Improve inferencing of Float16 operations with constant inputs

XMLWordPrintable

    • Type: Enhancement
    • Resolution: Fixed
    • Priority: P4
    • 26
    • Affects Version/s: 25
    • Component/s: hotspot
    • b05
    • generic

      Currently, we do not infer Float16-specific IR for the following source pattern.

      SRC = Float16.valueOf(src[i]);
      RES = Float16.max(SRC, Float16.NaN)

      The reason is that the current scheme to detect Float16 operations for some of the operations is based on generic pattern matching. In the above case, the second argument gets constant folded to a Float constant, thus bypassing the existing pattern, which expects the inputs to be ConvHF2F.

            Assignee:
            Jatin Bhateja
            Reporter:
            Jatin Bhateja
            Votes:
            0 Vote for this issue
            Watchers:
            4 Start watching this issue

              Created:
              Updated:
              Resolved: