def main():
num_qpus, unroll_shift, code_offset = map(int, sys.argv[1:])
for insn in assemble(qpu_stbmv, num_qpus=num_qpus,
unroll_shift=unroll_shift, code_offset=code_offset):
print(f'UINT64_C({insn:#018x}),')
def main():
num_qpus, code_offset = map(int, sys.argv[1:])
for insn in assemble(qpu_sgemm_rnn, num_qpus=num_qpus,
code_offset=code_offset):
print(f'UINT64_C({insn:#018x}),')
def main():
num_qpus, unroll, code_offset = map(int, sys.argv[1:])
for insn in assemble(qpu_somatcopy_n,
num_qpus=num_qpus,
unroll=unroll,
code_offset=code_offset):
print(f'UINT64_C({insn:#018x}),')
def main():
if len(sys.argv) == 1:
benchmark()
return
num_qpus, tile_rows, tile_cols, subtile_rows, subtile_cols, code_offset \
= map(int, sys.argv[1:])
for insn in assemble(qpu_comatcopy_t,
num_qpus=num_qpus,
tile_rows=tile_rows,
tile_cols=tile_cols,
subtile_rows=subtile_rows,
subtile_cols=subtile_cols,
code_offset=code_offset):
print(f'UINT64_C({insn:#018x}),')
def program(self, prog, *args, **kwargs):
asm = assemble(prog, *args, **kwargs)
offset = self.code_pos
code = Array(
shape=len(asm),
dtype=np.uint64,
phyaddr=self.memory.phyaddr + offset,
buffer=self.memory.buffer,
offset=offset,
)
self.code_pos += code.nbytes
if self.code_pos > self.code_area_base + self.code_area_size:
raise DriverError('Code too large')
code[:] = asm
return code
def dump_program(self, prog, *args, file=sys.stdout, **kwargs):
self.dump_code(assemble(prog, *args, **kwargs), file=file)
def main():
num_qpus, do_unroll = map(int, sys.argv[1:])
for insn in assemble(qpu_fft2, num_qpus=num_qpus, do_unroll=do_unroll):
print(f'UINT64_C({insn:#018x}),')
def main():
num_qpus, is_forward = map(int, sys.argv[1:])
for insn in assemble(qpu_fft4, num_qpus=num_qpus, is_forward=is_forward):
print(f'UINT64_C({insn:#018x}),')
