Type of some parameters should match the input data type #442
Description
-
pad(),
MLPadOptions::valueis float in current spec, which is used as padded value if theMLPadOptions::modeis "constant", its type should exactly match the input data type, otherwise this may cause precision loss.
(ONNX Pad-18 requires the input data and constant value (a single scalar input tensor) to be the same type T.) -
clamp(),
MLClampOptions::minValueandMLClampOptions::maxValue, current spec states them as a float scalar.
And for the v2 ops, we will also have to consider these: (thanks @fdwr for providing the list)
struct DML_FILL_VALUE_CONSTANT_OPERATOR_DESC
{
const DML_TENSOR_DESC* OutputTensor;
DML_TENSOR_DATA_TYPE ValueDataType;
DML_SCALAR_UNION Value;
};
struct DML_FILL_VALUE_SEQUENCE_OPERATOR_DESC
{
const DML_TENSOR_DESC* OutputTensor;
DML_TENSOR_DATA_TYPE ValueDataType;
DML_SCALAR_UNION ValueStart;
DML_SCALAR_UNION ValueDelta;
};
struct DML_DIAGONAL_MATRIX1_OPERATOR_DESC
{
_Maybenull_ const DML_TENSOR_DESC* InputTensor;
const DML_TENSOR_DESC* OutputTensor;
DML_TENSOR_DATA_TYPE ValueDataType;
DML_SCALAR_UNION Value;
INT DiagonalFillBegin;
INT DiagonalFillEnd;
};