def test_fake_values_concat(concat_test_graph):
G = create_graph(concat_test_graph, opts={"load_tensors": True})
G.add_dimensions()
G.adjust_order()
matcher = get_std_match_group()
matcher.match(G)
G.add_dimensions()
G.constant_store.fake = True
stats_collector = ActivationStatsCollector()
stats_collector.collect_stats(
G, [np.random.rand(*node.dims.shape) for node in G.inputs()])
astats = stats_collector.reduce_stats()
stats_collector = FilterStatsCollector()
fstats = stats_collector.collect_stats(G)
quantizer = SimpleQuantizer(astats, fstats, force_width=8)
qrecs = quantizer.quantize(G)
G.quantization = qrecs
with tempfile.TemporaryDirectory() as tempdir:
opts = {
'default_input_location': 'ARG_LOC_L2',
'default_output_location': 'ARG_LOC_L2',
'default_global_location': 'ARG_LOC_L3_HFLASH',
'default_local_location': '0',
'at_ver': 3,
'tensor_directory': tempdir
}
code_gen = CodeGenerator(G, DefaultNamingConvension(G), opts)
print(default_template(G, code_generator=code_gen))
code_gen.write_constants()
def generate_code(args):
LOG.propagate = False
nntool_shell = NNToolShell(args, allow_cli_args=False)
# handler = GeneratorShellLogHandler()
# formatter = logging.Formatter('%(module)s - %(message)s')
# handler.setFormatter(formatter)
# LOG.addHandler(handler)
# if args.log_level:
# LOG.setLevel(args.log_level.upper())
# else:
# LOG.setLevel('WARN')
LOG.info("Loading %s", args.graph_file)
nntool_shell.load_state_file(args.graph_file)
G = nntool_shell.G
opts = nntool_shell.settings
if args.model_file:
opts['model_file'] = args.model_file
if args.model_directory:
opts['model_directory'] = args.model_directory
if args.tensor_directory:
opts['tensor_directory'] = args.tensor_directory
opts['basic_kernel_header_file'] = args.model_file
opts['basic_kernel_source_file'] = args.model_file
os.makedirs(os.path.abspath(opts['model_directory']),
mode=0o750,
exist_ok=True)
os.makedirs(os.path.abspath(opts['tensor_directory']),
mode=0o750,
exist_ok=True)
code_gen = CodeGenerator(G, DefaultNamingConvension(G), opts)
if args.template_file:
code_template = dynamic_template(args.template_file)
else:
code_template = default_template
write_template(G, code_gen, opts['model_directory'], opts['model_file'],
code_template, "model")
if G.has_expressions:
write_template(G, code_gen, opts['model_directory'],
opts['basic_kernel_header_file'],
basic_kernel_header_template, "kernel headers")
write_template(G, code_gen, opts['model_directory'],
opts['basic_kernel_source_file'],
basic_kernel_source_template, "kernel source")
if args.header_file:
with open(os.path.join(opts['model_directory'], args.header_file),
"w") as output_fp:
output_fp.write(header_template(G, code_generator=code_gen))
if not args.dont_dump_tensors:
LOG.info("Writing constants to %s", opts['model_directory'])
code_gen.write_constants()
def do_gen(self, args):
"""
Generate AutoTiler model C code and optionally dump tensors. If no destination file is
given the generated code will be outputed to the screen. Check the 'set' command for
settings related to code generation."""
