(
method_name, op_type, reverse=False, scalar_method=None
)
| 636 | return _scalar_op_(var, 1.0 / value, 0.0) |
| 637 | |
| 638 | def _binary_creator_( |
| 639 | method_name, op_type, reverse=False, scalar_method=None |
| 640 | ): |
| 641 | def __impl__(self, other_var): |
| 642 | # 1. scalar exists cases |
| 643 | # we need combine the tensor.dtype and scalar.dtype, cast correct object |
| 644 | if isinstance(other_var, float): |
| 645 | # in all cases(+, -, *, /, **, //, %), we need cast tensor.dtype to float |
| 646 | if self.dtype in _supported_int_dtype_: |
| 647 | self = astype(self, 'float32') |
| 648 | # here use `scale` replace `elementwise` to get better performance |
| 649 | # but only +, -, *, / can use this method |
| 650 | if scalar_method is not None: |
| 651 | return scalar_method(self, other_var) |
| 652 | elif isinstance(other_var, int): |
| 653 | # in all cases(+, -, *, /, **, //, %), we can cast it to float |
| 654 | # because the output tensor.dtype depend on the type of input tensor |
| 655 | other_var = float(other_var) |
| 656 | # division is a special case |
| 657 | # NOTE(chenweihang): because we cast tensor to float32 instead float64, |
| 658 | # the division result can only guarantee the numerical accuracy of 6 digits |
| 659 | # after the decimal point. The result of numpy calculation is of float64 type, |
| 660 | # so the calculation result here and the calculation result of numpy are |
| 661 | # different after 6 decimal point. If necessary, we can also use float64 here. |
| 662 | # torch's behavior here is consistent with ours |
| 663 | if ( |
| 664 | op_type == 'elementwise_div' |
| 665 | and self.dtype in _supported_int_dtype_ |
| 666 | ): |
| 667 | self = astype(self, 'float32') |
| 668 | # bool(tensor) + int(scalar) will do type promotion to int64 |
| 669 | if self.dtype == core.VarDesc.VarType.BOOL: |
| 670 | self = astype(self, 'int64') |
| 671 | # here use `scale` replace `elementwise` to get better performance |
| 672 | # but only +, -, *, / can use this method |
| 673 | if scalar_method is not None: |
| 674 | return scalar_method(self, other_var) |
| 675 | elif isinstance(other_var, complex): |
| 676 | if self.dtype not in _supported_complex_dtype_: |
| 677 | self = astype(self, 'complex64') |
| 678 | other_var = create_new_tmp_var( |
| 679 | current_block(self), dtype='complex64' |
| 680 | ) |
| 681 | else: |
| 682 | # do nothing |
| 683 | pass |
| 684 | |
| 685 | # 2. create variable for scalar |
| 686 | lhs_dtype = safe_get_dtype(self) |
| 687 | if not isinstance(other_var, Variable): |
| 688 | if reverse: |
| 689 | for elem in self.shape: |
| 690 | if elem < 0: |
| 691 | other_var = create_tensor_with_batchsize( |
| 692 | self, other_var, lhs_dtype |
| 693 | ) |
| 694 | break |
| 695 | else: |
no test coverage detected