def generateLayerwiseJson(layername):
#args = xdnn_io.processCommandLine()
parser = xdnn_io.default_parser_args()
parser.add_argument('--layerindex', type=int, default=0, help='Index value for layer in json', required=True)
argvt = parser.parse_args()
args = xdnn_io.make_dict_args(argvt)
with open (args['netcfg'], 'r') as fp:
data = json.load(fp)
#print json.dumps(data, indent=2)
# Get layers from json
nodes = data['network']
#print "Total layers (nodes): ", len(nodes)
reachedNode = False
for node in nodes:
if node['active'] == 0:
continue
#print "Active: ", node['active'], " ", node['name']
if reachedNode == False and node['name'] == layername:
reachedNode = True
elif reachedNode and node['name'] != layername:
node['active'] = 0
fname = str(layername) + str('.json')
fjson = fname.replace('/', '_')
with open(fjson, 'w') as wfp:
json.dump(data, wfp, indent=2, sort_keys=True)
return fjson
def main():
parser = xdnn_io.default_parser_args()
parser.add_argument('--numprepproc', type=int, default=1,
help='number of parallel processes used to decode and quantize images')
parser.add_argument('--numstream', type=int, default=16,
help='number of FPGA streams')
parser.add_argument('--deviceID', type=int, default=0,
help='FPGA no. -> FPGA ID to run in case multiple FPGAs')
parser.add_argument('--benchmarkmode', type=int, default=0,
help='bypass pre/post processing for benchmarking')
args = parser.parse_args()
args = xdnn_io.make_dict_args(args)
ret = xdnn.createManager()
if ret != True:
sys.exit(1)
sharedInputArrs = []
fpgaOutputs = []
compilerJSONObj = xdnn.CompilerJsonParser( args['netcfg'])
qPrep = mp.Queue(maxsize=args['numprepproc']*10)
qFpga = mp.Queue(maxsize=100)
streamQ = mp.Queue(maxsize=args['numstream'])
prepProcQ = mp.Queue(maxsize=100)
firstOutputShape = compilerJSONObj.getOutputs().itervalues().next()
firstInputShape = compilerJSONObj.getInputs().itervalues().next()
for i in range( args['numstream'] ):
fpgaOutputs.append(mp.Array(ctypes.c_float, args['batch_sz'] * np.prod( tuple(firstOutputShape[1:]) ) ))
streamQ.put ( i )
for i in range(100):
bufSize = np.prod(tuple(firstInputShape))
sharedInputArrs.append( mp.Array(ctypes.c_float, bufSize ) )
prepProcQ.put (i)
img_paths = xdnn_io.getFilePaths(args['images'])
p = mp.Pool( initializer = init_prepImage, initargs = (args, qPrep, img_paths, sharedInputArrs, prepProcQ, compilerJSONObj, ), processes = args['numprepproc'])
xdnnProc = mp.Process(target=fpga_process_async, args=(qPrep, qFpga, args, len(img_paths), sharedInputArrs,prepProcQ, streamQ, fpgaOutputs, compilerJSONObj,))
xdnnProc.start()
postProc = mp.Process(target=post_process, args=(qFpga, args, img_paths,streamQ, fpgaOutputs,))
postProc.start()
if args['perpetual']:
while True:
res = [p.map_async(run_prepImage, range(len(img_paths)))]
for j in res:
j.wait()
del j
else:
p.map_async(run_prepImage, range(len(img_paths)))
xdnnProc.join()
postProc.join()
p.close()
p.join()
def __init__(self, **kwargs):
arglist = []
for k, v in kwargs.items():
arglist.append("--" + str(k))
arglist.append(str(v))
print arglist
parser = default_parser()
args = parser.parse_args(arglist)
self.args = xdnn_io.make_dict_args(args)
def setup(self, bottom, top):
self.param_dict = eval(self.param_str) # Get args from prototxt
self._args = xdnn_io.make_dict_args(self.param_dict)
self._numPE = self._args[
"batch_sz"] # Bryan hack to detremine number of PEs in FPGA
# Establish FPGA Communication, Load bitstream
ret, handles = xdnn.createHandle(self._args["xclbin"], "kernelSxdnn_0")
if ret != 0:
raise Exception("Failed to open FPGA handle.")
