If the result queue is not empty and we start a reduction, the state of the ALU goes to INTRA_LANE_REDUCTION. Here, the VRF write-back of the previous instruction does not happen. Moreover, the INTRA_LANE_REDUCTION needs the result queue for its own purposes.
Changelog
Fixed
Start integer reductions only if the result queue is empty
If the result queue is not empty and we start a reduction, the state of the ALU goes to INTRA_LANE_REDUCTION. Here, the VRF write-back of the previous instruction does not happen. Moreover, the INTRA_LANE_REDUCTION needs the result queue for its own purposes.
Changelog
Fixed
Checklist