self._check_graph()
self._check_quantized()
self._check_adjusted()
if args.checksums:
input_args = self._get_input_args(None)
LOG.info("input file %s", args.checksums)
data = import_data(args.checksums, **input_args)
executer = GraphExecuter(self.G, qrecs=self.G.quantization)
executer.execute([data], qmode=QuantizationMode.all())
self.settings['checksum_file'] = args.checksums
self.settings['generate_checksums'] = True
if args.tensor_directory:
self.settings['tensor_directory'] = args.tensor_directory
if args.model_directory:
self.settings['model_directory'] = args.model_directory
self.settings['basic_kernel_source_file'] = args.basic_kernel_source_file
self.settings['basic_kernel_header_file'] = args.basic_kernel_header_file
code_gen = CodeGenerator(self.G, DefaultNamingConvension(self.G), self.settings)
if self.settings['template_file']:
code_template = dynamic_template(self.settings['template_file'])
else:
code_template = default_template
if args.model_file:
with open(os.path.join(self.settings['model_directory'],
args.model_file), "w") as output_fp:
output_fp.write(code_template(self.G, code_generator=code_gen))
if self.G.has_expressions:
with open(os.path.join(self.settings['model_directory'],
args.basic_kernel_source_file), "w") as output_fp:
output_fp.write(basic_kernel_source_template(self.G, code_generator=code_gen))
with open(os.path.join(self.settings['model_directory'],
args.basic_kernel_header_file), "w") as output_fp:
output_fp.write(basic_kernel_header_template(self.G, code_generator=code_gen))
else:
self.ppaged(code_template(self.G, code_generator=code_gen))
if self.G.has_expressions:
self.ppaged(basic_kernel_source_template(self.G, code_generator=code_gen))
self.ppaged(basic_kernel_header_template(self.G, code_generator=code_gen))
if args.output_tensors:
code_gen.write_constants()
if args.header_file:
with open(os.path.join(self.settings['model_directory'], args.header_file), "w") as output_fp:
output_fp.write(header_template(self.G, code_generator=code_gen))
def generate_code(args):
LOG.propagate = False
handler = GeneratorShellLogHandler()
formatter = logging.Formatter('%(module)s - %(message)s')
handler.setFormatter(formatter)
LOG.addHandler(handler)
if args.log_level:
LOG.setLevel(args.log_level.upper())
else:
LOG.setLevel('WARN')
LOG.info("Loading %s", args.graph_file)
G, opts = load_state(args.graph_file, return_extra=True)
if args.model_file:
opts['model_file'] = args.model_file
if args.model_directory:
opts['model_directory'] = args.model_directory
if args.tensor_directory:
opts['tensor_directory'] = args.tensor_directory
os.makedirs(os.path.abspath(opts['model_directory']),
mode=0o750,
exist_ok=True)
os.makedirs(os.path.abspath(opts['tensor_directory']),
mode=0o750,
exist_ok=True)
model_path = os.path.join(opts['model_directory'], opts['model_file'])
LOG.info("Saving model to %s", model_path)
code_gen = CodeGenerator(G, DefaultNamingConvension(G), opts)
if args.template_file:
code_template = dynamic_template(args.template_file)
else:
code_template = default_template
model = code_template(G, code_generator=code_gen)
if not model:
LOG.error("error generating model code")
sys.exit(1)
with open(model_path, "w") as output_fp:
output_fp.write(model)
if args.header_file:
with open(os.path.join(opts['model_directory'], args.header_file),
"w") as output_fp:
output_fp.write(header_template(G, code_generator=code_gen))
if not args.dont_dump_tensors:
LOG.info("Writing constants to %s", opts['model_directory'])
code_gen.write_constants()
def test_fused_operational(caplog, mnist_fused_8bit_state):
caplog.set_level(logging.INFO)
G = load_state(mnist_fused_8bit_state)
opts = {
'default_input_location': 'ARG_LOC_L2',
'default_output_location': 'ARG_LOC_L2',
'default_global_location': 'ARG_LOC_L3_HFLASH',
'default_local_location': '0',
}
code_gen = CodeGenerator(G, DefaultNamingConvension(G), opts)
default_template(G, code_generator=code_gen)
def test_tensor_dump(mnist_fused_8bit_state):
with tempfile.TemporaryDirectory() as tempdir:
G = load_state(mnist_fused_8bit_state)
opts = {
'default_input_location': 'ARG_LOC_L2',
'default_output_location': 'ARG_LOC_L2',
'default_global_location': 'ARG_LOC_L3_HFLASH',
'default_local_location': '0',
'tensor_directory': tempdir
}
code_gen = CodeGenerator(G, DefaultNamingConvension(G), opts)
default_template(G, code_generator=code_gen)
code_gen.write_constants()
files_list = [
f for f in os.listdir(tempdir)
if os.path.isfile(os.path.join(tempdir, f))
]
assert set(files_list) == set([
'Step2Weights.tensor', 'Step1Weights.tensor', 'Step1Biases.tensor',
'Step3Weights.tensor', 'Step2Biases.tensor', 'Step3Biases.tensor'
])
def gen_project(G,
settings,
project_folder,
script_commands,
overwrite=False,
performance=False):
code_gen = CodeGenerator(G, DefaultNamingConvension(G), settings)
if not os.path.exists(project_folder):
os.mkdir(project_folder)
main = os.path.join(project_folder, f"{code_gen.project_name}")
main_c = main + '.c'
main_h = main + '.h'
common_mk = os.path.join(project_folder, "common.mk")
nntool_script = os.path.join(project_folder, "nntool_script")
if overwrite or not os.path.exists(main_c):
with open(os.path.join(project_folder, f"{code_gen.project_name}.c"),
"w") as output_fp:
output_fp.write(generate_main_appl_template(G, code_gen))
if overwrite or not os.path.exists(main_h):
with open(os.path.join(project_folder, f"{code_gen.project_name}.h"),
"w") as output_fp:
output_fp.write(generate_main_appl_header(G, code_gen))
if overwrite or not os.path.exists(common_mk):
with open(os.path.join(project_folder, "common.mk"), "w") as output_fp:
output_fp.write(generate_main_appl_make(G, code_gen))
if overwrite or not os.path.exists(nntool_script):
with open(nntool_script, 'w') as fp:
# NOTE - gen_template_project is excluded so that tests work. Normally it will not be in the
# history.
fp.writelines([
f"{line}\n" for line in script_commands
if not (line.startswith('open') or line.startswith(
'gen_project') or line.startswith('performance'))
])
if performance:
for setting in [
'set graph_produce_node_names true',
'set graph_produce_operinfos true',
'set graph_monitor_cycles true'
]:
fp.write(f'{setting}\n')
if script_commands[-1] != "save_state":
fp.write('save_state\n')
ignore_function = None if overwrite else skip_existing_files(
project_folder)
shutil.copytree('generation/project_template',
project_folder,
dirs_exist_ok=True,
ignore=ignore_function)
def test_gen_vergesense(caplog):
caplog.set_level(logging.DEBUG)
tfi = TfliteImporter()
G = tfi.create_graph("tests/graph/marco_17_04.tflite", {
'load_tensors': True,
'load_quantization': True
})
G.add_dimensions()
G.adjust_order()
matcher = get_scale8_match_group()
matcher.match(G)
G.add_dimensions()
with tempfile.TemporaryDirectory() as tempdir:
opts = {
'default_input_location': 'ARG_LOC_L2',
'default_output_location': 'ARG_LOC_L2',
'default_global_location': 'ARG_LOC_L3_HFLASH',
'default_local_location': 'AT_MEM_UNDEF',
'tensor_directory': tempdir
}
code_gen = CodeGenerator(G, DefaultNamingConvension(G), opts)
default_template(G, code_generator=code_gen)
code_gen.write_constants()
def test_activatiofusion(actfusion_graph):
G = actfusion_graph
matcher = get_fusion('scale8_match_group')
matcher.match(G)
G.add_dimensions()
astat_col = ActivationStatsCollector()
astats = astat_col.collect_stats(
G, [np.full([10, 10, 2], 1),
np.full([10, 10, 2], 1)])
astats = astat_col.reduce_stats()
quantizer = MultQuantizer(astats,
force_width=8,
quantized_dimension="channel")
G.quantization = quantizer.quantize(G)
with tempfile.TemporaryDirectory() as tempdir:
opts = {
'default_input_location': 'ARG_LOC_L2',
'default_output_location': 'ARG_LOC_L2',
'default_global_location': 'ARG_LOC_L3_HFLASH',
'default_local_location': 'AT_MEM_UNDEF',
'tensor_directory': tempdir
}
code_gen = CodeGenerator(G, DefaultNamingConvension(G), opts)
ATModel_code = default_template(G, code_generator=code_gen)
def gen_project(G,
settings,
project_folder,
script_commands,
overwrite=False,
performance=False,
quantized=False,
test_results=False,
save_inputs=False,
input_file=None,
input_args=None,
gen_atproject=False,
dump_tensors=False,
input_tensors=None,
tolerance=0.0):
settings = deepcopy(settings)
settings['graph_monitor_cycles'] = True
settings['graph_produce_node_names'] = True
settings['graph_produce_operinfos'] = True
code_gen = CodeGenerator(G, DefaultNamingConvension(G), settings)
if not os.path.exists(project_folder):
os.mkdir(project_folder)
qoutputs = None
if test_results:
np.random.seed(12345)
finput_tensors = []
input_tensors = []
for i, node in enumerate(G.input_nodes()):
out_q = G.quantization[NodeId(node)].out_qs[0]
if input_file:
file_per_input = glob_input_files(input_file, G.num_inputs)[0]
finput = import_data(file_per_input[i], **input_args)
else:
min_val = out_q.min if not out_q.is_floating else -1.0
max_val = out_q.max if not out_q.is_floating else 1.0
finput = get_rand(node.out_dims[0].shape,
