def __init__(self, jsonFile):
with open(jsonFile) as f:
data = json.load(f)
self.parameters = {"name": {"output": data['comm']['ports']['output'], "input": data['comm']['ports']['input']},
"mac_table": data['comm']['mac_table'],
"rank": data['comm']['src_rank'],
"config_hardware": data['config_hardware'],
"neural_net" : data['neural_net'],
"mode" : data['mode']
}
GalapagosNet.__init__(self, self.parameters)
for rank_id, rank in enumerate(self.parameters['neural_net']):
def __init__(self, jsonFile):
with open(jsonFile) as f:
data = json.load(f)
self.parameters = {
"name": {
"output": data['comm']['ports']['output'],
"input": data['comm']['ports']['input']
},
"mac_table": data['comm']['mac_table'],
"rank": data['comm']['src_rank'],
"config_hardware": data['config_hardware'],
"neural_net": data['neural_net'],
"mode": data['mode']
}
GalapagosNet.__init__(self, self.parameters)
for rank_id, rank in enumerate(self.parameters['neural_net']):
self.parameters['neural_net'][rank_id]['rank_info'][
'batch_size'] = 1
self.parameters['neural_net'][rank_id]['rank_info'][
'num_ranks'] = 1
self.parameters['neural_net'][rank_id]['rank_info'][
'ifm_baseaddr'] = int(
self.parameters['neural_net'][rank_id]['rank_info']
['ifm_baseaddr'], 16)
self.parameters['neural_net'][rank_id]['rank_info'][
'ofm_baseaddr'] = int(
self.parameters['neural_net'][rank_id]['rank_info']
['ofm_baseaddr'], 16)
self.parameters['neural_net'][rank_id]['rank_info'][
'biases_weights_baseaddr'] = int(
self.parameters['neural_net'][rank_id]['rank_info']
['biases_weights_baseaddr'], 16)
for cmd_id, cmd in enumerate(
self.parameters['neural_net'][rank_id]['command']):
num_weights = 0
weights = bytearray(0)
with open(rank['command'][cmd_id]['weight_file'], 'r') as fobj:
for line in fobj:
for num in line.split():
weights = bytearray(struct.pack(
'>h', int(num))) + weights
num_weights = num_weights + 1
#weights = [[struct.pack('>I',int(num)) for num in line.split()] for line in fobj]
#print binascii.hexlify(weights)
#self.parameters['neural_net'][rank_id]['command'][cmd_id].update({"weights": np.array(weights, dtype="int16")})
self.parameters['neural_net'][rank_id]['command'][cmd_id][
"weights"] = weights
print "num of weights " + str(num_weights)
if self.parameters['neural_net'][rank_id]['command'][cmd_id][
'bias'] == 1:
biases = bytearray(0)
with open(
self.parameters['neural_net'][rank_id]['command']
[cmd_id]['bias_file'], 'r') as fobj:
for line in fobj:
for num in line.split():
biases = bytearray(struct.pack(
'>h', int(num))) + biases
self.parameters['neural_net'][rank_id]['command'][
cmd_id].update({"biases": biases})
self.parameters['neural_net'][rank_id]['command'][
cmd_id].update({
"configParameters":
self.parameters['neural_net'][rank_id]['command']
[cmd_id]['biases'] +
self.parameters['neural_net'][rank_id]['weights']
})
else:
self.parameters['neural_net'][rank_id]['command'][
cmd_id].update({
"configParameters":
self.parameters['neural_net'][rank_id]['command']
[cmd_id]['weights']
})
print "length of weights2 " + str(
len(self.parameters['neural_net'][rank_id]['command']
[cmd_id]['weights']))
print "length of configparam " + str(
len(self.parameters['neural_net'][rank_id]['command']
[cmd_id]['configParameters']))
