Fix handling of negative seeds in random number generator op kernels for XLA

Casting negative s32 number to u64 directly will have leading 1s in the representation which is not what we want to get a single u64 out of two s32 seeds. Fixed this by first getting unsigned number of the same bit-width.

PiperOrigin-RevId: 345618958
This commit is contained in:
A. Unique TensorFlower 2020-12-04 00:03:15 -08:00 committed by TensorFlow MLIR Team
parent 9456af5880
commit e87d53742b
1 changed files with 2 additions and 10 deletions

View File

@ -518,19 +518,11 @@ void ConvertOp::build(OpBuilder& builder, OperationState& result, Value operand,
}
OpFoldResult ConvertOp::fold(ArrayRef<Attribute> operands) {
auto operand_ty = getOperand().getType().cast<TensorType>();
auto result_ty = getResult().getType().cast<TensorType>();
if (operand_ty == result_ty) return getOperand();
if (getOperand().getType() == getResult().getType()) return getOperand();
// If the result has non-static shape, a convert op is necessary to go from
// static shape to non-static shape.
if (!result_ty.hasStaticShape()) return {};
// TODO(hinsu): Handle unsigned types.
if (operand_ty.getElementType().isUnsignedInteger() ||
result_ty.getElementType().isUnsignedInteger()) {
return {};
}
if (!getResult().getType().cast<TensorType>().hasStaticShape()) return {};
// If the operand is constant, we can do the conversion now.
if (auto elementsAttr = operands.front().dyn_cast_or_null<ElementsAttr>()) {