wehub-resource-sync
50 lines
1.8 KiB
Python
50 lines
1.8 KiB
Python
#!/usr/bin/python
|
|
import sys
|
|
import json
|
|
import os
|
|
post_treat = {}
|
|
# neuropilot position
|
|
sdk_path = os.environ["NEURON_SDK"]
|
|
print('NEURON_SDK: ', sdk_path)
|
|
ncccomiple = sdk_path + '/host/bin/ncc-tflite '
|
|
extractshared = sdk_path + '/host/bin/extract-shared '
|
|
archoptions = ' --arch=mdla5.1 --l1-size-kb=7168 --num-mdla=4 '
|
|
print(archoptions)
|
|
clean_tmp = True
|
|
with open(sys.argv[1]) as f:
|
|
post_treat = json.load(f)
|
|
if (len(sys.argv)>=3):
|
|
archoptions = sys.argv[2]
|
|
print("archoptions: ", archoptions)
|
|
merges = post_treat["merge"]
|
|
for key in post_treat["merge"]:
|
|
srcs = merges[key]
|
|
dst = key
|
|
#options = " --opt=3 --opt-footprint --opt-aggressive --stable-linearize --gno=LTS --gno-exp --gno-non-4d-tiling --fc-to-conv --broadcast-act-wgt --split-large-conv-ic=1536 --mdla-int16-lut --suppress-input --suppress-output --show-memory-summary"
|
|
options = " --stable-linearize --gno=LTS --gno-exp --gno-non-4d-tiling --fc-to-conv --broadcast-act-wgt "
|
|
options += " --mdla-int16-lut "
|
|
options += " --split-large-conv-ic=2048"
|
|
options += ' --opt=3 --opt-accuracy --opt-footprint '
|
|
options += " --suppress-input --suppress-output "
|
|
IN_DLA = ""
|
|
OUT_DLA = ""
|
|
dlas = []
|
|
for i,src in enumerate(srcs):
|
|
fullcmd = ncccomiple + " " + src + options + archoptions
|
|
print(fullcmd)
|
|
print(os.popen(fullcmd).read())
|
|
dlas.append(src.replace('tflite', 'dla'))
|
|
IN_DLA += " -i " + src.replace('tflite', 'dla') + ' '
|
|
OUT_DLA += " -o " + key + ".shared" + "_%d" %i
|
|
if clean_tmp:
|
|
os.remove(src)
|
|
if len(srcs) > 1:
|
|
fullcmd = extractshared + " " + IN_DLA + ' ' + OUT_DLA + " -s " + key + '.weight'
|
|
print(fullcmd)
|
|
print(os.popen(fullcmd).read())
|
|
if clean_tmp:
|
|
for dla in dlas:
|
|
os.remove(dla)
|
|
|
|
|