Helpers for CMSSW emulation (fastmachinelearning#1336)

* attempt to add an defines for emulation in CMSSW

* add emulator wrapper

* add some support for io_stream on the hls4ml side

* add pytest for emulation

Author: jmitrevs

hls4ml/backends/vitis/vitis_backend.py

@@ -64,6 +64,7 @@ def create_initial_config(
         namespace=None,
         write_weights_txt=True,
         write_tar=False,
+        write_emulation_constants=False,
         tb_output_stream='both',
         **_,
     ):
@@ -79,6 +80,8 @@ def create_initial_config(
             write_weights_txt (bool, optional): If True, writes weights to .txt files which speeds up compilation.
                 Defaults to True.
             write_tar (bool, optional): If True, compresses the output directory into a .tar.gz file. Defaults to False.
+            write_emulation_constants (bool, optional): If True, write constants to define.h useful for emulation.
+                Defaults to False.
             tb_output_stream (str, optional): Controls where to write the output. Options are 'stdout', 'file' and 'both'.
                 Defaults to 'both'.
 
@@ -97,6 +100,7 @@ def create_initial_config(
             'WriteWeightsTxt': write_weights_txt,
             'WriteTar': write_tar,
             'TBOutputStream': tb_output_stream,
+            'WriteEmulationConstants': write_emulation_constants,
         }
 
         return config

hls4ml/templates/vivado/firmware/defines.h

@@ -4,8 +4,10 @@
 #include "ap_fixed.h"
 #include "ap_int.h"
 #include "nnet_utils/nnet_types.h"
+#include <array>
 #include <cstddef>
 #include <cstdio>
+#include <tuple>
 // hls-fpga-machine-learning insert headers
 
 // hls-fpga-machine-learning insert namespace-start
@@ -14,6 +16,8 @@
 
 // hls-fpga-machine-learning insert layer-precision
 
+// hls-fpga-machine-learning insert emulator-defines
+
 // hls-fpga-machine-learning insert namespace-end
 
 #endif

hls4ml/templates/vivado/firmware/myproject.h

@@ -14,6 +14,8 @@ void myproject(
     // hls-fpga-machine-learning insert header
 );
 
+// hls-fpga-machine-learning insert emulator-defines
+
 // hls-fpga-machine-learning insert namespace-end
 
 #endif

hls4ml/writer/vivado_writer.py

@@ -322,6 +322,34 @@ def write_project_header(self, model):
                 if namespace is not None:
                     newline += '}\n'
 
+            elif '// hls-fpga-machine-learning insert emulator-defines' in line:
+                newline = line
+
+                if model.config.get_writer_config().get('WriteEmulationConstants', False):
+                    brams_def_str = ', '.join([b.definition_cpp(as_reference=False) for b in model_brams])
+                    brams_call_str = ', '.join([b.name for b in model_brams])
+
+                    if model.config.get_config_value('IOType') == 'io_stream':
+                        input_call_str = ', '.join([f'std::get<{n}>(inputs)' for n in range(len(model_inputs))])
+                        output_call_str = ', '.join([f'std::get<{n}>(outputs)' for n in range(len(model_outputs))])
+                    else:
+                        input_call_str = ', '.join([f'std::get<{n}>(inputs).data()' for n in range(len(model_inputs))])
+                        output_call_str = ', '.join([f'std::get<{n}>(outputs).data()' for n in range(len(model_outputs))])
+
+                    newline += (
+                        f'\ninline void {model.config.get_project_name()}_emulator('
+                        'inputs_t& inputs, outputs_t& outputs'  # the inputs_t should ideally be const
+                    )
+                    if len(model_brams) > 0:
+                        newline += ',\n' + brams_def_str
+                    newline += ') {\n'
+                    newline += indent + model.config.get_project_name() + '(\n'
+                    newline += indent + indent + input_call_str + ',\n'
+                    newline += indent + indent + output_call_str
+                    if len(model_brams) > 0:
+                        newline += ',\n' + indent + indent + brams_call_str
+                    newline += '\n' + indent + ');\n}\n'
+
             else:
                 newline = line
             fout.write(newline)
@@ -385,6 +413,20 @@ def write_defines(self, model):
                 if namespace is not None:
                     newline += '}\n'
 
+            elif '// hls-fpga-machine-learning insert emulator-defines' in line:
+                newline = line
+
+                if model.config.get_writer_config().get('WriteEmulationConstants', False):
+                    if model.config.get_config_value('IOType') == 'io_stream':
+                        input_types = [f'hls::stream<{v.type.name}>' for v in model.get_input_variables()]
+                        output_types = [f'hls::stream<{v.type.name}>' for v in model.get_output_variables()]
+                    else:
+                        input_types = [f'std::array<{v.type.name}, {v.size_cpp()}>' for v in model.get_input_variables()]
+                        output_types = [f'std::array<{v.type.name}, {v.size_cpp()}>' for v in model.get_output_variables()]
+                    input_types_str = ', '.join(input_types)
+                    output_types_str = ', '.join(output_types)
+                    newline += '\n' + f'using inputs_t = std::tuple<{input_types_str}>;'
+                    newline += '\n' + f'using outputs_t = std::tuple<{output_types_str}>;\n'
             else:
                 newline = line
             fout.write(newline)

test/pytest/test_writer_config.py

@@ -33,6 +33,18 @@ def test_namespace(keras_model, namespace, io_type, backend):
     hls_model.compile()  # It's enough that the model compiles
 
 
+@pytest.mark.parametrize('io_type', ['io_stream', 'io_parallel'])
+@pytest.mark.parametrize('backend', ['Vitis'])  # Only Vitis is supported
+def test_emulator(keras_model, io_type, backend):
+
+    config = hls4ml.utils.config_from_keras_model(keras_model, granularity='name', backend=backend)
+    odir = str(test_root_path / f'hls4mlprj_emulation_{backend}_{io_type}')
+    hls_model = hls4ml.converters.convert_from_keras_model(
+        keras_model, hls_config=config, io_type=io_type, output_dir=odir, write_emulation_constants=True, backend=backend
+    )
+    hls_model.compile()  # It's enough that the model compiles
+
+
 @pytest.mark.parametrize('backend', ['Vivado', 'Vitis'])  # No Quartus for now
 @pytest.mark.parametrize('write_tar', [True, False])
 def test_write_tar(keras_model, write_tar, backend):