INNER CODE UNIT · Python
compute_sdot_vec_vec
PaddlePaddle/Paddle-Lite · lite/tools/convert_arm_sdot_to_machine_code.py:21
def compute_sdot_vec_vec(vd, vn, vm):
i = 0x4e809400 | int(vd) | (int(vn) << 5) | (int(vm) << 16)
return '".word 0x{:08x}\\n"'.format(i) + \
' /* sdot v{vd}.4s, v{vn}.16b, v{vm}.16b */'.format(
vd=vd, vn=vn, vm=vm)
def compute_sdot_vec_elem(vd, vn, vm, idx):
i = 0x4f80e000 | int(vd) | (int(vn) << 5) | (int(vm) << 16) | (int(
idx % 2) << 21) | (int(idx / 2) << 11)
return '".word 0x{:08x}\\n"'.format(i) + \
' /* sdot v{vd}.4s, v{vn}.16b, v{vm}.4b[{idx}] */\\\r\n'.format(
vd=vd, vn=vn, vm=vm, idx=idx)
def match_sdot_patten(line):
matched = re.search(
r'sdot\s+v(.*?).4s\s*,\s*v(.*?).16b\s*,\s*v(.*?).4b\[(.*?)\].*', line,