def XDLFloadWeights(args,
Weights,
outChans,
inChans,
kernH,
kernW,
layerName,
isxdnnv3=False):
print("Loading weights/bias/quant_params to FPGA...")
if isxdnnv3 == "True":
size = xdnn.v3computeWeightsBiasQuantSize(
kernW, kernH, outChans, int(math.ceil(float(inChans) / float(96))),
0, 0, False)
size = size * 2
else:
size = xdnn.computeWeightsBiasQuantSize(\
kernW, kernH, inChans, outChans, True if args['quantizecfg'] else False)
blob = xdnn.makeWeightsBiasQuantBlob(size)
bias = [0 for v in range(outChans)]
if isxdnnv3 == "True":
offset = xdnn.v3fillWeightsBiasQuantBlob(blob, 0, args['quantizecfg'],
Weights, args['scaleA'], bias,
args['scaleB'], kernW, kernH,
inChans, outChans, layerName)
else:
offset = xdnn.fillWeightsBiasQuantBlob(blob, 0, args['quantizecfg'],
Weights, args['scaleA'], bias,
args['scaleB'], kernW, kernH,
inChans, outChans, layerName)
layer2OffsetMap = "%s:%d" % (layerName, 0)
fps = xdnn.loadBlobToDdr(blob, size, layer2OffsetMap, int(args['PE']))
return (blob, fps)
def XDLFBunchloadWeightsBiasQuantLatestRepl(args, allLayerNames,
allLayersWeightsBiasQuantizeKey,
allConvLayerNamesParams, size,
isxdnnv3):
print("Loading weights/bias/quant_params to FPGA...")
blob = xdnn.makeWeightsBiasQuantBlob(size)
offset = 0
layer2OffsetMapStr = ""
for layerName in allLayerNames:
if layer2OffsetMapStr != "":
layer2OffsetMapStr += ","
layer2OffsetMapStr += "%s:%d" % (layerName, offset)
if isxdnnv3 == "True":
offset += xdnn.v3fillWeightsBiasQuantBlob(
blob, offset, args['quantizecfg'],
allLayersWeightsBiasQuantizeKey[layerName]['weights'],
args['scaleA'],
allLayersWeightsBiasQuantizeKey[layerName]['Bias'],
args['scaleB'],
allLayersWeightsBiasQuantizeKey[layerName]['kernW'],
allLayersWeightsBiasQuantizeKey[layerName]['kernH'],
allLayersWeightsBiasQuantizeKey[layerName]['inChans'],
allLayersWeightsBiasQuantizeKey[layerName]['outChans'],
int(allConvLayerNamesParams[layerName]['srcFullSectNum']),
int(allConvLayerNamesParams[layerName]['srcReplSectNum']),
int(allConvLayerNamesParams[layerName]['srcReplUnitNum']),
int(allConvLayerNamesParams[layerName]['srcReplUnitWidth']),
int(allConvLayerNamesParams[layerName]['convHalfRateMode']),
layerName)
else:
offset += xdnn.fillWeightsBiasQuantBlob(
blob, offset, args['quantizecfg'],
allLayersWeightsBiasQuantizeKey[layerName]['weights'],
args['scaleA'],
allLayersWeightsBiasQuantizeKey[layerName]['Bias'],
args['scaleB'],
allLayersWeightsBiasQuantizeKey[layerName]['kernW'],
allLayersWeightsBiasQuantizeKey[layerName]['kernH'],
allLayersWeightsBiasQuantizeKey[layerName]['inChans'],
allLayersWeightsBiasQuantizeKey[layerName]['outChans'],
layerName)
fps = xdnn.loadBlobToDdr(blob, size, layer2OffsetMapStr, int(args['PE']))
return (blob, fps)
def XDLFBunchloadWeightsBiasQuant(args, allLayerNames,
allLayersWeightsBiasQuantizeKey, size,
isxdnnv3):
print("Loading weights/bias/quant_params to FPGA...")