self._args["scaleB"] = 1
self._args["PE"] = -1
# Instantiate runtime interface object
self.fpgaRT = xdnn.XDNNFPGAOp(handles, self._args)
self._indictnames = self._args["input_names"]
self._outdictnames = self._args["output_names"]
self._parser = xdnn.CompilerJsonParser(self._args["netcfg"])
def getCurrentLayerByIndex(index = 0):
#args = xdnn_io.processCommandLine()
parser = xdnn_io.default_parser_args()
parser.add_argument('--layerindex', type=int, default=0, help='Index value for layer in json', required=True)
argvt = parser.parse_args()
args = xdnn_io.make_dict_args(argvt)
if 'layerindex' in args:
index = args['layerindex']
with open(args['netcfg']) as fp:
data = json.load(fp)
# Strip nodes that don't run in hardware
nodes = data['network']
nodes = [x for x in nodes if x['xdnn_kv'] and x['active'] == 1]
# Get layername
if index >= len(nodes):
return None, None
if nodes[index]['xdnn_kv']['slice'] == "0":
return nodes[index]['name'], "DBL"
return nodes[index]['name'], nodes[index]['xdnn_kv']['XNOp']
default=False,
help='batch size')
parser.add_argument('--cutAfter',
default="",
help='Node in graph to start cutting after')
return parser
if __name__ == "__main__":
# Run
parser = default_parser()
args = parser.parse_args()
args = xdnn_io.make_dict_args(args)
netpb = caffe_pb2.NetParameter()
with open(args["inproto"], "r") as f:
pbtf.Parse(f.read(), netpb)
srctensor1 = args["cutAfter"]
outpb = cut_subgraph(netpb, srctensor1, "subgraph0", args)
with open(args["outproto"], "w") as f:
f.write(str(outpb))
# if user passes the train_val prototxt, we will steal the data and accuracy layers
if (args["trainproto"]):
trainpb = caffe_pb2.NetParameter()
def networkForward(netcfg, layername):
#args = xdnn_io.processCommandLine()
parser = xdnn_io.default_parser_args()
parser.add_argument('--layerindex', type=int, default=0, help='Index value for layer in json', required=True)
argvt = parser.parse_args()
args = xdnn_io.make_dict_args(argvt)
args['netcfg'] = netcfg
# Hardcode these parameters, so we only have to look at performance of 1 PE
args["batch_sz"] = 1
args["PE"] = 0
#print "{:-^100}".format(' Before: createHandle ')
ret, handles = xdnn.createHandle(args['xclbin'], "kernelSxdnn_0")
#print "{:-^100}".format(' After: createHandle ')
if ret != 0:
sys.exit(1)
fpgaRT = xdnn.XDNNFPGAOp(handles, args)
#print "{:-^100}".format('1')
fpgaOutput = fpgaRT.getOutputs()
#print "{:-^100}".format('2')
fpgaInput = fpgaRT.getInputs()
#print "{:-^100}".format('3')
img_paths = xdnn_io.getFilePaths(args['images'])
inShape = (args['batch_sz'],) + tuple ( tuple (fpgaRT.getInputDescriptors().values() )[0][1:] )
firstInput = list(fpgaInput.values())[0]
firstOutput = list (fpgaOutput.values())[0]
for i in xrange(0, len(img_paths), args['batch_sz']):
pl = []
for j, p in enumerate(img_paths[i:i + args['batch_sz']]):
firstInput[0, ...], _ = xdnn_io.loadImageBlobFromFile(img_paths[0], args['img_raw_scale'], args['img_mean'], args['img_input_scale'], inShape[2], inShape[3])
pl.append(p)
with open(args['netcfg']) as fp:
data = json.load(fp)
#print json.dumps(data, indent=2)
# Strip nodes that don't run in hardware
nodes = data['network']
nodes = [x for x in nodes if x['xdnn_kv']]
nLayers = len(nodes)
# How many iterations to run, and average across
iterations = 1
# Initialize empty list to hold accumulated runtime
t1 = []
for k in range(iterations):
t1.append(0.0)
# Run N iterations of network permutations
for l in range(iterations):
fpgaRT.execute(fpgaInput, fpgaOutput)
t1[l] += (fpgaRT.get_exec_time())
#for node in nodes:
# print node['name']
# Average it
avetime = sum(t1)/iterations
#print "{:<25} = {:<25}".format(layername, avetime)
return avetime
xdnn.closeHandle()
del fpgaRT
del fpgaInput
del fpgaOutput
del ret
def main():
parser = xdnn_io.default_parser_args()
parser.add_argument(
'--numprepproc',
type=int,
default=1,
help='number of parallel processes used to decode and quantize images')
parser.add_argument('--numstream',
type=int,
default=16,
help='number of FPGA streams')
parser.add_argument(
'--deviceID',
type=int,
default=0,
help='FPGA no. -> FPGA ID to run in case multiple FPGAs')
args = parser.parse_args()
args = xdnn_io.make_dict_args(args)
ret = xdnn.createManager(args['xlnxlib'])
if ret != True:
sys.exit(1)
sharedInputArrs = []
fpgaOutputs = []
qPrep = mp.Queue(maxsize=args['numprepproc'] * 10)
qFpga = mp.Queue(maxsize=100)
streamQ = mp.Queue(maxsize=args['numstream'])
prepProcQ = mp.Queue(maxsize=100)
for i in range(args['numstream']):
shared_arr = mp.Array(ctypes.c_float,
args['batch_sz'] * args['fpgaoutsz'])
fpgaOutputs.append(shared_arr)
streamQ.put(i)
for i in range(100):
bufSize = np.prod(args['in_shape'])
sharedInputArrs.append(mp.Array(ctypes.c_float, bufSize))
prepProcQ.put(i)
img_paths = xdnn_io.getFilePaths(args['images'])
p = mp.Pool(initializer=init_prepImage,
initargs=(
args,
qPrep,
img_paths,
sharedInputArrs,
prepProcQ,
),
processes=args['numprepproc'])
xdnnProc = mp.Process(target=fpga_process_async,
args=(
qPrep,
qFpga,
args,
len(img_paths),
sharedInputArrs,
prepProcQ,
streamQ,
fpgaOutputs,
))
xdnnProc.start()
postProc = mp.Process(target=post_process,
args=(
qFpga,
args,
img_paths,
streamQ,
fpgaOutputs,
))
postProc.start()
if args['perpetual']:
while True:
res = [p.map_async(run_prepImage, range(len(img_paths)))]
for j in res:
j.wait()
del j
else:
p.map_async(run_prepImage, range(len(img_paths)))
xdnnProc.join()
postProc.join()
p.close()
p.join()