Skip to content
Adrian Sampson edited this page Aug 1, 2014 · 6 revisions

Big Additions

  • Intro rework. See the other to-do page. @sampsyo @luisceze
    • Ensure we sell ACCEPT's advantages over manual transformation (from PLDI reviews). Probably mention relative broadness of our "deployment" (i.e., undergrads have successfully used this). @sampsyo
  • Region selection in Section 4.2. @andreolb
  • Neural network "optimization" description in 4.3. @andreolb @tmoreau89
  • NPU implementation details in 6.2. @tmoreau89
  • Annotation section should emphasize that escape hatches are unsound and intentionally so (something that made PLDI reviewers angry). This is a practical system dealing with imperfect realities. Maybe also place more emphasis training/testing divide to build confidence in results. @sampsyo
  • Perspectives from undergrads in eval.

Details

  • New title? We currently have "software" but the NPU thing is pretty hardwarey. It would also be nice to emphasize the new "practical" part of the motivation: something you can use today.
  • Lots of cuts in 6.2 and 6.3 (too many irrelevant implementation details). @sampsyo
  • Experimental setup details for Zynq in 7.2.
  • Cut AA relaxation stuff.

Results

  • Add NPU results.
  • Add Wisp results.
  • Collect a new batch of x86 results?
  • Add some sort of quantification of the annotation burden?
  • Similarly, some sort of quantification of the value of auto-tuning? How many configurations were explored, and how hard was it to find a good one? Would an exhaustive search find anything better (we could do an exhaustive search for one benchmark as a case study)?
  • Time allowing: re-introduce and measure "nullify" optimization?
  • Error bars in the figure, and possibly statistical tests to show speedup.

Clone this wiki locally