def __init__(self,
xgraph,
arch,
work_dir=os.path.join(os.getcwd(), 'work'),
build_dir=os.getcwd(),
mode='debug'):
super(VAICompiler, self).__init__(xgraph)
if not os.path.isfile(arch):
raise ValueError("Arch file: {} does not exist".format(arch))
self.arch = arch
q_output = self.xgraph.get_quantizer_output()
self.netcfgs = {q_key: q_output.get_q_file(q_key)
for q_key in q_output.keys()}
assert(len(self.netcfgs) == 1)
self.work_dir = work_dir
if not os.path.exists(self.work_dir):
os.makedirs(self.work_dir)
self.build_dir = build_dir if build_dir is not None else work_dir
if not os.path.exists(self.build_dir):
os.makedirs(self.build_dir)
self.mode = mode
self.c_output = CompilerOutput(name=xgraph.get_name())
def __init__(self,
xgraph,
target,
arch,
work_dir=os.path.join(os.getcwd(), 'work'),
build_dir=None,
mode='debug'):
super(DPUCompiler, self).__init__(xgraph)
if not os.path.isfile(arch):
raise ValueError("Arch file: {} does not exist".format(arch))
warnings.warn("This compilation only works with one network"
" configuration at the moment!!")
q_output = self.xgraph.get_quantizer_output()
self.netcfgs = {
q_key: q_output.get_orig_pb(q_key)
for q_key in q_output.keys()
}
self.quant_info = {
q_key: q_output.get_q_info(q_key)
for q_key in q_output.keys()
}
assert (len(self.netcfgs) == 1)
self.work_dir = work_dir
self.build_dir = build_dir if build_dir is not None else work_dir
self.target = target
self.arch = arch
self.mode = mode
self.c_output = CompilerOutput(name=xgraph.get_name())
def __init__(self,
xgraph,
arch,
meta,
cpu_arch='arm64',
work_dir=os.path.join(os.getcwd(), 'work'),
build_dir=os.getcwd(),
mode='debug'):
super(VAICompiler, self).__init__(xgraph)
if not os.path.isfile(arch):
raise ValueError("Arch file: {} does not exist".format(arch))
if cpu_arch != 'arm64':
raise ValueError("Unsupported CPU architecture: {}. Supported"
" architectures are: 'arm64'")
warnings.warn("This compilation only works with one network"
" configuration at the moment!!")
q_output = self.xgraph.get_quantizer_output()
self.netcfgs = {
q_key: q_output.get_q_file(q_key)
for q_key in q_output.keys()
}
assert (len(self.netcfgs) == 1)
self.arch = arch
self.meta = meta
self.cpu_arch = cpu_arch
self.work_dir = work_dir
if not os.path.exists(self.work_dir):
os.makedirs(self.work_dir)
self.build_dir = build_dir if build_dir is not None else work_dir
if not os.path.exists(self.build_dir):
os.makedirs(self.build_dir)
self.mode = mode
self.c_output = CompilerOutput(name=xgraph.get_name())
class VAICompiler(XGraphBaseCompiler):
""" Vitis-AI compiler wrapper for DPUCAHX8H """
xgraph_partitioner = XGraphPartitioner()
xgraph_factory = XGraphFactory()
def __init__(self,
xgraph,
arch,
work_dir=os.path.join(os.getcwd(), 'work'),
build_dir=os.getcwd(),
mode='debug'):
super(VAICompiler, self).__init__(xgraph)
if not os.path.isfile(arch):
raise ValueError("Arch file: {} does not exist".format(arch))
self.arch = arch
q_output = self.xgraph.get_quantizer_output()
self.netcfgs = {q_key: q_output.get_q_file(q_key)
for q_key in q_output.keys()}
assert(len(self.netcfgs) == 1)
self.work_dir = work_dir
if not os.path.exists(self.work_dir):
os.makedirs(self.work_dir)
self.build_dir = build_dir if build_dir is not None else work_dir
if not os.path.exists(self.build_dir):
os.makedirs(self.build_dir)
self.mode = mode
self.c_output = CompilerOutput(name=xgraph.get_name())
def compile(self) -> None:
""" Start DPUCAHX8H compilation """
net_name = list(self.netcfgs.keys())[0]
netcfg = list(self.netcfgs.values())[0]