low_high=(min_val, max_val))
finput_tensors.append(finput)
executer = GraphExecuter(G, qrecs=G.quantization)
qoutput_tensors = executer.execute(finput_tensors.copy(),
qmode=QuantizationMode.all())
qoutputs = []
for params in G.outputs():
outp = qoutput_tensors[params.step_idx][0]
qoutputs.append(outp)
for i, params in enumerate(G.input_nodes()):
inp = qoutput_tensors[params.step_idx][0]
input_tensors.append(inp)
if save_inputs:
nodeq = G.quantization[NodeId(params, None)].out_qs[0]
np.save(os.path.join(project_folder, f"fake_input_{i}.npy"),
nodeq.dequantize(inp))
main = os.path.join(project_folder, f"{code_gen.project_name}")
main_c = main + '.c'
main_h = main + '.h'
common_mk = os.path.join(project_folder, "common.mk")
nntool_script = os.path.join(project_folder, "nntool_script")
if overwrite or not os.path.exists(main_c):
with open(os.path.join(project_folder, f"{code_gen.project_name}.c"),
"w") as output_fp:
output_fp.write(
generate_main_appl_template(G, code_gen, input_tensors,
qoutputs, tolerance))
if overwrite or not os.path.exists(main_h):
with open(os.path.join(project_folder, f"{code_gen.project_name}.h"),
"w") as output_fp:
output_fp.write(generate_main_appl_header(G, code_gen))
if overwrite or not os.path.exists(common_mk):
open_args = parse_last_open(script_commands)
open_args = build_last_open_args(open_args) if open_args else ""
with open(os.path.join(project_folder, "common.mk"), "w") as output_fp:
if gen_atproject:
output_fp.write(
generate_main_appl_make_atproject(G, code_gen, quantized,
'Model.c'))
else:
output_fp.write(
generate_main_appl_make(G,
code_gen,
quantized,
open_args=open_args))
if overwrite or not os.path.exists(nntool_script):
with open(nntool_script, 'w') as fp:
# NOTE - gen_template_project is excluded so that tests work. Normally it will not be in the
# history.
fp.writelines(process_script(script_commands))
# always add performance since the main template uses it
for setting in [
'set graph_produce_node_names true',
'set graph_produce_operinfos true',
'set graph_monitor_cycles true'
]:
fp.write(f'{setting}\n')
if dump_tensors:
fp.write('set graph_dump_tensor 7\n')
if script_commands[-1] != "save_state":
fp.write('save_state\n')
if gen_atproject:
code_gen = CodeGenerator(G, DefaultNamingConvension(G), settings)
with open(os.path.join(project_folder, 'Model.c'), "w") as output_fp:
output_fp.write(default_template(G, code_generator=code_gen))
if G.has_expressions:
with open(os.path.join(project_folder, "Expression_Kernels.c"),
"w") as output_fp:
output_fp.write(
basic_kernel_source_template(G, code_generator=code_gen))
with open(os.path.join(project_folder, "Expression_Kernels.h"),
"w") as output_fp:
output_fp.write(
basic_kernel_header_template(G, code_generator=code_gen))
code_gen.write_constants(tensor_directory=project_folder)
ignore_function = None if overwrite else skip_existing_files(
project_folder)
shutil.copytree(os.path.join(os.environ.get("NNTOOL_PATH"),
'generation/project_template'),
project_folder,
dirs_exist_ok=True,
ignore=ignore_function)
if not gen_atproject:
try:
shutil.copy(
G.graph_identity.filename,
os.path.join(project_folder,
os.path.split(G.graph_identity.filename)[1]))
except shutil.SameFileError:
pass
