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 init_fpga():
# Instead of using command line, we hard code it here.
# Typing correct args is almost impossible so either do it in .sh or .py
#
global g_args
global g_ctxt
print(" --- INIT FPGA --- \n")
xdnnArgs = build_xdnn_args()
print(xdnnArgs)
g_args = xdnn_io.processCommandLine(xdnnArgs)
print(" --- After parsing --- \n")
print(g_args)
print(" --- Create handle --- \n")
ret, handles = xdnn.createHandle(g_args['xclbin'], "kernelSxdnn_0")
if ret != 0:
print(" --- !!! FAILED: Cannot create handle. --- \n")
sys.exit(1)
print(" --- Create fpgaRT --- \n")
fpgaRT = xdnn.XDNNFPGAOp(handles, g_args)
g_ctxt["fpgaRT"] = fpgaRT
print(" --- Weight and Bias --- \n")
fcWeight, fcBias = xdnn_io.loadFCWeightsBias(g_args)
g_ctxt["fcWeight"] = fcWeight
g_ctxt["fcBias"] = fcBias
print(" --- Init input input/output area --- \n")
if is_deploymode():
g_ctxt['fpgaOutput'] = fpgaRT.getOutputs()
g_ctxt['fpgaInput'] = fpgaRT.getInputs()
g_ctxt['inShape'] = (g_args['batch_sz'], ) + tuple(
fpgaRT.getInputDescriptors().itervalues().next()[1:])
else:
g_ctxt['fpgaOutput'] = np.empty((
g_args['batch_sz'],
g_args['fpgaoutsz'],
),
dtype=np.float32,
order='C')
g_ctxt['batch_array'] = np.empty(
((g_args['batch_sz'], ) + g_args['in_shape']),
dtype=np.float32,
order='C')
g_ctxt['fcOutput'] = np.empty((
g_args['batch_sz'],
g_args['outsz'],
),
dtype=np.float32,
order='C')
print(" --- Get lables --- \n")
g_ctxt['labels'] = xdnn_io.get_labels(g_args['labels'])
# golden? What is that?
# Seems we are done.
print(" --- FPGA INITIALIZED! ---\n")
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 fpga_process_async(qFrom, qTo, args, num_img, sharedInputArrs, prepProcQ,
streamQ, fpgaOutputs):
ret, handles = xdnn.createHandle(args['xclbin'], "kernelSxdnn_0",
[args["deviceID"]])
if ret != 0:
sys.exit(1)
fpgaRT = xdnn.XDNNFPGAOp(handles, args)
qWait = mp.Queue(maxsize=100)
numStreams = args['numstream']
bsz = args['batch_sz']
input_ptrs = []
for i in range(numStreams):
input_ptrs.append([])
numProcessed = 0
t = threading.Thread(target=xdnn_wait,
args=(
fpgaRT,
qWait,
qTo,
prepProcQ,
))
t.start()
#startTime = time.time()
while numProcessed < num_img or args['perpetual']:
img_list = np.full((bsz, ), -1, dtype=np.int32)
sId = streamQ.get()
input_ptrs[sId] = []
shMemIdxArr = []
for j in range(bsz):
(sMemIdx, img_idx) = qFrom.get()
numProcessed += 1
img_list[j] = img_idx
nparr_view = np.frombuffer(sharedInputArrs[sMemIdx].get_obj(),
dtype=np.float32)
nparr_view = nparr_view[np.newaxis, ...]