# We only handle one partition at the moment
Xp = VAICompiler.xgraph_partitioner\
.get_subgraphs(self.xgraph)[0]
subxg_layers = Xp.subgraph_data
xgraph = VAICompiler.xgraph_factory.build_from_xlayer(subxg_layers)
# assert xgraph.get_name() == net_name
input_names = xgraph.get_input_names()
input_shapes = [xgraph.get(in_name).shapes[:]
for in_name in input_names]
output_names = list(Xp.attrs['__top_tensors'].keys()) # xgraph.get_output_names()
output_shapes = [xgraph.get(out_name).shapes[:]
for out_name in output_names]
if len(input_names) > 1:
raise NotImplementedError("VAICompiler only handles models with"
" one input at the moment but found: {}"
.format(len(input_names)))
netcfg=netcfg.replace('deploy_model.pb', 'quantize_eval_model.pb')
command = """
vai_c_tensorflow \
--frozen_pb {} \
--arch {} \
--output_dir {} \
--net_name {} \
--options "{}"
""".format(netcfg, self.arch, self.build_dir, net_name, str(dict()))
logger.info("Command: {}".format(command))
process = subprocess.Popen(command,
shell=True,
cwd=FILE_PATH,
stdout=subprocess.PIPE)
output, error = process.communicate()
logger.debug("{} {}".format(output, error))
if error is not None:
error = error.decode('utf-8')
raise ValueError(error)
in_map = {in_name: in_name for in_name in input_names}
out_map = {out_name: out_name for out_name in output_names}
self.c_output.add(net_name, ['libvart-runner.so'], in_map, out_map)
self.xgraph.set_compiler_output(self.c_output)
# TODO
self.xgraph.meta_attrs['compiled'] = True
self.xgraph.meta_attrs['compiler_libs'] = ['libvart-runner.so']
self.xgraph.meta_attrs['compiler_in_map'] = in_map
self.xgraph.meta_attrs['compiler_out_map'] = out_map
return self.xgraph
class VAICompiler(XGraphBaseCompiler):
""" Vitis-AI compiler wrapper for DPUCZDX8G """
xgraph_partitioner = XGraphPartitioner()
xgraph_factory = XGraphFactory()
def __init__(self,
xgraph,
arch,
meta,
dcf,
cpu_arch='arm64',
work_dir=os.path.join(os.getcwd(), 'work'),
build_dir=os.getcwd(),
mode='debug'):
super(VAICompiler, self).__init__(xgraph)
if not os.path.isfile(arch):
raise ValueError("Arch file: {} does not exist".format(arch))
if cpu_arch != 'arm64':
raise ValueError("Unsupported CPU architecture: {}. Supported"
" architectures are: 'arm64'")
q_output = self.xgraph.get_quantizer_output()
self.netcfgs = {q_key: q_output.get_q_file(q_key)
for q_key in q_output.keys()}
assert(len(self.netcfgs) == 1)
self.arch = arch
self.meta = meta
self.dcf = dcf
self.cpu_arch = cpu_arch
self.work_dir = work_dir
if not os.path.exists(self.work_dir):
os.makedirs(self.work_dir)
self.build_dir = build_dir if build_dir is not None else work_dir
if not os.path.exists(self.build_dir):
os.makedirs(self.build_dir)
self.mode = mode
self.c_output = CompilerOutput(name=xgraph.get_name())
def compile(self) -> None:
""" Start DPUv2 compilation """
net_name = list(self.netcfgs.keys())[0]
netcfg = list(self.netcfgs.values())[0]
# We only handle one partition at the moment
Xp = VAICompiler.xgraph_partitioner\
.get_subgraphs(self.xgraph)[0]
subxg_layers = Xp.subgraph_data
xgraph = VAICompiler.xgraph_factory.build_from_xlayer(subxg_layers)
# assert xgraph.get_name() == net_name
input_names = xgraph.get_input_names()
input_shapes = [xgraph.get(in_name).shapes[:]
for in_name in input_names]
output_names = list(Xp.attrs['__top_tensors'].keys()) # xgraph.get_output_names()
output_shapes = [xgraph.get(out_name).shapes[:]
for out_name in output_names]
if len(input_names) > 1:
raise NotImplementedError("VAICompiler only handles models with"
" one input at the moment but found: {}"
.format(len(input_names)))