print("MNDBG total A size %d" % size)
blob = xdnn.makeWeightsBiasQuantBlob(size)
layer2OffsetMapStr = ""
offset = 0
for layerName in allLayerNames:
if layer2OffsetMapStr != "":
layer2OffsetMapStr += ","
layer2OffsetMapStr += "%s:%d" % (layerName, offset)
if isxdnnv3 == "True":
offset += xdnn.v3fillWeightsBiasQuantBlob(
blob, offset, args['quantizecfg'],
allLayersWeightsBiasQuantizeKey[layerName]['weights'],
args['scaleA'],
allLayersWeightsBiasQuantizeKey[layerName]['Bias'],
args['scaleB'],
allLayersWeightsBiasQuantizeKey[layerName]['kernW'],
allLayersWeightsBiasQuantizeKey[layerName]['kernH'],
allLayersWeightsBiasQuantizeKey[layerName]['inChans'],
allLayersWeightsBiasQuantizeKey[layerName]['outChans'],
layerName)
else:
offset += xdnn.fillWeightsBiasQuantBlob(
blob, offset, args['quantizecfg'],
allLayersWeightsBiasQuantizeKey[layerName]['weights'],
args['scaleA'],
allLayersWeightsBiasQuantizeKey[layerName]['Bias'],
args['scaleB'],
allLayersWeightsBiasQuantizeKey[layerName]['kernW'],
allLayersWeightsBiasQuantizeKey[layerName]['kernH'],
allLayersWeightsBiasQuantizeKey[layerName]['inChans'],
allLayersWeightsBiasQuantizeKey[layerName]['outChans'],
layerName)
fps = xdnn.loadBlobToDdr(blob, size, layer2OffsetMapStr, int(args['PE']))
return (blob, fps)
def loadWeightsBiasQuantv3(args):
print("Loading weights/bias/quant_params to FPGA...")
allConvLayerNames, allConvLayerNamesParams = parseCompilerFile(
args['netcfg'])
size = 0
fi = 0
while True:
fname = "%s/fwbqb_%d" % (args['datadir'], fi)
if not os.path.isfile(fname):
break
layerName = None
weights = None
bias = None
kern = None
inChans = None
outChans = None
with open(fname, 'r') as f:
data = f.read()
vals = data.strip().split(' ')
layerName = vals[0]
kern = int(vals[1])
assert kern == int(allConvLayerNamesParams[layerName]['kernW'])
assert kern == int(allConvLayerNamesParams[layerName]['kernH'])
inChans = int(vals[2])
assert inChans == int(
allConvLayerNamesParams[layerName]['inChans'])
outChans = int(vals[3])
# assert outChans == int(allConvLayerNamesParams[layerName]['outChans'])
srcFullSectNum = int(
allConvLayerNamesParams[layerName]['srcFullSectNum'])
srcReplSectNum = int(
allConvLayerNamesParams[layerName]['srcReplSectNum'])
srcReplUnitNum = int(
allConvLayerNamesParams[layerName]['srcReplUnitNum'])
size += xdnn.v3computeWeightsBiasQuantSize(kern, kern, outChans,
srcFullSectNum,
srcReplSectNum,
srcReplUnitNum, False)
fi += 1
size = size * 2
blob = xdnn.makeWeightsBiasQuantBlob(size)
fi = 0
offset = 0
layer2OffsetMapStr = ""
while True:
fname = "%s/fwbqb_%d" % (args['datadir'], fi)
if not os.path.isfile(fname):
break
layerName = None
weights = None
bias = None
kern = None
inChans = None
outChans = None
with open(fname, 'r') as f:
data = f.read()
vals = data.strip().split(' ')
layerName = vals[0]
kern = int(vals[1])
assert kern == int(allConvLayerNamesParams[layerName]['kernW'])
assert kern == int(allConvLayerNamesParams[layerName]['kernH'])
inChans = int(vals[2])
assert inChans == int(
allConvLayerNamesParams[layerName]['inChans'])
outChans = int(vals[3])
# assert outChans == int(allConvLayerNamesParams[layerName]['outChans'])
srcFullSectNum = int(
allConvLayerNamesParams[layerName]['srcFullSectNum'])
srcReplSectNum = int(
allConvLayerNamesParams[layerName]['srcReplSectNum'])
srcReplUnitNum = int(
allConvLayerNamesParams[layerName]['srcReplUnitNum'])
srcReplUnitWidth = int(
allConvLayerNamesParams[layerName]['srcReplUnitWidth'])
convHalfRateMode = int(
allConvLayerNamesParams[layerName]['convHalfRateMode'])
vals = vals[4:]
weights = [float(v) for v in vals]
fname = "%s/fwbqb_bias_%d" % (args['datadir'], fi)
with open(fname, 'r') as f:
data = f.read()
vals = data.strip().split(' ')
vals = vals[4:]
bias = [float(v) for v in vals]
if layer2OffsetMapStr != "":
layer2OffsetMapStr += ","
layer2OffsetMapStr += "%s:%d" % (layerName, offset)
offset += xdnn.v3fillWeightsBiasQuantBlob(
blob, offset, args['quantizecfg'], weights, args['scaleA'], bias,
args['scaleB'], kern, kern, inChans, outChans, srcFullSectNum,
srcReplSectNum, srcReplUnitNum, srcReplUnitWidth, convHalfRateMode,
layerName)
fi += 1
xdnn.loadBlobToDdr(blob, size, layer2OffsetMapStr, int(args['PE']))
return blob