input_ptrs[sId].append(
nparr_view.ctypes.data_as(ctypes.POINTER(ctypes.c_float)))
shMemIdxArr.append(sMemIdx)
if numProcessed == num_img:
break
npout_view = np.frombuffer(fpgaOutputs[sId].get_obj(),
dtype=np.float32)
fpgaRT.exec_async(input_ptrs[sId], npout_view, sId)
qWait.put((sId, img_list, shMemIdxArr))
qWait.put((None, None, None))
#elapsedTime = ( time.time() - startTime )
#print ( "FPGA_process: ", float(numProcessed)/elapsedTime, "img/s")
t.join()
xdnn.closeHandle()
def fpga_process_async (qFrom, qTo, args, num_img, sharedInputArrs, prepProcQ, streamQ, fpgaOutputs, compJson):
ret, handles = xdnn.createHandle(args['xclbin'], "kernelSxdnn_0", [args["deviceID"]])
if ret != 0:
sys.exit(1)
fpgaRT = xdnn.XDNNFPGAOp(handles, args)
qWait = mp.Queue(maxsize=100)
numStreams = args['numstream']
bsz = args['batch_sz']
input_ptrs = [[] for i in range(numStreams)]
numProcessed = 0
t = threading.Thread(target=xdnn_wait, args=(fpgaRT, qWait, qTo, prepProcQ, ))
t.start()
firstInputName = compJson.getInputs().iterkeys().next()
firstOutputName = compJson.getOutputs().iterkeys().next()
firstOutputShape = compJson.getOutputs().itervalues().next()
firstInputShape = compJson.getInputs().itervalues().next()
#startTime = time.time()
while numProcessed < num_img or args['perpetual']:
img_list = np.full( (bsz,), -1, dtype = np.int32 )
sId = streamQ.get()
input_ptrs[sId] = []
shMemIdxArr = []
for j in range(bsz):
(sMemIdx, img_idx) = qFrom.get()
numProcessed += 1
img_list[j] = img_idx
nparr_view = np.frombuffer(sharedInputArrs[sMemIdx].get_obj(), dtype = np.float32)
#nparr_view = np.frombuffer(sharedInputArrs[sMemIdx].get_obj(), dtype = np.float32).reshape ( tuple ( firstInputShape ))
input_ptrs[sId].append( nparr_view )
shMemIdxArr.append(sMemIdx)
if numProcessed == num_img:
break
npout_view = np.frombuffer(fpgaOutputs[sId].get_obj(), dtype = np.float32).reshape( (args['batch_sz'],) + tuple ( firstOutputShape[1:]) )
fpgaRT.exec_async( {firstInputName : input_ptrs[sId]}, {firstOutputName : npout_view}, sId)
qWait.put((sId, img_list, shMemIdxArr))
qWait.put ((None, None, None))
#elapsedTime = ( time.time() - startTime )
#print ( "FPGA_process: ", float(numProcessed)/elapsedTime, "img/s")
t.join()
xdnn.closeHandle()
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 __init__(self, maxNumStreams):
self._maxNumStreams = maxNumStreams
self._streamsAvailable = []
self._streamInputs = []
self._streamOutputs = []
self._config = xdnn_io.processCommandLine()
ret, handles = xdnn.createHandle(self._config['xclbin'])
if ret != 0:
sys.exit(1)
self._fpgaRT = xdnn.XDNNFPGAOp(handles, self._config)
self._fcWeight, self._fcBias = xdnn_io.loadFCWeightsBias(self._config)
self._labels = xdnn_io.get_labels(self._config['labels'])
for i in range(maxNumStreams):
self._streamsAvailable.append(i)
self._streamInputs.append(None)
self._streamOutputs.append(None)
def executeOnFPGA(sProtoBufPath, Qmode, Inference_Data, handle, name,
num_models):
TOTAL_IMAGES = 128
# Create handle for FPGA
ret, handle = xdnn.createHandle(
"../overlaybins/" + "aws" + "/overlay_1.xclbin", "kernelSxdnn_0")
#Initialize objects to store results
fpgaRT = {}
fpgaOutput = {}
fcWeight = {}
fcBias = {}
netFiles = {}
confNames = []
#Generate batch
batch_array = generateRandomBatch(TOTAL_IMAGES, None)
#Get Image batch to start inference
for i in range(0, num_models):
confNames += [str(i)]
#Generate batch 10 * batchsize
config = initializeFpgaModel(sProtoBufPath, Qmode)
config["PE"] = i
config["name"] = config["name"] + "_" + str(i)
# Load weights to FPGA
config = TransferWeightsFPGA(len(batch_array), config, handle, i)
fpgaRT[str(i)] = xdnn.XDNNFPGAOp(handle, config)
(fcWeight[str(i)], fcBias[str(i)]) = xdnn_io.loadFCWeightsBias(config)
fpgaOutput[str(i)], fcOutput, config = AllocateMemoryToHost(config)
start0 = time.time()
# Schedule FPGA execution asynchronously
for i in range(0, num_models):
fpgaRT[str(i)].exec_async(batch_array, fpgaOutput[str(i)], i)
start1 = time.time()
#Fetch results of all parallel executions
for i in range(0, num_models):
#Get FPGA output
ret = fpgaRT[str(i)].get_result(i)
#Compute Inner product - fully connected layer
xdnn.computeFC(fcWeight[str(i)], fcBias[str(i)], fpgaOutput[str(i)],
config['batch_sz'], config['outsz'],
config['fpgaoutsz'], fcOutput)
#Compute output softmax
softmaxOut = xdnn.computeSoftmax(fcOutput)
#xdnn_io.printClassification(softmaxOut, config['images'], labels);
end = time.time()
print("throughput", (num_models * len(batch_array) / (end - start0)),
"duration", end - start0)
Inference_result = []
#Append results
Inference_Data.append({
"experiment":
str(Qmode) + "_bit_mode",
"duration_overall":
end - start0,
"imgsPerSecAll":
num_models * len(batch_array) / (end - start0),
"num_models_parallel":
num_models
})
xdnn.closeHandle()
Inference_Data = pd.DataFrame(Inference_Data)
# Inference_Data.to_csv('multinet_results.csv')
result = pd.read_csv('multinet_results.csv')
result = result.append(Inference_Data)
result.to_csv('multinet_results.csv')
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