import torch
import numpy as np
import torch_npu
from torch_npu.testing.testcase import TestCase, run_tests
from torch_npu.testing.common_utils import create_common_tensor
torch.npu.set_compile_mode(jit_compile=False)
torch.npu.config.allow_internal_format = False
class TestAddbmm(TestCase):
def cpu_op_exec(self, input1, input2, input3, scalar1, scalar2):
output = torch.addbmm(input1, input2, input3, beta=scalar1, alpha=scalar2)
output = output.numpy()
return output
def npu_op_exec(self, input1, input2, input3, scalar1, scalar2):
output = torch.addbmm(input1, input2, input3, beta=scalar1, alpha=scalar2)
output = output.to("cpu")
output = output.numpy()
return output
def npu_op_exec_out(self, input1, input2, input3, scalar1, scalar2):
output = torch.ones_like(input1)
torch.addbmm(input1, input2, input3, beta=scalar1, alpha=scalar2, out=output)
output = output.to("cpu")
output = output.numpy()
return output
def npu_op_exec_inplace(self, input1, input2, input3, scalar1, scalar2):
input1.addbmm_(input2, input3, beta=scalar1, alpha=scalar2)
output = input1.to("cpu")
output = output.numpy()
return output
def cpu_op_transpose_exec(self, input1, input2, input3, scalar1, scalar2):
input3_t = input3.permute(0, 2, 1)
output = torch.addbmm(input1, input2, input3_t, beta=scalar1, alpha=scalar2)
output = output.numpy()
return output
def npu_op_transpose_exec(self, input1, input2, input3, scalar1, scalar2):
input3_t = input3.permute(0, 2, 1)
output = torch.addbmm(input1, input2, input3_t, beta=scalar1, alpha=scalar2)
output = output.to("cpu")
output = output.numpy()
return output
def test_addbmm(self):
shape_format = [
[[np.float16, 0, [3, 5]], [np.float16, 0, [10, 3, 4]], [np.float16, 0, [10, 4, 5]]],
]
for item in shape_format:
cpu_input1, npu_input1 = create_common_tensor(item[0], 0, 100)
cpu_input2, npu_input2 = create_common_tensor(item[1], 0, 100)
cpu_input3, npu_input3 = create_common_tensor(item[2], 0, 100)
scalar1 = np.random.uniform(0, 10)
scalar2 = np.random.uniform(0, 10)
cpu_input1 = cpu_input1.float()
cpu_input2 = cpu_input2.float()
cpu_input3 = cpu_input3.float()
npu_input1 = npu_input1.float()
npu_input2 = npu_input2.float()
npu_input3 = npu_input3.float()
cpu_output = self.cpu_op_exec(cpu_input1, cpu_input2, cpu_input3, scalar1, scalar2)
npu_output = self.npu_op_exec(npu_input1, npu_input2, npu_input3, scalar1, scalar2)
npu_output1 = self.npu_op_exec_out(npu_input1, npu_input2, npu_input3, scalar1, scalar2)
npu_output2 = self.npu_op_exec_inplace(npu_input1, npu_input2, npu_input3, scalar1, scalar2)
self.assertRtolEqual(cpu_output, npu_output, prec=1.e-3)
self.assertRtolEqual(cpu_output, npu_output1, prec=1.e-3)
self.assertRtolEqual(cpu_output, npu_output2, prec=1.e-3)
def test_addbmm_transpose(self):
shape_format = [
[[np.float16, 0, [4, 5]], [np.float16, 0, [10, 4, 7]], [np.float16, 0, [10, 5, 7]]],
]
for item in shape_format:
cpu_input1, npu_input1 = create_common_tensor(item[0], 0, 100)
cpu_input2, npu_input2 = create_common_tensor(item[1], 0, 100)
cpu_input3, npu_input3 = create_common_tensor(item[2], 0, 100)
scalar1 = np.random.uniform(0, 10)
scalar2 = np.random.uniform(0, 10)
cpu_transpose_output = self.cpu_op_transpose_exec(
cpu_input1.float(), cpu_input2.float(), cpu_input3.float(), scalar1, scalar2)
npu_transpose_output = self.npu_op_transpose_exec(
npu_input1.float(), npu_input2.float(), npu_input3.float(), scalar1, scalar2)
self.assertRtolEqual(cpu_transpose_output, npu_transpose_output, prec=1.e-3)
if __name__ == "__main__":
run_tests()