend/catapult
     backend/quartus
+    backend/bambu
+    backend/nanoxplore
     backend/sr
 
 .. toctree::

hls4ml/backends/__init__.py

@@ -13,6 +13,12 @@
 
 from hls4ml.backends.vitis.vitis_backend import VitisBackend  # isort: skip
 
+from hls4ml.backends.bambu.bambu_backend import BambuBackend  # isort: skip
+
+from hls4ml.backends.bambu_accelerator.bambu_accelerator_backend import BambuAcceleratorBackend  # isort: skip  # noqa: F401
+
+from hls4ml.backends.nanoxplore_accelerator.nanoxplore_accelerator_backend import NanoXploreAcceleratorBackend  # isort: skip  # noqa: E501,F401
+
 
 def _register_builtin_backends():
     register_backend('Vivado', VivadoBackend)
@@ -23,6 +29,8 @@ def _register_builtin_backends():
     register_backend('SymbolicExpression', SymbolicExpressionBackend)
     register_backend('oneAPI', OneAPIBackend)
     register_backend('Libero', LiberoBackend)
+    register_backend('Bambu', BambuBackend)
+    register_backend('NanoXploreAccelerator', NanoXploreAcceleratorBackend)
 
 
 _register_builtin_backends()