#command = """
#vai_c_tensorflow \
# --frozen_pb {} \
# --arch {} \
# --output_dir {} \
# --net_name {} \
# --options "{}"
#""".format(netcfg, self.arch, self.work_dir, net_name, str(dict()))
# import pdb; pdb.set_trace()
command = """
dnnc-dpuv2 --parser tensorflow\
--frozen_pb {} \
--cpu_arch {} \
--output_dir {} \
--net_name {} \
--dcf {}
""".format(netcfg, self.cpu_arch, self.work_dir, net_name, self.dcf)
logger.info("Command: {}".format(command))
process = subprocess.Popen(command,
shell=True,
cwd=FILE_PATH,
stdout=subprocess.PIPE)
output, error = process.communicate()
logger.debug("{} {}".format(output, error))
if output is not None:
output = output.decode('utf-8')
logger.info("Output: {}".format(output))
logger.info("Output names: {}".format(output_names))
do = DNNCOutput(str(repr(output)))
dpu_input_nodes = do.get_input_nodes()
dpu_output_nodes = do.get_output_nodes()
dpu_output_nodes_on_shapes = do.get_output_nodes_on_shapes()
in_shapes_log = ["{}*{}*{}".format(ishape[1], ishape[2], ishape[3])
for ishape in input_shapes]
out_shapes_log = ["{}*{}*{}".format(os[1], os[2], os[3])
for os in output_shapes]
in_map = {in_name: in_name + ':0' for in_name, _ in zip(input_names, in_shapes_log)}
out_map = {}
for out_name, out_shape_str in zip(output_names, out_shapes_log):
# DNNC changes naming
dnnc_out_name = do.get_dnnc_str(out_name)
if dnnc_out_name in dpu_output_nodes:
out_map[out_name] = dpu_output_nodes[dnnc_out_name]
# out_name: dpu_output_nodes[out_shape_str] + ':0'
else:
assert len(dpu_output_nodes_on_shapes) == len(output_names),\
"Can't retrieve right out tensor names from DNNC compiler output"
out_map[out_name] = dpu_output_nodes_on_shapes[out_shape_str]
logger.info("DPU kernel in_map: {}".format(in_map))
logger.info("DPU kernel out_map: {}".format(out_map))
if error is not None:
error = error.decode('utf-8')
raise ValueError(error)
logger.info("VAI_C Output: {}".format(output))
logger.info("VAI_C Error: {}".format(error))
logger.debug("CROSS COMPILATION")
command = """
aarch64-linux-gnu-gcc -fPIC -shared {}/dpu_{}.elf -o {}/libdpumodel{}.so
""".format(self.work_dir, net_name, self.work_dir, net_name)
logger.debug("Command: {}".format(command))
process = subprocess.Popen(command.split(),
cwd=FILE_PATH,
stdout=subprocess.PIPE)
output, error = process.communicate()
if output is not None:
output = output.decode('utf-8')
if error is not None:
error = error.decode('utf-8')
raise ValueError(error)
logger.debug("Output: {}".format(output))
logger.debug("Error: {}".format(error))
lib_file = "{}/libdpumodel{}.so".format(self.work_dir, net_name)
to_lib_file = "{}/libdpumodel{}.so".format(self.build_dir, net_name)
shutil.move(lib_file, to_lib_file)
# meta_file = "{}/meta.json".format(self.work_dir)
self.meta["vitis_dpu_kernel"] = net_name
to_meta_file = "{}/meta.json".format(self.build_dir)
# shutil.move(meta_file, to_meta_file)
with open(to_meta_file, 'w') as f:
json.dump(self.meta, f)
self.c_output.add(net_name, [to_lib_file], in_map, out_map)
self.xgraph.set_compiler_output(self.c_output)
return self.xgraph
class DPUCompiler(XGraphBaseCompiler):
""" TODO """
xgraph_partitioner = XGraphPartitioner()
xgraph_factory = XGraphFactory()
tf_generator = TfGenerator()
def __init__(self,
xgraph,
target,
arch,
work_dir=os.path.join(os.getcwd(), 'work'),
build_dir=None,
mode='debug'):
super(DPUCompiler, self).__init__(xgraph)
if not os.path.isfile(arch):
raise ValueError("Arch file: {} does not exist".format(arch))
warnings.warn("This compilation only works with one network"
" configuration at the moment!!")
