Inventor(s)

Abstract

One embodiment provides a dot-product engine that includes a reconfigurable adder tree whose effective accumulation width is controlled by a runtime alignment-bit setting. The engine can operate in a first mode that accumulates N operands using a wider internal alignment/precision, and in a second mode that accumulates 2N operands using a narrower internal alignment/precision on the same physical adder structure by segmenting carry chains or otherwise partitioning adder resources. The runtime alignment-bit control selects among modes (including precision/throughput combinations), enabling the same numeric data type to execute at different precision-throughput points without duplicating the adder tree, and optionally coordinating rounding and buffering to maintain acceptable numerical quality and bandwidth utilization.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Share

COinS