# if(cmd_id == (0)):
# #self.parameters['neural_net'][rank_id]['command'][cmd_id].update({"biases": np.array(biases, dtype="int16").byteswap()})
# self.parameters['neural_net'][rank_id]['command'][cmd_id].update({"biases": biases})
# #self.parameters['neural_net'][rank_id]['command'][cmd_id].update({"configParameters" :self.parameters['neural_net'][rank_id]['command'][cmd_id]['biases'].tobytes() + self.parameters['neural_net'][rank_id]['weights'].tobytes()})
# self.parameters['neural_net'][rank_id]['command'][cmd_id].update({"configParameters" :self.parameters['neural_net'][rank_id]['command'][cmd_id]['biases'] + self.parameters['neural_net'][rank_id]['weights']})
#
#
# else:
# self.parameters['neural_net'][rank_id]['command'][cmd_id].update({"configParameters" : self.parameters['neural_net'][rank_id]['command'][cmd_id]['weights']})
##A self.parameters['neural_net'][rank_id]['command'][cmd_id].update({"weights_baseaddr": self.parameters['neural_net'][rank_id]['command'][cmd_id]["biases_weights_baseaddr"]})
#
if (cmd_id == (0)):
if self.parameters['neural_net'][rank_id]['command'][
cmd_id]['bias'] == 1:
self.parameters['neural_net'][rank_id]['command'][
cmd_id][
'biases_weights_baseaddr'] = self.parameters[
'neural_net'][rank_id]['rank_info'][
'biases_weights_baseaddr']
self.parameters['neural_net'][rank_id]['command'][
cmd_id]['weights_baseaddr'] = self.parameters[
'neural_net'][rank_id]['rank_info'][
'biases_weights_baseaddr'] + len(
self.parameters['neural_net'][rank_id]
['command'][cmd_id]['biases'])
else:
self.parameters['neural_net'][rank_id]['command'][
cmd_id][
'biases_weights_baseaddr'] = self.parameters[
'neural_net'][rank_id]['rank_info'][
'biases_weights_baseaddr']
self.parameters['neural_net'][rank_id]['command'][
cmd_id]['weights_baseaddr'] = self.parameters[
'neural_net'][rank_id]['rank_info'][
'biases_weights_baseaddr']
if (cmd_id !=
(len(self.parameters['neural_net'][rank_id]['command']) -
1)):
if self.parameters['neural_net'][rank_id]['command'][
cmd_id]['bias'] == 1:
self.parameters['neural_net'][rank_id]['command'][
cmd_id +
1]['biases_weights_baseaddr'] = self.parameters[
'neural_net'][rank_id]['command'][cmd_id][
'biases_weights_baseaddr'] + len(
self.parameters['neural_net'][rank_id]
['command'][cmd_id]
['configParameters'])
self.parameters['neural_net'][rank_id]['command'][
cmd_id + 1]['weights_baseaddr'] = self.parameters[
'neural_net'][rank_id]['command'][
cmd_id +
1]['biases_weights_baseaddr'] + len(
self.parameters['neural_net'][rank_id]
['command'][cmd_id + 1]['biases'])
else:
self.parameters['neural_net'][rank_id]['command'][
cmd_id +
1]['biases_weights_baseaddr'] = self.parameters[
'neural_net'][rank_id]['command'][cmd_id][
'biases_weights_baseaddr'] + len(
self.parameters['neural_net'][rank_id]
['command'][cmd_id]
['configParameters'])
self.parameters['neural_net'][rank_id]['command'][
cmd_id + 1]['weights_baseaddr'] = self.parameters[
'neural_net'][rank_id]['command'][
cmd_id + 1]['biases_weights_baseaddr']
#print self.parameters['neural_net'][rank_id]['command'][cmd_id]["biases_weights_baseaddr"]
self.parameters['neural_net'][rank_id]['command'][
cmd_id].update({
"ofm_height":
math.ceil((self.parameters['neural_net'][rank_id]
['command'][cmd_id]['ifm_height'] +
2 * self.parameters['neural_net'][rank_id]
['command'][cmd_id]['pad'] -
self.parameters['neural_net'][rank_id]
['command'][cmd_id]['kernel_height']) /
self.parameters['neural_net'][rank_id]
['command'][cmd_id]['stride'] + 1)
})
self.parameters['neural_net'][rank_id]['command'][
cmd_id].update({
"ofm_width":
math.ceil((self.parameters['neural_net'][rank_id]
['command'][cmd_id]['ifm_width'] +
2 * self.parameters['neural_net'][rank_id]
['command'][cmd_id]['pad'] -
self.parameters['neural_net'][rank_id]
['command'][cmd_id]['kernel_width']) /
self.parameters['neural_net'][rank_id]