q_output = self.xgraph.get_quantizer_output()
self.netcfgs = {
q_key: q_output.get_orig_pb(q_key)
for q_key in q_output.keys()
}
self.quant_info = {
q_key: q_output.get_q_info(q_key)
for q_key in q_output.keys()
}
assert (len(self.netcfgs) == 1)
self.work_dir = work_dir
self.build_dir = build_dir if build_dir is not None else work_dir
self.target = target
self.arch = arch
self.mode = mode
self.c_output = CompilerOutput(name=xgraph.get_name())
def Getopts(self, input_shapes):
return {
"maximumasrelu": True,
"pipelineconvmaxpool": False,
"bytesperpixels": 1,
"dsp": 96,
"memory": 9,
"ddr": "256",
"cpulayermustgo": True,
"forceweightsfullyconnected": True,
"mixmemorystrategy": True,
"maximumasrelu": True,
"pipelineconvmaxpool": True,
'bridges': ['bytype', 'Concat'],
"usedeephi": True,
'placeholdershape': input_shapes
}
def compile(self):
# type: () -> None
""" """
layout_transform_pass = \
XGraphLayoutTransformationPass('NHWC', target=self.target)
self.xgraph = layout_transform_pass.execute(self.xgraph,
subgraphs_only=False)
# netcfg = list(self.netcfgs.values())[0] # orig pb file
quant_info_file = list(self.quant_info.values())[0] # quant info file
subxg_layers = DPUCompiler.xgraph_partitioner\
.get_subgraphs(self.xgraph)[0].subgraph_data
xgraph = DPUCompiler.xgraph_factory.build_from_xlayer(subxg_layers)
net_name = list(self.netcfgs.keys())[0]
fs = DPUCompiler.tf_generator.generate(xgraph,
'graph',
subgraphs_only=True,
layout='NHWC',
batch_size=1,
placeholder=True,
out_dir=self.work_dir)
netcfg = list(fs.values())[0]
input_names = xgraph.get_input_names()
input_shapes = [
xgraph.get(in_name).shapes.tolist()[:] for in_name in input_names
]
output_names = xgraph.get_output_names()
output_shapes = [
xgraph.get(out_name).shapes.tolist()[:]
for out_name in output_names
]
if len(input_names) > 1:
raise NotImplementedError(
"DPUCompiler only handles models with"
" one input at the moment but found: {}".format(
len(input_names)))
opt_input_shapes = {
in_name: [e if e != -1 else 1 for e in input_shape]
for in_name, input_shape in zip(input_names, input_shapes)
}
opts = self.Getopts(opt_input_shapes)
if not os.path.isfile(quant_info_file):
raise ValueError(
"quant file: {} does not exist".format(quant_info_file))
opts['quant_cfgfile'] = quant_info_file
opts = str(opts)
command = """
vai_c_tensorflow \
--frozen_pb {} \
--arch {} \
--output_dir {} \
--net_name {}\
--options "{}"
""".format(netcfg, self.arch, self.build_dir, 'compiler', opts)
logger.info("command: {}".format(command))
process = subprocess.Popen(command,
shell=True,
cwd=FILE_PATH,
stdout=subprocess.PIPE)
output, error = process.communicate()
if output is not None:
output = output.decode('utf-8')
if 'SUCCESSFUL COMPILATION' not in output:
logger.info(output)
raise ValueError('compiler is failed. Please see the log for'
' more details')
if error is not None:
error = error.decode('utf-8')
# raise ValueError(error)
logger.debug("Output: {}".format(output))
logger.debug("Error: {}".format(error))
compiler_json_file = self.build_dir + '/compiler.json'
with open(compiler_json_file) as json_file:
json_graph = json.load(json_file)
graph_inputs = json_graph["inputs"]
graph_outputs = json_graph["outputs"]
logger.debug("{} {}".format(input_names, graph_inputs))
logger.debug("{} {}".format(output_names, graph_outputs))
in_map = {in_name: in_name for in_name in input_names}
out_node_merged = []
out_nodes = [
graph_output['previous_layers'][0]
for graph_output in graph_outputs
]
for i in range(len(out_nodes)):
out_node_merged.append([
layer['merged'][-1] for layer in json_graph['network']
if layer['name'] == out_nodes[i]
][0])
in_map = {in_name: in_name for in_name in input_names}
out_map = {out_name: t for out_name, t in zip(output_names, out_nodes)}
#out_map = {out_name: out_name for out_name in output_names}
self.c_output.add(net_name, ['dpuv1lib.so'], in_map, out_map)
self.xgraph.set_compiler_output(self.c_output)
# TODO
self.xgraph.meta_attrs['compiled'] = True
self.xgraph.meta_attrs['compiler_libs'] = ['dpuv1lib.so']
self.xgraph.meta_attrs['compiler_in_map'] = in_map
self.xgraph.meta_attrs['compiler_out_map'] = out_map
return self.xgraph