Competition
Continue optimizing your open-ended implementation to compete for prizes awarded to the top five submissions. Complete the competition individually. Participation is optional and carries no additional course credit.
Winning submissions require an approximately 30-minute oral exam with staff during the week after the lab deadline. Be prepared to explain your design choices, experiments, and results. Staff or Annapurna can make final decisions on submissions eligible for prizes.
The open-ended permissions and workflow apply here as well.
Scoring and Prizes
Prizes will be given to the top 5 submissions of this competition.
Submissions will be benchmarked on generated patterns, photographs, and scanned graphics at a variety of dimensions. Your implementation must pass all correctness checks to be ranked.
For each workload, speedup is the starter’s mean total kernel time divided by yours. Input loading and output are excluded. Your score is:
\[\text{score} = 100\left(1 + \operatorname{mean}(\log_2(\text{speedup}))\right)\]The mean is taken across all benchmark workloads, with each weighted equally.
Try the other presets in workloads.yaml and add your own images. Focus on
images on the order of thousands of pixels on a side, including photographs
and scanned documents or drawings. An optimization that helps clean boundaries
may behave differently on textured scenes. The
Benchmarking guide explains how to add inputs
and compare results.
The staff suite includes additional images and generated patterns that you will not have access to. You may generate your own patterns and test stimuli as well. Do recall that LLMs are NOT permitted for optimizing the kernels themselves.
Prizes will be announced soon, but they have historically included Amazon giftcards and AWS merch. In the words of Professor Fletcher, internships at Annapurna are also on the table :)
Submission
Submit this portion by pushing to your student GitHub repository.
Submit the source from your final reported experiment and a report explaining what you tried and why. Include profiles showing where the program spent its time, correctness results, and before-and-after timings. State which versions you compared and which images you used. The saved experiment numbers and results identify these for you. If you used multiple threads, give the thread count.
Explain whether your changes helped and how the results varied with the input. Include unsuccessful attempts; a well-explained negative result is useful.
Put the report and selected evidence in a tracked directory such as
submission/; the generated experiments/ directory is ignored by Git.
Copy any timing tables and profile figures you cite into that directory. Check
that your committed src/ matches the final experiment you report before pushing.
Use VS Code Remote SSH or scp to retrieve files needed on your laptop.
Download any experiments you want to retain before losing access to the instance;
pushing your code does not back up the ignored experiment history.