def main(argv):
args = xdnn_io.processCommandLine(argv)
ret, handles = xdnn.createHandle(args['xclbin'], "kernelSxdnn_0")
# ret = xdnn.createHandle(g_xclbin, "kernelSxdnn_0", g_xdnnLib)
if ret != 0:
sys.exit(1)
labels = xdnn_io.get_labels(args['labels'])
# TODO dict of tuples instead?
fpgaRT = {}
fpgaOutputs = {}
fcWeights = {}
fcBiases = {}
netFiles = {}
confNames = []
args = args['jsoncfg'] # we do not use other args' keys
for netconf_args in args:
confName = str(netconf_args['name'])
confNames += [confName]
# netconf_args['netcfg'] = './data/{}_{}.json'.format(netconf_args['net'], netconf_args['dsp'])
fpgaRT[confName] = xdnn.XDNNFPGAOp(handles, netconf_args)
netconf_args['in_shape'] = tuple((netconf_args['batch_sz'],) + tuple(fpgaRT[confName].getInputDescriptors().itervalues().next()[1:] ))
(fcWeights[confName],
fcBiases[confName]) = xdnn_io.loadFCWeightsBias(netconf_args)
fpgaOutputs[confName] = np.empty ((netconf_args['batch_sz'], int(netconf_args['fpgaoutsz']),), dtype=np.float32, order='C')
netFiles[confName] = str(netconf_args['netcfg'])
batchArrays = []
for streamId, netconf_args in enumerate(args):
batchArrays.append(np.empty(netconf_args['in_shape'], dtype=np.float32, order='C'))
pl = []
img_paths = xdnn_io.getFilePaths(netconf_args['images'])
for j, p in enumerate(img_paths[:netconf_args['batch_sz']]):
batchArrays[-1][j, ...], _ = xdnn_io.loadImageBlobFromFile(p, netconf_args['img_raw_scale'],
netconf_args['img_mean'],
netconf_args['img_input_scale'],
netconf_args['in_shape'][2],
netconf_args['in_shape'][3])
pl.append(p)
confName = str(netconf_args['name'])
firstInputName = fpgaRT[confName].getInputs().iterkeys().next()
firstOutputName = fpgaRT[confName].getOutputs().iterkeys().next()
fpgaRT[confName].exec_async({ firstInputName : batchArrays[-1] }, { firstOutputName : fpgaOutputs[confName] }, streamId)
for streamId, confName in enumerate(confNames):
fpgaRT[confName].get_result (streamId)
for netconf_args in args:
confName = str(netconf_args['name'])
fcOut = np.empty( (netconf_args['batch_sz'], netconf_args['outsz']), dtype=np.float32, order = 'C')
xdnn.computeFC (fcWeights[confName], fcBiases[confName], fpgaOutputs[confName], fcOut)
softmaxOut = xdnn.computeSoftmax(fcOut)
xdnn_io.printClassification(softmaxOut, netconf_args['images'], labels);
xdnn.closeHandle()
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 main():
args = xdnn_io.processCommandLine()
ret, handles = xdnn.createHandle(args['xclbin'], "kernelSxdnn_0")
if ret != 0:
sys.exit(1)
fpgaRT = xdnn.XDNNFPGAOp(handles, args)
fcWeight, fcBias = xdnn_io.loadFCWeightsBias(args)
img_paths = xdnn_io.getFilePaths(args['images'])
fpgaOutput = np.empty((
args['batch_sz'],
args['fpgaoutsz'],
),
dtype=np.float32,
order='C')
fcOutput = np.empty((
args['batch_sz'],
args['outsz'],
),
dtype=np.float32,
order='C')
batch_array = np.empty(((args['batch_sz'], ) + args['in_shape']),
dtype=np.float32,
order='C')
labels = xdnn_io.get_labels(args['labels'])
if args['golden']:
goldenMap = xdnn_io.getGoldenMap(args['golden'])
top5Count = 0
top1Count = 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']]):
batch_array[j, ...], _ = xdnn_io.loadImageBlobFromFile(
p, args['img_raw_scale'], args['img_mean'],
args['img_input_scale'], args['in_shape'][2],
args['in_shape'][1])
pl.append(p)
fpgaRT.execute(batch_array, fpgaOutput)
xdnn.computeFC(fcWeight, fcBias, fpgaOutput, args['batch_sz'],
args['outsz'], args['fpgaoutsz'], fcOutput)
softmaxOut = xdnn.computeSoftmax(fcOutput)
xdnn_io.printClassification(softmaxOut, pl, labels)
if args['golden']:
for j, p in enumerate(img_paths[i:i + args['batch_sz']]):
top1Count += xdnn_io.isTopK(softmaxOut[j], goldenMap, p,
labels, 1)
top5Count += xdnn_io.isTopK(softmaxOut[j], goldenMap, p,
labels, 5)
xdnn.closeHandle()
if args['golden']:
print("\nAverage accuracy (n=%d) Top-1: %.1f%%, Top-5: %.1f%%\n") % (
len(img_paths), float(top1Count) / float(len(img_paths)) * 100.,
float(top5Count) / float(len(img_paths)) * 100.)
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()