import os
import math
import pypto
from numpy.testing import assert_allclose
import torch
def test_gcd_onboard():
device_id = int(os.environ.get('TILE_FWK_DEVICE_ID', 0))
torch.npu.set_device(device_id)
shape = (10, 10)
view_shape = (10, 8)
tile_shape = (10, 8)
pypto.runtime._device_init()
input1 = pypto.tensor(shape, pypto.DT_INT32, "pypto_TENSOR_input1")
input2 = pypto.tensor(shape, pypto.DT_INT32, "pypto_TENSOR_input2")
output = pypto.tensor(shape, pypto.DT_INT32, "pypto_TENSOR_output")
b_loop_num = math.ceil(shape[0] / view_shape[0])
s_loop_num = math.ceil(shape[1] / view_shape[1])
with pypto.function("MAIN", input1, input2, output):
for b_idx in pypto.loop(b_loop_num, name="b0", idx_name="bidx"):
for s_idx in pypto.loop(s_loop_num, name="s0", idx_name="sidx"):
view_tensor_a = pypto.view(input1, view_shape,
[b_idx * view_shape[0], s_idx * view_shape[1]],
valid_shape=[
pypto.min(pypto.symbolic_scalar(shape[0]) - b_idx * view_shape[0],
pypto.symbolic_scalar(view_shape[0])),
pypto.min(pypto.symbolic_scalar(shape[1]) - s_idx * view_shape[1],
pypto.symbolic_scalar(view_shape[1])),
],
)
view_tensor_b = pypto.view(input2, view_shape,
[b_idx * view_shape[0], s_idx * view_shape[1]],
valid_shape=[
pypto.min(pypto.symbolic_scalar(shape[0]) - b_idx * view_shape[0],
pypto.symbolic_scalar(view_shape[0])),
pypto.min(pypto.symbolic_scalar(shape[1]) - s_idx * view_shape[1],
pypto.symbolic_scalar(view_shape[1])),
],
)
pypto.set_vec_tile_shapes(tile_shape[0], tile_shape[1])
view_tensor_a.move(pypto.gcd(view_tensor_a, view_tensor_b))
pypto.assemble(view_tensor_a, [b_idx * view_shape[0], s_idx * view_shape[1]], output)
del view_tensor_a
assert isinstance(output, pypto.tensor)
a_tensor = torch.randint(
low=-10, high=10, size=[shape[0], shape[1]], dtype=torch.int32)
b_tensor = torch.randint(
low=-10, high=10, size=[shape[0], shape[1]], dtype=torch.int32)
out_tensor = torch.zeros(shape[0], shape[1], dtype=torch.int32)
pto_a_tensor = pypto.from_torch(a_tensor, "a_tensor")
pto_b_tensor = pypto.from_torch(b_tensor, "b_tensor")
pto_out_tensor = pypto.from_torch(out_tensor, "out_tensor")
pypto.runtime._device_run_once_data_from_host(pto_a_tensor, pto_b_tensor, pto_out_tensor)
golden = torch.gcd(a_tensor, b_tensor)
assert_allclose(out_tensor.flatten(), golden.flatten(), rtol=0, atol=0)
pypto.runtime._device_fini()