Draft

Reply draft

Edit the reply on the left. The right side shows how the numbered points and indented bullets will render on OpenReview. Copy pastes the plain text exactly as written.

Reply text
PreviewSub-bullets need 3 leading spaces

Checked against the paper

  1. App. A.1 says: "Token usage and exact compile counts were not retained, so the baseline comparisons should not be interpreted as matched-compute experiments." The figures 7.78M and ~247K do not appear in the paper. The response does not say whether OpenCode ran inside the same compile-validate-measure loop.
  2. App. B.2, 29 cBench programs: OpenTuner faster on 18, AICORE faster on 10, one tie (consumer_tiff2rgba, 1.03 each). OpenTuner leads by more than 5% on 11, for example office_stringsearch1 2.41 vs 1.00 and consumer_tiff2dither 1.41 vs 1.04. Means: AICORE 1.30, OpenTuner 1.14.
  3. PolyBench ablation: full method 4.49x, without compiler evidence 3.50x, rewrite-only 3.01x, parameter-only 1.03x.

Your edits are kept in this browser only.