['command'][cmd_id]['stride'] + 1)
})
self.parameters['neural_net'][rank_id]['command'][
cmd_id].update({
"ifm_slices":
math.ceil(
self.parameters['neural_net'][rank_id]['command']
[cmd_id]['ifm_depth'] /
self.parameters['config_hardware']['C_NUM_OF_ROWS']
)
})
self.parameters['neural_net'][rank_id]['command'][
cmd_id].update({
"ofm_slices":
math.ceil(
self.parameters['neural_net'][rank_id]['command']
[cmd_id]['output_channels'] /
self.parameters['config_hardware']['C_NUM_OF_ROWS']
)
})
print "pool kernel height " + str(
self.parameters['neural_net'][rank_id]['command'][cmd_id]
['pool_kernel_height'])
print "pool kernel width " + str(
self.parameters['neural_net'][rank_id]['command'][cmd_id]
['pool_kernel_width'])
print "pool stride " + str(
self.parameters['neural_net'][rank_id]['command'][cmd_id]
['pool_stride'])
self.parameters['neural_net'][rank_id]['command'][
cmd_id].update({
"ifm_packet_length":
self.parameters['neural_net'][rank_id]['command']
[cmd_id]['ifm_width'] * self.parameters['neural_net']
[rank_id]['command'][cmd_id]['ifm_height'] *
self.parameters['neural_net'][rank_id]['command']
[cmd_id]['ifm_slices']
})
print "ifm slices " + str(
self.parameters['neural_net'][rank_id]['command'][cmd_id]
['ifm_slices'])
self.parameters['neural_net'][rank_id]['command'][
cmd_id].update({
"ifm_depth_offset":
self.parameters['neural_net'][rank_id]['command']
[cmd_id]['ifm_width'] * self.parameters['neural_net']
[rank_id]['command'][cmd_id]['ifm_height'] *
self.parameters['neural_net'][rank_id]['command']
[cmd_id]['ifm_depth']
})
if (cmd_id == 0):
self.parameters['neural_net'][rank_id]['command'][0][
'ifm_baseaddr'] = self.parameters['neural_net'][
rank_id]['rank_info']['ifm_baseaddr']
else:
self.parameters['neural_net'][rank_id]['command'][cmd_id][
'ifm_baseaddr'] = self.parameters['neural_net'][
rank_id]['command'][cmd_id - 1]['ofm_baseaddr']
#self.parameters['neural_net'][rank_id]['command'][cmd_id]['ofm_baseaddr'] = int(self.parameters['neural_net'][rank_id]['command'][cmd_id]['ifm_baseaddr'],16) + self.parameters['neural_net'][rank_id]['command'][cmd_id]['ifm_packet_length']*2
self.parameters['neural_net'][rank_id]['command'][cmd_id][
'ofm_baseaddr'] = self.parameters['neural_net'][rank_id][
'command'][cmd_id]['ifm_baseaddr'] + self.parameters[
'neural_net'][rank_id]['command'][cmd_id][
'ifm_height'] * self.parameters['neural_net'][
rank_id]['command'][cmd_id][
'ifm_width'] * self.parameters[
'neural_net'][rank_id]['command'][
cmd_id]['ifm_depth'] * 2
if (cmd_id == len(
self.parameters['neural_net'][rank_id]['command']) -
1):
self.parameters['neural_net'][rank_id]['rank_info'][
'ofm_baseaddr'] = self.parameters['neural_net'][
rank_id]['command'][cmd_id]['ofm_baseaddr']
self.parameters['neural_net'][rank_id]['command'][
cmd_id].update({
"weights_packet_length":
self.parameters['neural_net'][rank_id]['command']
[cmd_id]['kernel_height'] *
self.parameters['neural_net'][rank_id]['command']
[cmd_id]['kernel_width'] *
self.parameters['neural_net'][rank_id]['command']
[cmd_id]['ifm_depth']
})
self.parameters['neural_net'][rank_id]['command'][
cmd_id].update({
"weight_depth_offset":
self.parameters['neural_net'][rank_id]['command']
[cmd_id]['kernel_height'] *
self.parameters['neural_net'][rank_id]['command']
[cmd_id]['kernel_width'] *
self.parameters['neural_net'][rank_id]['command']
[cmd_id]['ifm_depth'] *
self.parameters['config_hardware']['C_NUM_OF_COLS']
})
self.parameters['neural_net'][rank_id]['command'][
cmd_id].update({
"pool_input_height":
self.parameters['neural_net'][rank_id]['command']
[cmd_id]['ofm_height']
})
self.parameters['neural_net'][rank_id]['command'][
cmd_id].update({
"pool_input_width":
self.parameters['neural_net'][rank_id]['command']
[cmd_id]['ofm_width']
})
print "pool input height " + str(
self.parameters['neural_net'][rank_id]['command'][cmd_id]
['pool_input_height'])
print "pool input width " + str(
self.parameters['neural_net'][rank_id]['command'][cmd_id]
['pool_input_width'])
#self.parameters['neural_net'][rank_id]['command'][cmd_id].update({"pool_input_height" : self.parameters['neural_net'][rank_id]['command'][cmd_id]['ofm_height']})
#self.parameters['neural_net'][rank_id]['command'][cmd_id].update({"pool_input_width" : self.parameters['neural_net'][rank_id]['command'][cmd_id]['ofm_width']})
try:
self.parameters['neural_net'][rank_id]['command'][
cmd_id].update({
"pool_output_height":
math.ceil(
((self.parameters['neural_net'][rank_id]
['command'][cmd_id]['pool_input_height'] -
self.parameters['neural_net'][rank_id]
['command'][cmd_id]['pool_kernel_height']) /
1.0) / self.parameters['neural_net'][rank_id]
['command'][cmd_id]['pool_stride'] / 1.0 + 1)
})
self.parameters['neural_net'][rank_id]['command'][
cmd_id].update({
"pool_output_width":
math.ceil(
((self.parameters['neural_net'][rank_id]
['command'][cmd_id]['pool_input_width'] -
self.parameters['neural_net'][rank_id]
['command'][cmd_id]['pool_kernel_width']) /
1.0) / self.parameters['neural_net'][rank_id]
['command'][cmd_id]['pool_stride'] / 1.0 + 1)
})
except ZeroDivisionError:
self.parameters['neural_net'][rank_id]['command'][
cmd_id].update({"pool_output_height": 0})
self.parameters['neural_net'][rank_id]['command'][
cmd_id].update({"pool_output_width": 0})
# pool_stride has to be a multiple of 2
if (self.parameters['neural_net'][rank_id]['command'][cmd_id]['pool_stride'] != 0 \
and self.parameters['neural_net'][rank_id]['command'][cmd_id]['pool_output_height'] > 5 and self.parameters['neural_net'][rank_id]['command'][cmd_id]['pool_output_width'] > 5 \
and self.parameters['neural_net'][rank_id]['command'][cmd_id]['pool_output_width'] * self.parameters['neural_net'][rank_id]['command'][cmd_id]['pool_kernel_width'] < 1 << 9 \
and self.parameters['neural_net'][rank_id]['command'][cmd_id]['pool_output_width'] < 1 << 8):
# WHEN MAXPOOL IS ENABLED, THE OUTPUT SIZE WILL BE SMALLER
# THERFORE, THE OFM_PACKET_LENGTH HAS TO BE ADJUSTED ACCORDINGLY
self.parameters['neural_net'][rank_id]['command'][
cmd_id].update({
"ofm_packet_length":
self.parameters['neural_net'][rank_id]['command']
[cmd_id]['pool_output_height'] *
self.parameters['neural_net'][rank_id]['command']
[cmd_id]['pool_output_width'] *
self.parameters['neural_net'][rank_id]['command']
[cmd_id]['ofm_slices']
})
else:
self.parameters['neural_net'][rank_id]['command'][cmd_id][
'pool_output_height'] = 0
self.parameters['neural_net'][rank_id]['command'][cmd_id][
'pool_output_width'] = 0
self.parameters['neural_net'][rank_id]['command'][cmd_id][
'pool_kernel_height'] = 0
self.parameters['neural_net'][rank_id]['command'][cmd_id][
'pool_kernel_width'] = 0
self.parameters['neural_net'][rank_id]['command'][cmd_id][
'pool_stride'] = 0
self.parameters['neural_net'][rank_id]['command'][
cmd_id].update({
"ofm_packet_length":
self.parameters['neural_net'][rank_id]['command']
[cmd_id]['ofm_height'] *
self.parameters['neural_net'][rank_id]['command']
[cmd_id]['ofm_width'] *
self.parameters['neural_net'][rank_id]['command']
[cmd_id]['ofm_slices']
})
self.configParametersFPGA(rank_id)
self.configDMAFPGA(rank_id)
self.configCmdFPGA(rank_id)